diff options
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU')
69 files changed, 14072 insertions, 15076 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/dereferenceable-declaration.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/dereferenceable-declaration.ll index c92e5c5f8bff..edb36079fdfb 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/dereferenceable-declaration.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/dereferenceable-declaration.ll @@ -1,5 +1,5 @@ ; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=fiji -stop-after=irtranslator -o - %s | FileCheck %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=fiji -stop-after=irtranslator -o - %s | FileCheck %s declare align(8) dereferenceable(8) ptr @declared_with_ret_deref() #0 declare align(8) ptr @unknown_decl() #0 diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll index 1aee6ab24eea..1b879a604d71 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll @@ -403,40 +403,38 @@ define half @v_neg_rcp_f16(half %x) { ; GFX6-IEEE-LABEL: v_neg_rcp_f16: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; ; GFX6-FLUSH-LABEL: v_neg_rcp_f16: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] @@ -460,40 +458,38 @@ define half @v_rcp_f16(half %x) { ; GFX6-IEEE-LABEL: v_rcp_f16: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; ; GFX6-FLUSH-LABEL: v_rcp_f16: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] @@ -517,40 +513,38 @@ define half @v_rcp_f16_arcp(half %x) { ; GFX6-IEEE-LABEL: v_rcp_f16_arcp: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; ; GFX6-FLUSH-LABEL: v_rcp_f16_arcp: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] @@ -575,9 +569,7 @@ define half @v_rcp_f16_arcp_afn(half %x) { ; GFX6: ; %bb.0: ; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-NEXT: v_rcp_f32_e32 v0, v0 -; GFX6-NEXT: v_mul_f32_e32 v0, v1, v0 ; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-NEXT: s_setpc_b64 s[30:31] ; @@ -600,40 +592,38 @@ define half @v_rcp_f16_ulp25(half %x) { ; GFX6-IEEE-LABEL: v_rcp_f16_ulp25: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; ; GFX6-FLUSH-LABEL: v_rcp_f16_ulp25: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] @@ -1454,70 +1444,67 @@ define <2 x half> @v_rcp_v2f16(<2 x half> %x) { ; GFX6-IEEE-LABEL: v_rcp_v2f16: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v3, v0, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v1, v1, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v2, v1, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, 1.0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v3, v1, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; ; GFX6-FLUSH-LABEL: v_rcp_v2f16: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v0, v2 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v4, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v2, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -1526,30 +1513,27 @@ define <2 x half> @v_rcp_v2f16(<2 x half> %x) { ; GFX8-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v6, -v1, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v4, v7, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -1561,26 +1545,23 @@ define <2 x half> @v_rcp_v2f16(<2 x half> %x) { ; GFX8-FLUSH-LABEL: v_rcp_v2f16: ; GFX8-FLUSH: ; %bb.0: ; GFX8-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v7, v8, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v9, v8, v6 -; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, -v1, v4, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, -v3, v5, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v6, v4, v4 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v7, v5, v5 +; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v6, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, 1.0 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v9 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-FLUSH-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -1594,30 +1575,27 @@ define <2 x half> @v_rcp_v2f16(<2 x half> %x) { ; GFX9-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v6, -v1, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v4, v7, v5 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX9-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX9-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -1628,26 +1606,24 @@ define <2 x half> @v_rcp_v2f16(<2 x half> %x) { ; GFX9-FLUSH-LABEL: v_rcp_v2f16: ; GFX9-FLUSH: ; %bb.0: ; GFX9-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-FLUSH-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 +; GFX9-FLUSH-NEXT: v_mov_b32_e32 v4, 1.0 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v1, v1 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v3, v3 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v1 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v5, v6, v1 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v4, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, -v0, v1, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v3, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v5, v5, v1, v1 +; GFX9-FLUSH-NEXT: v_mad_f32 v6, v6, v3, v3 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v4, -v0, v6, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; GFX9-FLUSH-NEXT: v_mul_f32_e32 v1, v7, v1 +; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v1, 0xff800000, v1 -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v4, v6, v3 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v1, v1, v5 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, -v0, v4, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v5, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v1, v1, v5 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-FLUSH-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -1660,30 +1636,27 @@ define <2 x half> @v_rcp_v2f16(<2 x half> %x) { ; GFX10-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-IEEE-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v6, 1.0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v9, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v9, v9, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v10, v10, v6 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v9, v9, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v10, v10, v5 -; GFX10-IEEE-NEXT: v_add_f32_e32 v7, v9, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v8, v10, v8 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v6, -v2, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX10-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_add_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v7 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, 1.0, v2 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-IEEE-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 @@ -1696,24 +1669,21 @@ define <2 x half> @v_rcp_v2f16(<2 x half> %x) { ; GFX10-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-FLUSH-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v6, 1.0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v6 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v4 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v6 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, -v2, v4, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, -v3, v5, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v6, v4, v4 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v7, v5, v5 +; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, 1.0 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-FLUSH-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 @@ -1726,27 +1696,25 @@ define <2 x half> @v_rcp_v2f16(<2 x half> %x) { ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX11-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX11-NEXT: v_cvt_f32_f16_e32 v4, 1.0 +; GFX11-NEXT: v_mov_b32_e32 v4, 1.0 ; GFX11-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX11-NEXT: v_rcp_f32_e32 v2, v2 ; GFX11-NEXT: v_rcp_f32_e32 v3, v3 ; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX11-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v4, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v4, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX11-NEXT: v_mul_f32_e32 v3, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_dual_mul_f32 v2, v6, v2 :: v_dual_and_b32 v3, 0xff800000, v3 -; GFX11-NEXT: v_dual_add_f32 v3, v3, v4 :: v_dual_and_b32 v2, 0xff800000, v2 -; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 -; GFX11-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, 1.0 +; GFX11-NEXT: v_fma_mix_f32 v5, -v0, v2, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v3, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v5, v5, v2, v2 +; GFX11-NEXT: v_fma_f32 v6, v6, v3, v3 +; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v4, -v0, v6, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_dual_mul_f32 v2, v7, v2 :: v_dual_mul_f32 v3, v4, v3 +; GFX11-NEXT: v_and_b32_e32 v2, 0xff800000, v2 +; GFX11-NEXT: v_dual_add_f32 v2, v2, v5 :: v_dual_and_b32 v3, 0xff800000, v3 +; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX11-NEXT: v_cvt_f16_f32_e32 v2, v2 +; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 +; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, 1.0 ; GFX11-NEXT: v_pack_b32_f16 v0, v0, v1 ; GFX11-NEXT: s_setpc_b64 s[30:31] %fdiv = fdiv <2 x half> <half 1.0, half 1.0>, %x @@ -1757,70 +1725,67 @@ define <2 x half> @v_neg_rcp_v2f16(<2 x half> %x) { ; GFX6-IEEE-LABEL: v_neg_rcp_v2f16: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v3, v0, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v1, v1, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v2, v1, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, -1.0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, -1.0, v1, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v3, v1, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v1, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; ; GFX6-FLUSH-LABEL: v_neg_rcp_v2f16: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v0, v2 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, -1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v4, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, -1.0, v1, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v2, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -1829,30 +1794,27 @@ define <2 x half> @v_neg_rcp_v2f16(<2 x half> %x) { ; GFX8-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v1, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, -1.0, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v3, v5 +; GFX8-IEEE-NEXT: v_sub_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, -1.0, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, -1.0, v1 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-IEEE-NEXT: v_sub_f32_e32 v4, v7, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, -1.0, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, -1.0 @@ -1864,26 +1826,23 @@ define <2 x half> @v_neg_rcp_v2f16(<2 x half> %x) { ; GFX8-FLUSH-LABEL: v_neg_rcp_v2f16: ; GFX8-FLUSH: ; %bb.0: ; GFX8-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v7, v8, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v9, v8, v6 -; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v1, v4, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v3, v5, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v6, v4, -v4 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v7, v5, -v5 +; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v6, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, -1.0 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v9 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-FLUSH-NEXT: v_div_fixup_f16 v0, v1, v0, -1.0 @@ -1897,30 +1856,27 @@ define <2 x half> @v_neg_rcp_v2f16(<2 x half> %x) { ; GFX9-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v6, v1, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v6, -1.0, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v3, v5 +; GFX9-IEEE-NEXT: v_sub_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v7, -1.0, v7 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, -1.0, v1 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX9-IEEE-NEXT: v_sub_f32_e32 v4, v7, v5 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, -1.0, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX9-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX9-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, -1.0 @@ -1931,26 +1887,24 @@ define <2 x half> @v_neg_rcp_v2f16(<2 x half> %x) { ; GFX9-FLUSH-LABEL: v_neg_rcp_v2f16: ; GFX9-FLUSH: ; %bb.0: ; GFX9-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-FLUSH-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 +; GFX9-FLUSH-NEXT: v_mov_b32_e32 v4, -1.0 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v1, v1 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v3, v3 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v1 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v5, v6, v1 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v4, -1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, v0, v1, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, v0, v3, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v5, v5, v1, -v1 +; GFX9-FLUSH-NEXT: v_mad_f32 v6, v6, v3, -v3 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v4, -v0, v6, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; GFX9-FLUSH-NEXT: v_mul_f32_e32 v1, v7, v1 +; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v1, 0xff800000, v1 -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v4, v6, v3 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v1, v1, v5 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, -v0, v4, -1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v5, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v1, v1, v5 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-FLUSH-NEXT: v_div_fixup_f16 v0, v1, v0, -1.0 @@ -1963,30 +1917,27 @@ define <2 x half> @v_neg_rcp_v2f16(<2 x half> %x) { ; GFX10-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-IEEE-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v6, -1.0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v9, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v9, v9, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v10, v10, v6 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v9, v9, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v10, v10, v5 -; GFX10-IEEE-NEXT: v_add_f32_e32 v7, v9, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v8, v10, v8 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v6, v2, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v3, v5 +; GFX10-IEEE-NEXT: v_add_f32_e32 v6, -1.0, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v7, -1.0, v7 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_sub_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_sub_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v7 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, -1.0, v2 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, -1.0, v3 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-IEEE-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 @@ -1999,24 +1950,21 @@ define <2 x half> @v_neg_rcp_v2f16(<2 x half> %x) { ; GFX10-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-FLUSH-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v6, -1.0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v6 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v4 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v6 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v2, v4, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v3, v5, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v6, v4, -v4 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v7, v5, -v5 +; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, -1.0 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-FLUSH-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 @@ -2029,27 +1977,25 @@ define <2 x half> @v_neg_rcp_v2f16(<2 x half> %x) { ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX11-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX11-NEXT: v_cvt_f32_f16_e32 v4, -1.0 +; GFX11-NEXT: v_mov_b32_e32 v4, -1.0 ; GFX11-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX11-NEXT: v_rcp_f32_e32 v2, v2 ; GFX11-NEXT: v_rcp_f32_e32 v3, v3 ; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX11-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v4, -1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v4, -1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX11-NEXT: v_mul_f32_e32 v3, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_dual_mul_f32 v2, v6, v2 :: v_dual_and_b32 v3, 0xff800000, v3 -; GFX11-NEXT: v_dual_add_f32 v3, v3, v4 :: v_dual_and_b32 v2, 0xff800000, v2 -; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 -; GFX11-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 +; GFX11-NEXT: v_fma_mix_f32 v5, v0, v2, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v6, v0, v3, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v5, v5, v2, -v2 +; GFX11-NEXT: v_fma_f32 v6, v6, v3, -v3 +; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v4, -v0, v6, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_dual_mul_f32 v2, v7, v2 :: v_dual_mul_f32 v3, v4, v3 +; GFX11-NEXT: v_and_b32_e32 v2, 0xff800000, v2 +; GFX11-NEXT: v_dual_add_f32 v2, v2, v5 :: v_dual_and_b32 v3, 0xff800000, v3 +; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX11-NEXT: v_cvt_f16_f32_e32 v2, v2 +; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 +; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 ; GFX11-NEXT: v_pack_b32_f16 v0, v0, v1 ; GFX11-NEXT: s_setpc_b64 s[30:31] %fdiv = fdiv <2 x half> <half -1.0, half -1.0>, %x @@ -2064,33 +2010,32 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX6-IEEE-NEXT: v_and_b32_e32 v0, 0xffff, v0 ; GFX6-IEEE-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX6-IEEE-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 +; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX6-IEEE-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v2, v2, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v1, v2, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v5, v0 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v0, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v0, v2, v1 -; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v5, v5, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 ; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v5, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v2, v6, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v6, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v5, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v4, v0 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v0, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v0, v1, 1.0 +; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v4, v4, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v4, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v5, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v1, v5, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v2, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v5, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v1, v4, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -2101,39 +2046,37 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX6-FLUSH-NEXT: v_and_b32_e32 v0, 0xffff, v0 ; GFX6-FLUSH-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX6-FLUSH-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 +; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX6-FLUSH-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v2, v2, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v1, v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v5, v0 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v2, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v0 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v1, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v5, v5, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v3, v3, 1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, v4, v5, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v3, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v1, v2, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, v6, v2, v2 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v3, v2 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v1, v6, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v2, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v6, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v5, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v1, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v2, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v1, v5, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v1, v3, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -2143,30 +2086,27 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX8-IEEE-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v6, -v1, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v4, v7, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -2179,26 +2119,23 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX8-FLUSH: ; %bb.0: ; GFX8-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-FLUSH-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v7, v8, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v9, v8, v6 -; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, -v1, v4, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, -v3, v5, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v6, v4, v4 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v7, v5, v5 +; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v6, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, 1.0 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v9 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-FLUSH-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -2213,30 +2150,27 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX9-IEEE-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v6, -v1, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v4, v7, v5 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX9-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX9-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -2248,26 +2182,24 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX9-FLUSH: ; %bb.0: ; GFX9-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-FLUSH-NEXT: v_and_b32_e32 v1, 0x7fff7fff, v0 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX9-FLUSH-NEXT: v_lshrrev_b32_e32 v3, 16, v1 +; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, v3 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v5, 1.0 +; GFX9-FLUSH-NEXT: v_mov_b32_e32 v5, 1.0 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v4, v4 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v2 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v1, v6, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v5, v5, v4 -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v6, v7, v2 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -|v0|, v5, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v5, v7, v4 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v8, -v1, v6, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v0, -|v0|, v5, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v1, v2, v5 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -|v0|, v4, v5 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v6, v6, v2, v2 +; GFX9-FLUSH-NEXT: v_mad_f32 v7, v7, v4, v4 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v8, -v1, v6, v5 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v0, -|v0|, v7, v5 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; GFX9-FLUSH-NEXT: v_mul_f32_e32 v2, v8, v2 ; GFX9-FLUSH-NEXT: v_mul_f32_e32 v0, v0, v4 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v0, 0xff800000, v0 ; GFX9-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v0, v0, v5 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v0, v0, v7 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX9-FLUSH-NEXT: v_div_fixup_f16 v1, v2, v1, 1.0 @@ -2279,32 +2211,29 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX10-IEEE: ; %bb.0: ; GFX10-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-IEEE-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v6, 1.0 ; GFX10-IEEE-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v9, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v9, v9, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v10, v10, v6 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v9, v9, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v10, v10, v5 -; GFX10-IEEE-NEXT: v_add_f32_e32 v7, v9, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v8, v10, v8 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v6, -v2, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX10-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_add_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v7 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, 1.0, v2 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-IEEE-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 @@ -2316,26 +2245,23 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX10-FLUSH: ; %bb.0: ; GFX10-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-FLUSH-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v6, 1.0 ; GFX10-FLUSH-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v6 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v4 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v6 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, -v2, v4, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, -v3, v5, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v6, v4, v4 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v7, v5, v5 +; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, 1.0 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-FLUSH-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 @@ -2346,30 +2272,30 @@ define <2 x half> @v_rcp_v2f16_fabs(<2 x half> %x) { ; GFX11-LABEL: v_rcp_v2f16_fabs: ; GFX11: ; %bb.0: ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_cvt_f32_f16_e32 v5, 1.0 ; GFX11-NEXT: v_and_b32_e32 v1, 0x7fff7fff, v0 -; GFX11-NEXT: v_cvt_f32_f16_e32 v3, v1 -; GFX11-NEXT: v_rcp_f32_e32 v3, v3 -; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v6, v5, v3 ; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v1 -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v6, 1.0 op_sel_hi:[1,0,1] ; GFX11-NEXT: v_cvt_f32_f16_e32 v4, v2 -; GFX11-NEXT: v_fmac_f32_e32 v6, v7, v3 ; GFX11-NEXT: v_rcp_f32_e32 v4, v4 -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v6, 1.0 op_sel_hi:[1,0,1] +; GFX11-NEXT: v_mov_b32_e32 v5, 1.0 ; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v5, v5, v4 -; GFX11-NEXT: v_fma_mix_f32 v8, -|v0|, v5, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v5, v8, v4 -; GFX11-NEXT: v_fma_mix_f32 v0, -|v0|, v5, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_dual_mul_f32 v3, v7, v3 :: v_dual_mul_f32 v0, v0, v4 +; GFX11-NEXT: v_fma_mix_f32 v7, -|v0|, v4, v5 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v7, v7, v4, v4 +; GFX11-NEXT: v_fma_mix_f32 v0, -|v0|, v7, v5 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_mul_f32_e32 v0, v0, v4 ; GFX11-NEXT: v_and_b32_e32 v0, 0xff800000, v0 -; GFX11-NEXT: v_dual_add_f32 v0, v0, v5 :: v_dual_and_b32 v3, 0xff800000, v3 -; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX11-NEXT: v_add_f32_e32 v0, v0, v7 +; GFX11-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 +; GFX11-NEXT: v_rcp_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v0, v0, v2, 1.0 +; GFX11-NEXT: s_waitcnt_depctr 0xfff +; GFX11-NEXT: v_fma_mix_f32 v6, -v1, v3, v5 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v6, v6, v3, v3 +; GFX11-NEXT: v_fma_mix_f32 v8, -v1, v6, v5 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_mul_f32_e32 v3, v8, v3 +; GFX11-NEXT: v_and_b32_e32 v3, 0xff800000, v3 +; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, 1.0 ; GFX11-NEXT: v_pack_b32_f16 v0, v1, v0 ; GFX11-NEXT: s_setpc_b64 s[30:31] @@ -2386,33 +2312,32 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX6-IEEE-NEXT: v_and_b32_e32 v0, 0xffff, v0 ; GFX6-IEEE-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX6-IEEE-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, -1.0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 +; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX6-IEEE-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v2, v2, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v1, v2, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v5, v0 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v0, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v0, v2, v1 -; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v5, v5, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, -1.0 ; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v5, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v2, v6, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v6, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v5, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, -1.0, v1, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v4, v0 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v0, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v0, v1, -1.0 +; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v4, v4, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v4, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v5, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v1, v5, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v2, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v5, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v1, v4, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -2423,39 +2348,37 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX6-FLUSH-NEXT: v_and_b32_e32 v0, 0xffff, v0 ; GFX6-FLUSH-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX6-FLUSH-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, -1.0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 +; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX6-FLUSH-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v2, v2, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v1, v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, -1.0, v1, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v5, v0 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v2, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v0 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v1, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v5, v5, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v3, v3, -1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, v4, v5, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, -1.0, v3, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v1, v2, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, v6, v2, v2 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v3, v2 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v1, v6, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v2, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v6, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v5, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v1, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v2, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v1, v5, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v1, v3, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -2465,30 +2388,27 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX8-IEEE-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v1, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, -1.0, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v3, v5 +; GFX8-IEEE-NEXT: v_sub_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, -1.0, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, -1.0, v1 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-IEEE-NEXT: v_sub_f32_e32 v4, v7, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, -1.0, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, -1.0 @@ -2501,26 +2421,23 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX8-FLUSH: ; %bb.0: ; GFX8-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-FLUSH-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v7, v8, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v9, v8, v6 -; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v1, v4, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v3, v5, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v6, v4, -v4 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v7, v5, -v5 +; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v6, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, -1.0 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v9 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-FLUSH-NEXT: v_div_fixup_f16 v0, v1, v0, -1.0 @@ -2535,30 +2452,27 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX9-IEEE-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v6, v1, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v6, -1.0, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v3, v5 +; GFX9-IEEE-NEXT: v_sub_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v7, -1.0, v7 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, -1.0, v1 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX9-IEEE-NEXT: v_sub_f32_e32 v4, v7, v5 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, -1.0, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX9-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX9-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, -1.0 @@ -2570,26 +2484,24 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX9-FLUSH: ; %bb.0: ; GFX9-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-FLUSH-NEXT: v_and_b32_e32 v1, 0x7fff7fff, v0 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX9-FLUSH-NEXT: v_lshrrev_b32_e32 v3, 16, v1 +; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, v3 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v5, -1.0 +; GFX9-FLUSH-NEXT: v_mov_b32_e32 v5, -1.0 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v4, v4 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v2 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v1, v6, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v5, v5, v4 -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v6, v7, v2 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -|v0|, v5, -1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v5, v7, v4 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v8, -v1, v6, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v0, -|v0|, v5, -1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, v1, v2, v5 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, |v0|, v4, v5 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v6, v6, v2, -v2 +; GFX9-FLUSH-NEXT: v_mad_f32 v7, v7, v4, -v4 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v8, -v1, v6, v5 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v0, -|v0|, v7, v5 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; GFX9-FLUSH-NEXT: v_mul_f32_e32 v2, v8, v2 ; GFX9-FLUSH-NEXT: v_mul_f32_e32 v0, v0, v4 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v0, 0xff800000, v0 ; GFX9-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v0, v0, v5 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v0, v0, v7 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX9-FLUSH-NEXT: v_div_fixup_f16 v1, v2, v1, -1.0 @@ -2601,32 +2513,29 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX10-IEEE: ; %bb.0: ; GFX10-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-IEEE-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v6, -1.0 ; GFX10-IEEE-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v9, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v9, v9, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v10, v10, v6 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v9, v9, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v10, v10, v5 -; GFX10-IEEE-NEXT: v_add_f32_e32 v7, v9, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v8, v10, v8 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v6, v2, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v3, v5 +; GFX10-IEEE-NEXT: v_add_f32_e32 v6, -1.0, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v7, -1.0, v7 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_sub_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_sub_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v7 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, -1.0, v2 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, -1.0, v3 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-IEEE-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 @@ -2638,26 +2547,23 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX10-FLUSH: ; %bb.0: ; GFX10-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-FLUSH-NEXT: v_and_b32_e32 v0, 0x7fff7fff, v0 -; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v6, -1.0 ; GFX10-FLUSH-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v6 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v4 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v6 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v2, v4, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v3, v5, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v6, v4, -v4 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v7, v5, -v5 +; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, -1.0 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-FLUSH-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 @@ -2668,30 +2574,30 @@ define <2 x half> @v_neg_rcp_v2f16_fabs(<2 x half> %x) { ; GFX11-LABEL: v_neg_rcp_v2f16_fabs: ; GFX11: ; %bb.0: ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-NEXT: v_cvt_f32_f16_e32 v5, -1.0 ; GFX11-NEXT: v_and_b32_e32 v1, 0x7fff7fff, v0 -; GFX11-NEXT: v_cvt_f32_f16_e32 v3, v1 -; GFX11-NEXT: v_rcp_f32_e32 v3, v3 -; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v6, v5, v3 ; GFX11-NEXT: v_lshrrev_b32_e32 v2, 16, v1 -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v6, -1.0 op_sel_hi:[1,0,1] ; GFX11-NEXT: v_cvt_f32_f16_e32 v4, v2 -; GFX11-NEXT: v_fmac_f32_e32 v6, v7, v3 ; GFX11-NEXT: v_rcp_f32_e32 v4, v4 -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v6, -1.0 op_sel_hi:[1,0,1] +; GFX11-NEXT: v_mov_b32_e32 v5, -1.0 ; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v5, v5, v4 -; GFX11-NEXT: v_fma_mix_f32 v8, -|v0|, v5, -1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v5, v8, v4 -; GFX11-NEXT: v_fma_mix_f32 v0, -|v0|, v5, -1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_dual_mul_f32 v3, v7, v3 :: v_dual_mul_f32 v0, v0, v4 +; GFX11-NEXT: v_fma_mix_f32 v7, |v0|, v4, v5 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v7, v7, v4, -v4 +; GFX11-NEXT: v_fma_mix_f32 v0, -|v0|, v7, v5 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_mul_f32_e32 v0, v0, v4 ; GFX11-NEXT: v_and_b32_e32 v0, 0xff800000, v0 -; GFX11-NEXT: v_dual_add_f32 v0, v0, v5 :: v_dual_and_b32 v3, 0xff800000, v3 -; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX11-NEXT: v_add_f32_e32 v0, v0, v7 +; GFX11-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX11-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 +; GFX11-NEXT: v_rcp_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v0, v0, v2, -1.0 +; GFX11-NEXT: s_waitcnt_depctr 0xfff +; GFX11-NEXT: v_fma_mix_f32 v6, v1, v3, v5 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v6, v6, v3, -v3 +; GFX11-NEXT: v_fma_mix_f32 v8, -v1, v6, v5 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_mul_f32_e32 v3, v8, v3 +; GFX11-NEXT: v_and_b32_e32 v3, 0xff800000, v3 +; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 ; GFX11-NEXT: v_pack_b32_f16 v0, v1, v0 ; GFX11-NEXT: s_setpc_b64 s[30:31] @@ -2704,70 +2610,67 @@ define <2 x half> @v_rcp_v2f16_arcp(<2 x half> %x) { ; GFX6-IEEE-LABEL: v_rcp_v2f16_arcp: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v3, v0, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v1, v1, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v2, v1, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, 1.0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v3, v1, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; ; GFX6-FLUSH-LABEL: v_rcp_v2f16_arcp: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v0, v2 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v4, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v2, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -2814,11 +2717,8 @@ define <2 x half> @v_rcp_v2f16_arcp_afn(<2 x half> %x) { ; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-NEXT: v_rcp_f32_e32 v0, v0 ; GFX6-NEXT: v_rcp_f32_e32 v1, v1 -; GFX6-NEXT: v_mul_f32_e32 v0, v2, v0 -; GFX6-NEXT: v_mul_f32_e32 v1, v2, v1 ; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-NEXT: s_setpc_b64 s[30:31] @@ -2864,70 +2764,67 @@ define <2 x half> @v_rcp_v2f16_ulp25(<2 x half> %x) { ; GFX6-IEEE-LABEL: v_rcp_v2f16_ulp25: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v3, v0, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v1, v1, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v2, v1, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, 1.0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v3, v1, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; ; GFX6-FLUSH-LABEL: v_rcp_v2f16_ulp25: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v0, v2 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v4, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v2, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -2936,30 +2833,27 @@ define <2 x half> @v_rcp_v2f16_ulp25(<2 x half> %x) { ; GFX8-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v6, -v1, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v4, v7, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -2971,26 +2865,23 @@ define <2 x half> @v_rcp_v2f16_ulp25(<2 x half> %x) { ; GFX8-FLUSH-LABEL: v_rcp_v2f16_ulp25: ; GFX8-FLUSH: ; %bb.0: ; GFX8-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v1 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v7, v8, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v9, v8, v6 -; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v7, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v4, v1 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, -v1, v4, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, -v3, v5, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v6, v4, v4 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v7, v5, v5 +; GFX8-FLUSH-NEXT: v_mad_f32 v1, -v1, v6, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, 1.0 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v9 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-FLUSH-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -3004,30 +2895,27 @@ define <2 x half> @v_rcp_v2f16_ulp25(<2 x half> %x) { ; GFX9-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-IEEE-NEXT: v_lshrrev_b32_e32 v2, 16, v0 -; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v1 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v8, -v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v7 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v10, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v5, v5, v6 -; GFX9-IEEE-NEXT: v_add_f32_e32 v5, v5, v9 -; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v4, v1 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v6, -v1, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v1, -v1, v6 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v1, v1, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v4, v7, v5 +; GFX9-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v4 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX9-IEEE-NEXT: v_and_b32_e32 v1, 0xff800000, v1 ; GFX9-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 +; GFX9-IEEE-NEXT: v_add_f32_e32 v1, v1, v6 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-IEEE-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -3038,26 +2926,24 @@ define <2 x half> @v_rcp_v2f16_ulp25(<2 x half> %x) { ; GFX9-FLUSH-LABEL: v_rcp_v2f16_ulp25: ; GFX9-FLUSH: ; %bb.0: ; GFX9-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-FLUSH-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v2 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 +; GFX9-FLUSH-NEXT: v_mov_b32_e32 v4, 1.0 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v1, v1 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v3, v3 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v1 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v5, v6, v1 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v4, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, -v0, v1, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v3, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v5, v5, v1, v1 +; GFX9-FLUSH-NEXT: v_mad_f32 v6, v6, v3, v3 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v4, -v0, v6, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] ; GFX9-FLUSH-NEXT: v_mul_f32_e32 v1, v7, v1 +; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v1, 0xff800000, v1 -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v4, v6, v3 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v1, v1, v5 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, -v0, v4, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v5, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v1, v1, v5 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-FLUSH-NEXT: v_div_fixup_f16 v0, v1, v0, 1.0 @@ -3070,30 +2956,27 @@ define <2 x half> @v_rcp_v2f16_ulp25(<2 x half> %x) { ; GFX10-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-IEEE-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v6, 1.0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v9, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v9, v9, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v10, v10, v6 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v9, v9, v4 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v10, v10, v5 -; GFX10-IEEE-NEXT: v_add_f32_e32 v7, v9, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v8, v10, v8 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v7 -; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v8 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v6, -v2, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX10-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX10-IEEE-NEXT: v_add_f32_e32 v7, v7, v5 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v6 +; GFX10-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v7 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, 1.0, v2 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-IEEE-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 @@ -3106,24 +2989,21 @@ define <2 x half> @v_rcp_v2f16_ulp25(<2 x half> %x) { ; GFX10-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-FLUSH-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v6, 1.0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v6 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v4 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v6 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, -v2, v4, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, -v3, v5, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v6, v4, v4 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v7, v5, v5 +; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, 1.0 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-FLUSH-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 @@ -3136,27 +3016,25 @@ define <2 x half> @v_rcp_v2f16_ulp25(<2 x half> %x) { ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX11-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX11-NEXT: v_cvt_f32_f16_e32 v4, 1.0 +; GFX11-NEXT: v_mov_b32_e32 v4, 1.0 ; GFX11-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX11-NEXT: v_rcp_f32_e32 v2, v2 ; GFX11-NEXT: v_rcp_f32_e32 v3, v3 ; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX11-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v4, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v4, 1.0 op_sel:[1,0,0] op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX11-NEXT: v_mul_f32_e32 v3, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_dual_mul_f32 v2, v6, v2 :: v_dual_and_b32 v3, 0xff800000, v3 -; GFX11-NEXT: v_dual_add_f32 v3, v3, v4 :: v_dual_and_b32 v2, 0xff800000, v2 -; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 -; GFX11-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, 1.0 +; GFX11-NEXT: v_fma_mix_f32 v5, -v0, v2, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v3, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v5, v5, v2, v2 +; GFX11-NEXT: v_fma_f32 v6, v6, v3, v3 +; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v4, -v0, v6, v4 op_sel:[1,0,0] op_sel_hi:[1,0,0] +; GFX11-NEXT: v_dual_mul_f32 v2, v7, v2 :: v_dual_mul_f32 v3, v4, v3 +; GFX11-NEXT: v_and_b32_e32 v2, 0xff800000, v2 +; GFX11-NEXT: v_dual_add_f32 v2, v2, v5 :: v_dual_and_b32 v3, 0xff800000, v3 +; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX11-NEXT: v_cvt_f16_f32_e32 v2, v2 +; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 +; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, 1.0 ; GFX11-NEXT: v_pack_b32_f16 v0, v0, v1 ; GFX11-NEXT: s_setpc_b64 s[30:31] %fdiv = fdiv <2 x half> <half 1.0, half 1.0>, %x @@ -4033,40 +3911,38 @@ define amdgpu_ps i32 @s_fdiv_v2f16(i32 inreg %a.arg, i32 inreg %b.arg) { define amdgpu_ps i16 @s_rcp_f16(i16 inreg %a.arg) { ; GFX6-IEEE-LABEL: s_rcp_f16: ; GFX6-IEEE: ; %bb.0: -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, 1.0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, s0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[0:1], v1, v1, v0 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v0, v1, v0 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v1, v0 +; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, s0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[0:1], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_readfirstlane_b32 s0, v0 ; GFX6-IEEE-NEXT: ; return to shader part epilog ; ; GFX6-FLUSH-LABEL: s_rcp_f16: ; GFX6-FLUSH: ; %bb.0: -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, 1.0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, s0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[0:1], v1, v1, v0 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v0, v1, v0 +; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, s0 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[0:1], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v1, v0 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_readfirstlane_b32 s0, v0 @@ -4099,40 +3975,38 @@ define amdgpu_ps i16 @s_rcp_f16(i16 inreg %a.arg) { define amdgpu_ps i16 @s_neg_rcp_f16(i16 inreg %a.arg) { ; GFX6-IEEE-LABEL: s_neg_rcp_f16: ; GFX6-IEEE: ; %bb.0: -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, -1.0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, s0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[0:1], v1, v1, v0 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v0, v1, v0 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v1, v0 +; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, s0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[0:1], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_readfirstlane_b32 s0, v0 ; GFX6-IEEE-NEXT: ; return to shader part epilog ; ; GFX6-FLUSH-LABEL: s_neg_rcp_f16: ; GFX6-FLUSH: ; %bb.0: -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, -1.0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, s0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[0:1], v1, v1, v0 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v0, v1, v0 +; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, s0 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[0:1], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v1, v0 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_readfirstlane_b32 s0, v0 @@ -4166,21 +4040,20 @@ define amdgpu_ps i16 @s_rsq_f16(i16 inreg %a.arg) { ; GFX6-IEEE-LABEL: s_rsq_f16: ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, s0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[0:1], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[0:1], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_readfirstlane_b32 s0, v0 ; GFX6-IEEE-NEXT: ; return to shader part epilog @@ -4188,24 +4061,23 @@ define amdgpu_ps i16 @s_rsq_f16(i16 inreg %a.arg) { ; GFX6-FLUSH-LABEL: s_rsq_f16: ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, s0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[0:1], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[0:1], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_readfirstlane_b32 s0, v0 ; GFX6-FLUSH-NEXT: ; return to shader part epilog @@ -4241,36 +4113,35 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, s0 ; GFX6-IEEE-NEXT: s_lshr_b32 s0, s0, 16 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, s0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, -1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[0:1], v0, v0, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v5, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v2, v0, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v6, s[0:1], v1, v1, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v9, -v3, v5, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v9, v5, v5 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v9, v4, v5 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v8, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v10, -v3, v9, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v9, v10, v5, v9 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v9, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v5, v9 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v3, v0, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v6, v8, 1.0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v7, s[0:1], v2, v1, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v3, v8, v8 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v7, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v6, v4, v7 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v3, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v6, v4, v7 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[0:1], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v5, s[0:1], v1, v1, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v8, -v2, v4, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v8, v4, v4 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v6, v5 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v9, v3, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v10, -v2, v9, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v9, v10, v4, v9 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v9, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v8, -v5, v6, 1.0 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v4, v9 +; GFX6-IEEE-NEXT: v_div_scale_f32 v7, s[0:1], -1.0, v1, -1.0 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v8, v6, v6 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v3, v7, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v5, v3, v7 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v4, v2, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v5, v3, v7 ; GFX6-IEEE-NEXT: s_mov_b64 vcc, s[0:1] -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v5, v3, v4 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v3, v1, v2 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v4, v2, v3 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v1, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_lshlrev_b32_e32 v1, 16, v1 @@ -4283,42 +4154,40 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, s0 ; GFX6-FLUSH-NEXT: s_lshr_b32 s0, s0, 16 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, s0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, -1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[0:1], v0, v0, v2 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[0:1], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v0, v2 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[0:1], v1, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[0:1], v1, v1, -1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v4, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, -1.0, v1, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v2, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; GFX6-FLUSH-NEXT: v_or_b32_e32 v0, v0, v1 @@ -4330,31 +4199,28 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX8-IEEE-NEXT: v_sqrt_f16_e32 v0, s0 ; GFX8-IEEE-NEXT: s_lshr_b32 s0, s0, 16 ; GFX8-IEEE-NEXT: v_sqrt_f16_e32 v1, s0 -; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v1 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v2 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v8, -v2, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v8, v8, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v8, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v9 -; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v8, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v10, v10, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v10, v6 -; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v2, v2, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v5, v8, v9 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v4, v2 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v2, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, -1.0, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, -1.0, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_sub_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_sub_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v6 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, -1.0, v3 +; GFX8-IEEE-NEXT: v_add_f32_e32 v2, -1.0, v2 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX8-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX8-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX8-IEEE-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 @@ -4369,25 +4235,22 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX8-FLUSH-NEXT: v_sqrt_f16_e32 v0, s0 ; GFX8-FLUSH-NEXT: s_lshr_b32 s0, s0, 16 ; GFX8-FLUSH-NEXT: v_sqrt_f16_e32 v1, s0 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v1 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v2 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v2, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v7, v8, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v8, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v9, v8, v6 -; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v9, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v6 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v5 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v2, v4, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v6, v4, -v4 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v3, v5, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v7, v5, -v5 +; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, -1.0 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v9 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX8-FLUSH-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 @@ -4402,25 +4265,22 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX9-IEEE-NEXT: v_sqrt_f16_e32 v0, s0 ; GFX9-IEEE-NEXT: s_lshr_b32 s0, s0, 16 ; GFX9-IEEE-NEXT: v_sqrt_f16_e32 v1, s0 -; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v1 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v2 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX9-IEEE-NEXT: v_fma_f32 v8, -v2, v7, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX9-IEEE-NEXT: v_fma_f32 v7, v8, v5, v7 -; GFX9-IEEE-NEXT: v_fma_f32 v8, -v3, v9, v4 -; GFX9-IEEE-NEXT: v_fma_f32 v8, v8, v6, v9 -; GFX9-IEEE-NEXT: v_fma_f32 v2, -v2, v7, v4 -; GFX9-IEEE-NEXT: v_fma_f32 v3, -v3, v8, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v2, v2, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v4, v2 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX9-IEEE-NEXT: v_fma_f32 v6, v2, v4, -1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v6, v6, v4, -v4 +; GFX9-IEEE-NEXT: v_fma_f32 v7, v3, v5, -1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v7, v7, v5, -v5 +; GFX9-IEEE-NEXT: v_fma_f32 v2, -v2, v6, -1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v3, -v3, v7, -1.0 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX9-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX9-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX9-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-IEEE-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 @@ -4434,25 +4294,23 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX9-FLUSH-NEXT: v_sqrt_f16_e32 v0, s0 ; GFX9-FLUSH-NEXT: s_lshr_b32 s0, s0, 16 ; GFX9-FLUSH-NEXT: v_sqrt_f16_e32 v1, s0 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 +; GFX9-FLUSH-NEXT: v_mov_b32_e32 v4, -1.0 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v3, v3 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v5, v6, v2 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v1, v4, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v4, v6, v3 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v1, v4, -1.0 op_sel_hi:[1,0,1] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, v0, v2, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v5, v5, v2, -v2 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, v1, v3, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v6, v6, v3, -v3 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v4, -v1, v6, v4 op_sel_hi:[1,0,0] ; GFX9-FLUSH-NEXT: v_mul_f32_e32 v2, v7, v2 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v6, v3 +; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX9-FLUSH-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-FLUSH-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 @@ -4466,25 +4324,23 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX10-IEEE-NEXT: s_lshr_b32 s1, s0, 16 ; GFX10-IEEE-NEXT: v_sqrt_f16_e32 v0, s0 ; GFX10-IEEE-NEXT: v_sqrt_f16_e32 v1, s1 -; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 +; GFX10-IEEE-NEXT: v_mov_b32_e32 v4, -1.0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v3, v3 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v1, v4, -1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX10-IEEE-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v1, v4, -1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v6, v2 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v7, v3 +; GFX10-IEEE-NEXT: v_fma_mix_f32 v5, v0, v2, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, v1, v3, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_f32 v5, v5, v2, -v2 +; GFX10-IEEE-NEXT: v_fma_f32 v6, v6, v3, -v3 +; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_mix_f32 v4, -v1, v6, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v7, v2 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX10-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-IEEE-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 @@ -4498,25 +4354,22 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX10-FLUSH-NEXT: s_lshr_b32 s1, s0, 16 ; GFX10-FLUSH-NEXT: v_sqrt_f16_e32 v0, s0 ; GFX10-FLUSH-NEXT: v_sqrt_f16_e32 v1, s1 -; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v6, -1.0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v6 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v4 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v6 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v2, v4, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v3, v5, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v6, v4, -v4 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v7, v5, -v5 +; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, -1.0 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-FLUSH-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 @@ -4530,29 +4383,27 @@ define amdgpu_ps i32 @s_rsq_v2f16(i32 inreg %a.arg) { ; GFX11-NEXT: s_lshr_b32 s1, s0, 16 ; GFX11-NEXT: v_sqrt_f16_e32 v0, s0 ; GFX11-NEXT: v_sqrt_f16_e32 v1, s1 -; GFX11-NEXT: v_cvt_f32_f16_e32 v4, -1.0 +; GFX11-NEXT: v_mov_b32_e32 v4, -1.0 ; GFX11-NEXT: s_waitcnt_depctr 0xfff ; GFX11-NEXT: v_cvt_f32_f16_e32 v2, v0 ; GFX11-NEXT: v_cvt_f32_f16_e32 v3, v1 ; GFX11-NEXT: v_rcp_f32_e32 v2, v2 ; GFX11-NEXT: v_rcp_f32_e32 v3, v3 ; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX11-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v4, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v4, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX11-NEXT: v_mul_f32_e32 v3, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_dual_mul_f32 v2, v6, v2 :: v_dual_and_b32 v3, 0xff800000, v3 -; GFX11-NEXT: v_dual_add_f32 v3, v3, v4 :: v_dual_and_b32 v2, 0xff800000, v2 -; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 -; GFX11-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 +; GFX11-NEXT: v_fma_mix_f32 v5, v0, v2, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v6, v1, v3, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v5, v5, v2, -v2 +; GFX11-NEXT: v_fma_f32 v6, v6, v3, -v3 +; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v4, -v1, v6, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_dual_mul_f32 v2, v7, v2 :: v_dual_mul_f32 v3, v4, v3 +; GFX11-NEXT: v_and_b32_e32 v2, 0xff800000, v2 +; GFX11-NEXT: v_dual_add_f32 v2, v2, v5 :: v_dual_and_b32 v3, 0xff800000, v3 +; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX11-NEXT: v_cvt_f16_f32_e32 v2, v2 +; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 +; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 ; GFX11-NEXT: v_pack_b32_f16 v0, v0, v1 ; GFX11-NEXT: v_readfirstlane_b32 s0, v0 ; GFX11-NEXT: ; return to shader part epilog @@ -4568,21 +4419,20 @@ define half @v_rsq_f16(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -4590,24 +4440,23 @@ define half @v_rsq_f16(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -4632,21 +4481,20 @@ define half @v_neg_rsq_f16(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -4654,24 +4502,23 @@ define half @v_neg_rsq_f16(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -4706,21 +4553,20 @@ define { half, half } @v_rsq_f16_multi_use(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v2, v2, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, vcc, v1, v2, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v6, v5 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v3, v2, v1 +; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -4728,24 +4574,23 @@ define { half, half } @v_rsq_f16_multi_use(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v2, v2, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v1, v2, v1 +; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v0 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v3, v2, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -4785,21 +4630,20 @@ define half @v_rsq_f16_missing_contract0(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -4807,24 +4651,23 @@ define half @v_rsq_f16_missing_contract0(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -4859,21 +4702,20 @@ define half @v_rsq_f16_missing_contract1(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -4881,24 +4723,23 @@ define half @v_rsq_f16_missing_contract1(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -4933,21 +4774,20 @@ define half @v_neg_rsq_f16_missing_contract0(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -4955,24 +4795,23 @@ define half @v_neg_rsq_f16_missing_contract0(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -5007,21 +4846,20 @@ define half @v_neg_rsq_f16_missing_contract1(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -5029,24 +4867,23 @@ define half @v_neg_rsq_f16_missing_contract1(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -5081,21 +4918,20 @@ define half @v_neg_rsq_f16_fabs(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e64 v0, |v0| -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -5103,24 +4939,23 @@ define half @v_neg_rsq_f16_fabs(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e64 v0, |v0| -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -5156,21 +4991,20 @@ define half @v_rsq_f16_arcp(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -5178,24 +5012,23 @@ define half @v_rsq_f16_arcp(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -5220,21 +5053,20 @@ define half @v_neg_rsq_f16_arcp(half %a) { ; GFX6-IEEE: ; %bb.0: ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v5, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-IEEE-NEXT: v_fma_f32 v1, -v1, v4, v3 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] ; @@ -5242,24 +5074,23 @@ define half @v_neg_rsq_f16_arcp(half %a) { ; GFX6-FLUSH: ; %bb.0: ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, -1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, v1 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, v1, v0, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v2, v1 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, v4, v2, v2 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v4, v3, v2 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v1, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v4, v5, v2, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v1, -v1, v4, v3 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, v1 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v1, v0, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -5294,12 +5125,10 @@ define half @v_rsq_f16_afn(half %a) { ; GFX6: ; %bb.0: ; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-NEXT: v_rcp_f32_e32 v0, v0 -; GFX6-NEXT: v_mul_f32_e32 v0, v1, v0 ; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-NEXT: s_setpc_b64 s[30:31] ; @@ -5324,12 +5153,10 @@ define half @v_rsq_f16_afn_nocontract(half %a) { ; GFX6: ; %bb.0: ; GFX6-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GFX6-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-NEXT: v_rcp_f32_e32 v0, v0 -; GFX6-NEXT: v_mul_f32_e32 v0, v1, v0 ; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-NEXT: s_setpc_b64 s[30:31] ; @@ -5365,36 +5192,35 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v2, v0, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, s[4:5], v1, v1, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v9, -v3, v6, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v9, v6, v6 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v8, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v10, -v3, v9, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v9, v10, v6, v9 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v9, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v6, v9 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v3, v0, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v5, v8, 1.0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v7, s[4:5], v2, v1, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v3, v8, v8 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v7, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v5, v4, v7 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v3, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v5, v4, v7 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, 1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v5, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, s[4:5], v1, v1, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v8, -v2, v5, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v8, v5, v5 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v6, v4 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v8, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v10, -v2, v8, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v8, v10, v5, v8 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v8, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v9, -v4, v6, 1.0 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v5, v8 +; GFX6-IEEE-NEXT: v_div_scale_f32 v7, s[4:5], 1.0, v1, 1.0 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v9, v6, v6 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v3, v7, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v4, v3, v7 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v2, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v4, v3, v7 ; GFX6-IEEE-NEXT: s_mov_b64 vcc, s[4:5] -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v5, v3, v4 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v3, v1, v2 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v4, v2, v3 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] @@ -5404,42 +5230,40 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, 1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v0, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v0, v2 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, 1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v4, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, 1.0, v1, 1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v2, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, 1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -5448,31 +5272,28 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX8-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-IEEE-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX8-IEEE-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v0 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v2 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v4, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v8 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v9, -v2, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v10, v10, v4 -; GFX8-IEEE-NEXT: v_add_f32_e32 v9, v9, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v10, v10, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v9, v9, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v8, v10, v8 -; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v9, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v8 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v2, v2, v5 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v4, v2 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v6, -v2, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v7, -v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, 1.0, v7 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, 1.0, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v6 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, 1.0, v3 +; GFX8-IEEE-NEXT: v_add_f32_e32 v2, 1.0, v2 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX8-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX8-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 -; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX8-IEEE-NEXT: v_div_fixup_f16 v0, v3, v0, 1.0 @@ -5486,25 +5307,22 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX8-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-FLUSH-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v0 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v2 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v8, v4, v6 -; GFX8-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v4 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v6 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v6 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v5 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, -v2, v4, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, -v3, v5, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v7, v5, v5 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v6, v4, v4 +; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, 1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, 1.0 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX8-FLUSH-NEXT: v_div_fixup_f16 v0, v3, v0, 1.0 @@ -5518,25 +5336,22 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX9-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-IEEE-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX9-IEEE-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v0 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v2 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v8, v4, v6 -; GFX9-IEEE-NEXT: v_fma_f32 v9, -v2, v7, v4 -; GFX9-IEEE-NEXT: v_fma_f32 v10, -v3, v8, v4 -; GFX9-IEEE-NEXT: v_fma_f32 v7, v9, v5, v7 -; GFX9-IEEE-NEXT: v_fma_f32 v8, v10, v6, v8 -; GFX9-IEEE-NEXT: v_fma_f32 v2, -v2, v7, v4 -; GFX9-IEEE-NEXT: v_fma_f32 v3, -v3, v8, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v2, v2, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v4, v2 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX9-IEEE-NEXT: v_fma_f32 v6, -v2, v4, 1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v7, -v3, v5, 1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v6, v6, v4, v4 +; GFX9-IEEE-NEXT: v_fma_f32 v7, v7, v5, v5 +; GFX9-IEEE-NEXT: v_fma_f32 v2, -v2, v6, 1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v3, -v3, v7, 1.0 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX9-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX9-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX9-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-IEEE-NEXT: v_div_fixup_f16 v1, v2, v1, 1.0 @@ -5549,25 +5364,23 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX9-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-FLUSH-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX9-FLUSH-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, 1.0 +; GFX9-FLUSH-NEXT: v_mov_b32_e32 v4, 1.0 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v0 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v3, v3 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v1, v5, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v4, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v5, v6, v2 -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v4, v7, v3 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v1, v5, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v4, 1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v2, v6, v2 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v7, v3 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, -v1, v2, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v0, v3, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v5, v5, v2, v2 +; GFX9-FLUSH-NEXT: v_mad_f32 v6, v6, v3, v3 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v1, v5, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v4, -v0, v6, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mul_f32_e32 v2, v7, v2 +; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX9-FLUSH-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-FLUSH-NEXT: v_div_fixup_f16 v1, v2, v1, 1.0 @@ -5580,25 +5393,23 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX10-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-IEEE-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX10-IEEE-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v4, 1.0 +; GFX10-IEEE-NEXT: v_mov_b32_e32 v4, 1.0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v0 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v3, v3 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, -v1, v5, 1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v0, v4, 1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX10-IEEE-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, -v1, v5, 1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v0, v4, 1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v6, v2 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v7, v3 +; GFX10-IEEE-NEXT: v_fma_mix_f32 v5, -v1, v2, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, -v0, v3, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_f32 v5, v5, v2, v2 +; GFX10-IEEE-NEXT: v_fma_f32 v6, v6, v3, v3 +; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v1, v5, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_mix_f32 v4, -v0, v6, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v7, v2 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX10-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-IEEE-NEXT: v_div_fixup_f16 v1, v2, v1, 1.0 @@ -5611,25 +5422,22 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX10-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-FLUSH-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX10-FLUSH-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v6, 1.0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v0 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v6 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v4 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v6 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, -v2, v4, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, -v3, v5, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v6, v4, v4 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v7, v5, v5 +; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, 1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, 1.0 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-FLUSH-NEXT: v_div_fixup_f16 v1, v2, v1, 1.0 @@ -5642,7 +5450,7 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX11-NEXT: v_sqrt_f16_e32 v0, v0 -; GFX11-NEXT: v_cvt_f32_f16_e32 v4, 1.0 +; GFX11-NEXT: v_mov_b32_e32 v4, 1.0 ; GFX11-NEXT: v_sqrt_f16_e32 v1, v1 ; GFX11-NEXT: s_waitcnt_depctr 0xfff ; GFX11-NEXT: v_cvt_f32_f16_e32 v2, v0 @@ -5650,22 +5458,20 @@ define <2 x half> @v_rsq_v2f16(<2 x half> %a) { ; GFX11-NEXT: v_rcp_f32_e32 v2, v2 ; GFX11-NEXT: v_rcp_f32_e32 v3, v3 ; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX11-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v4, 1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v4, 1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX11-NEXT: v_mul_f32_e32 v3, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, 1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_dual_mul_f32 v2, v6, v2 :: v_dual_and_b32 v3, 0xff800000, v3 -; GFX11-NEXT: v_dual_add_f32 v3, v3, v4 :: v_dual_and_b32 v2, 0xff800000, v2 -; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 -; GFX11-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, 1.0 +; GFX11-NEXT: v_fma_mix_f32 v5, -v0, v2, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v6, -v1, v3, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v5, v5, v2, v2 +; GFX11-NEXT: v_fma_f32 v6, v6, v3, v3 +; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v4, -v1, v6, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_dual_mul_f32 v2, v7, v2 :: v_dual_mul_f32 v3, v4, v3 +; GFX11-NEXT: v_and_b32_e32 v2, 0xff800000, v2 +; GFX11-NEXT: v_dual_add_f32 v2, v2, v5 :: v_dual_and_b32 v3, 0xff800000, v3 +; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX11-NEXT: v_cvt_f16_f32_e32 v2, v2 +; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v0, v2, v0, 1.0 +; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, 1.0 ; GFX11-NEXT: v_pack_b32_f16 v0, v0, v1 ; GFX11-NEXT: s_setpc_b64 s[30:31] %sqrt = call contract <2 x half> @llvm.sqrt.v2f16(<2 x half> %a) @@ -5679,36 +5485,35 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX6-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v2, -1.0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_sqrt_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-IEEE-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX6-IEEE-NEXT: v_div_scale_f32 v4, vcc, v2, v0, v2 -; GFX6-IEEE-NEXT: v_div_scale_f32 v5, s[4:5], v1, v1, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v9, -v3, v6, 1.0 -; GFX6-IEEE-NEXT: v_fma_f32 v6, v9, v6, v6 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v9, v4, v6 -; GFX6-IEEE-NEXT: v_rcp_f32_e32 v8, v5 -; GFX6-IEEE-NEXT: v_fma_f32 v10, -v3, v9, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v9, v10, v6, v9 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v3, v9, v4 -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v3, v6, v9 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v3, v0, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v3, -v5, v8, 1.0 -; GFX6-IEEE-NEXT: v_div_scale_f32 v7, s[4:5], v2, v1, v2 -; GFX6-IEEE-NEXT: v_fma_f32 v3, v3, v8, v8 -; GFX6-IEEE-NEXT: v_mul_f32_e32 v4, v7, v3 -; GFX6-IEEE-NEXT: v_fma_f32 v6, -v5, v4, v7 -; GFX6-IEEE-NEXT: v_fma_f32 v4, v6, v3, v4 -; GFX6-IEEE-NEXT: v_fma_f32 v5, -v5, v4, v7 +; GFX6-IEEE-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, -1.0 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v5, v2 +; GFX6-IEEE-NEXT: v_div_scale_f32 v3, vcc, -1.0, v0, -1.0 +; GFX6-IEEE-NEXT: v_div_scale_f32 v4, s[4:5], v1, v1, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v8, -v2, v5, 1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v5, v8, v5, v5 +; GFX6-IEEE-NEXT: v_rcp_f32_e32 v6, v4 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v8, v3, v5 +; GFX6-IEEE-NEXT: v_fma_f32 v10, -v2, v8, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v8, v10, v5, v8 +; GFX6-IEEE-NEXT: v_fma_f32 v2, -v2, v8, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v9, -v4, v6, 1.0 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v2, v5, v8 +; GFX6-IEEE-NEXT: v_div_scale_f32 v7, s[4:5], -1.0, v1, -1.0 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v0, v2, v0, -1.0 +; GFX6-IEEE-NEXT: v_fma_f32 v2, v9, v6, v6 +; GFX6-IEEE-NEXT: v_mul_f32_e32 v3, v7, v2 +; GFX6-IEEE-NEXT: v_fma_f32 v5, -v4, v3, v7 +; GFX6-IEEE-NEXT: v_fma_f32 v3, v5, v2, v3 +; GFX6-IEEE-NEXT: v_fma_f32 v4, -v4, v3, v7 ; GFX6-IEEE-NEXT: s_mov_b64 vcc, s[4:5] -; GFX6-IEEE-NEXT: v_div_fmas_f32 v3, v5, v3, v4 -; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v3, v1, v2 +; GFX6-IEEE-NEXT: v_div_fmas_f32 v2, v4, v2, v3 +; GFX6-IEEE-NEXT: v_div_fixup_f32 v1, v2, v1, -1.0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-IEEE-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-IEEE-NEXT: s_setpc_b64 s[30:31] @@ -5718,42 +5523,40 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX6-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, -1.0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_sqrt_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 2, 2), 0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v3, s[4:5], v0, v0, v2 -; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v4, v3 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v2, v0, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v0, v0, -1.0 +; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, -1.0, v0, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v3, v4, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v4, v6, v4, v4 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v4 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v3, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v4, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, -v3, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v3, v3, v4, v6 -; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v3, v0, v2 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v0, v2, v0, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v2, s[4:5], v1, v1, -1.0 ; GFX6-FLUSH-NEXT: v_rcp_f32_e32 v3, v2 -; GFX6-FLUSH-NEXT: v_div_scale_f32 v5, vcc, v4, v1, v4 +; GFX6-FLUSH-NEXT: v_div_scale_f32 v4, vcc, -1.0, v1, -1.0 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v3, 1.0 -; GFX6-FLUSH-NEXT: v_fma_f32 v3, v6, v3, v3 -; GFX6-FLUSH-NEXT: v_mul_f32_e32 v6, v5, v3 -; GFX6-FLUSH-NEXT: v_fma_f32 v7, -v2, v6, v5 -; GFX6-FLUSH-NEXT: v_fma_f32 v6, v7, v3, v6 -; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v6, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, -v2, v3, 1.0 +; GFX6-FLUSH-NEXT: v_fma_f32 v3, v5, v3, v3 +; GFX6-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v3 +; GFX6-FLUSH-NEXT: v_fma_f32 v6, -v2, v5, v4 +; GFX6-FLUSH-NEXT: v_fma_f32 v5, v6, v3, v5 +; GFX6-FLUSH-NEXT: v_fma_f32 v2, -v2, v5, v4 ; GFX6-FLUSH-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v6 -; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, v4 +; GFX6-FLUSH-NEXT: v_div_fmas_f32 v2, v2, v3, v5 +; GFX6-FLUSH-NEXT: v_div_fixup_f32 v1, v2, v1, -1.0 ; GFX6-FLUSH-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GFX6-FLUSH-NEXT: s_setpc_b64 s[30:31] ; @@ -5762,31 +5565,28 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX8-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-IEEE-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX8-IEEE-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX8-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v0 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v2 -; GFX8-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v8, v4, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v10, -v3, v8 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v9, -v2, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v10, v10, v4 -; GFX8-IEEE-NEXT: v_add_f32_e32 v9, v9, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v10, v10, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v9, v9, v5 -; GFX8-IEEE-NEXT: v_add_f32_e32 v8, v10, v8 -; GFX8-IEEE-NEXT: v_add_f32_e32 v7, v9, v7 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v8 -; GFX8-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v7 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 -; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v4 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 -; GFX8-IEEE-NEXT: v_mul_f32_e32 v2, v2, v5 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v4, v2 +; GFX8-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v2, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v3, v5 +; GFX8-IEEE-NEXT: v_add_f32_e32 v7, -1.0, v7 +; GFX8-IEEE-NEXT: v_add_f32_e32 v6, -1.0, v6 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_sub_f32_e32 v7, v7, v5 +; GFX8-IEEE-NEXT: v_sub_f32_e32 v6, v6, v4 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v3, -v3, v7 +; GFX8-IEEE-NEXT: v_mul_f32_e64 v2, -v2, v6 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, -1.0, v3 +; GFX8-IEEE-NEXT: v_add_f32_e32 v2, -1.0, v2 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 +; GFX8-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX8-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX8-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 -; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 -; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 +; GFX8-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX8-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX8-IEEE-NEXT: v_div_fixup_f16 v0, v3, v0, -1.0 @@ -5800,25 +5600,22 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX8-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX8-FLUSH-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX8-FLUSH-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX8-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v0 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v2 -; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v6, v3 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v8, v4, v6 -; GFX8-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v4 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v6 -; GFX8-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v5 -; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v4 -; GFX8-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v4 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v6 -; GFX8-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v5 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 +; GFX8-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v2, v4, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v3, v5, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v7, v7, v5, -v5 +; GFX8-FLUSH-NEXT: v_mad_f32 v6, v6, v4, -v4 +; GFX8-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, -1.0 +; GFX8-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, -1.0 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 +; GFX8-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX8-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 -; GFX8-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 +; GFX8-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX8-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX8-FLUSH-NEXT: v_div_fixup_f16 v0, v3, v0, -1.0 @@ -5832,25 +5629,22 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX9-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-IEEE-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX9-IEEE-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX9-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v0 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v2 -; GFX9-IEEE-NEXT: v_rcp_f32_e32 v6, v3 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v7, v4, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v8, v4, v6 -; GFX9-IEEE-NEXT: v_fma_f32 v9, -v2, v7, v4 -; GFX9-IEEE-NEXT: v_fma_f32 v10, -v3, v8, v4 -; GFX9-IEEE-NEXT: v_fma_f32 v7, v9, v5, v7 -; GFX9-IEEE-NEXT: v_fma_f32 v8, v10, v6, v8 -; GFX9-IEEE-NEXT: v_fma_f32 v2, -v2, v7, v4 -; GFX9-IEEE-NEXT: v_fma_f32 v3, -v3, v8, v4 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v2, v2, v5 -; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v6 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v4, v2 +; GFX9-IEEE-NEXT: v_rcp_f32_e32 v5, v3 +; GFX9-IEEE-NEXT: v_fma_f32 v6, v2, v4, -1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v7, v3, v5, -1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v6, v6, v4, -v4 +; GFX9-IEEE-NEXT: v_fma_f32 v7, v7, v5, -v5 +; GFX9-IEEE-NEXT: v_fma_f32 v2, -v2, v6, -1.0 +; GFX9-IEEE-NEXT: v_fma_f32 v3, -v3, v7, -1.0 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v2, v2, v4 +; GFX9-IEEE-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX9-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX9-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX9-IEEE-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX9-IEEE-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX9-IEEE-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX9-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-IEEE-NEXT: v_div_fixup_f16 v1, v2, v1, -1.0 @@ -5863,25 +5657,23 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX9-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-FLUSH-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX9-FLUSH-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v4, -1.0 +; GFX9-FLUSH-NEXT: v_mov_b32_e32 v4, -1.0 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX9-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v0 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_rcp_f32_e32 v3, v3 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v1, v5, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v4, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v5, v6, v2 -; GFX9-FLUSH-NEXT: v_mac_f32_e32 v4, v7, v3 -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, -v1, v5, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v0, v4, -1.0 op_sel_hi:[1,0,1] -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v2, v6, v2 -; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v7, v3 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v5, v1, v2, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v6, v0, v3, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_f32 v5, v5, v2, -v2 +; GFX9-FLUSH-NEXT: v_mad_f32 v6, v6, v3, -v3 +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v7, -v1, v5, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mad_mix_f32 v4, -v0, v6, v4 op_sel_hi:[1,0,0] +; GFX9-FLUSH-NEXT: v_mul_f32_e32 v2, v7, v2 +; GFX9-FLUSH-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX9-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX9-FLUSH-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX9-FLUSH-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX9-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX9-FLUSH-NEXT: v_div_fixup_f16 v1, v2, v1, -1.0 @@ -5894,25 +5686,23 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX10-IEEE-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-IEEE-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX10-IEEE-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v4, -1.0 +; GFX10-IEEE-NEXT: v_mov_b32_e32 v4, -1.0 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX10-IEEE-NEXT: v_cvt_f32_f16_e32 v3, v0 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_rcp_f32_e32 v3, v3 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, -v1, v5, -1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v0, v4, -1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX10-IEEE-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, -v1, v5, -1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v0, v4, -1.0 op_sel_hi:[1,0,1] -; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v6, v2 -; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v7, v3 +; GFX10-IEEE-NEXT: v_fma_mix_f32 v5, v1, v2, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_mix_f32 v6, v0, v3, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_f32 v5, v5, v2, -v2 +; GFX10-IEEE-NEXT: v_fma_f32 v6, v6, v3, -v3 +; GFX10-IEEE-NEXT: v_fma_mix_f32 v7, -v1, v5, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_fma_mix_f32 v4, -v0, v6, v4 op_sel_hi:[1,0,0] +; GFX10-IEEE-NEXT: v_mul_f32_e32 v2, v7, v2 +; GFX10-IEEE-NEXT: v_mul_f32_e32 v3, v4, v3 ; GFX10-IEEE-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-IEEE-NEXT: v_and_b32_e32 v3, 0xff800000, v3 ; GFX10-IEEE-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v4 +; GFX10-IEEE-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-IEEE-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-IEEE-NEXT: v_div_fixup_f16 v1, v2, v1, -1.0 @@ -5925,25 +5715,22 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX10-FLUSH-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-FLUSH-NEXT: v_sqrt_f16_e32 v1, v0 ; GFX10-FLUSH-NEXT: v_sqrt_f16_sdwa v0, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 -; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v6, -1.0 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v2, v1 ; GFX10-FLUSH-NEXT: v_cvt_f32_f16_e32 v3, v0 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v4, v2 ; GFX10-FLUSH-NEXT: v_rcp_f32_e32 v5, v3 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v7, v6, v4 -; GFX10-FLUSH-NEXT: v_mul_f32_e32 v8, v6, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v9, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v10, -v3, v8, v6 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v7, v9, v4 -; GFX10-FLUSH-NEXT: v_mac_f32_e32 v8, v10, v5 -; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v7, v6 -; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v8, v6 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v2, v4, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v3, v5, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v6, v6, v4, -v4 +; GFX10-FLUSH-NEXT: v_mad_f32 v7, v7, v5, -v5 +; GFX10-FLUSH-NEXT: v_mad_f32 v2, -v2, v6, -1.0 +; GFX10-FLUSH-NEXT: v_mad_f32 v3, -v3, v7, -1.0 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v2, v2, v4 ; GFX10-FLUSH-NEXT: v_mul_f32_e32 v3, v3, v5 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v2, 0xff800000, v2 ; GFX10-FLUSH-NEXT: v_and_b32_e32 v3, 0xff800000, v3 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v7 -; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v8 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v2, v2, v6 +; GFX10-FLUSH-NEXT: v_add_f32_e32 v3, v3, v7 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GFX10-FLUSH-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX10-FLUSH-NEXT: v_div_fixup_f16 v1, v2, v1, -1.0 @@ -5956,7 +5743,7 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GFX11-NEXT: v_sqrt_f16_e32 v0, v0 -; GFX11-NEXT: v_cvt_f32_f16_e32 v4, -1.0 +; GFX11-NEXT: v_mov_b32_e32 v4, -1.0 ; GFX11-NEXT: v_sqrt_f16_e32 v1, v1 ; GFX11-NEXT: s_waitcnt_depctr 0xfff ; GFX11-NEXT: v_cvt_f32_f16_e32 v2, v0 @@ -5964,22 +5751,20 @@ define <2 x half> @v_neg_rsq_v2f16(<2 x half> %a) { ; GFX11-NEXT: v_rcp_f32_e32 v2, v2 ; GFX11-NEXT: v_rcp_f32_e32 v3, v3 ; GFX11-NEXT: s_waitcnt_depctr 0xfff -; GFX11-NEXT: v_mul_f32_e32 v5, v4, v2 -; GFX11-NEXT: v_mul_f32_e32 v4, v4, v3 -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v4, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v4, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fma_mix_f32 v7, -v1, v4, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_fmac_f32_e32 v5, v6, v2 -; GFX11-NEXT: v_mul_f32_e32 v3, v7, v3 -; GFX11-NEXT: v_fma_mix_f32 v6, -v0, v5, -1.0 op_sel_hi:[1,0,1] -; GFX11-NEXT: v_dual_mul_f32 v2, v6, v2 :: v_dual_and_b32 v3, 0xff800000, v3 -; GFX11-NEXT: v_dual_add_f32 v3, v3, v4 :: v_dual_and_b32 v2, 0xff800000, v2 -; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 -; GFX11-NEXT: v_add_f32_e32 v2, v2, v5 -; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 +; GFX11-NEXT: v_fma_mix_f32 v5, v0, v2, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v6, v1, v3, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_f32 v5, v5, v2, -v2 +; GFX11-NEXT: v_fma_f32 v6, v6, v3, -v3 +; GFX11-NEXT: v_fma_mix_f32 v7, -v0, v5, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_fma_mix_f32 v4, -v1, v6, v4 op_sel_hi:[1,0,0] +; GFX11-NEXT: v_dual_mul_f32 v2, v7, v2 :: v_dual_mul_f32 v3, v4, v3 +; GFX11-NEXT: v_and_b32_e32 v2, 0xff800000, v2 +; GFX11-NEXT: v_dual_add_f32 v2, v2, v5 :: v_dual_and_b32 v3, 0xff800000, v3 +; GFX11-NEXT: v_add_f32_e32 v3, v3, v6 ; GFX11-NEXT: v_cvt_f16_f32_e32 v2, v2 +; GFX11-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GFX11-NEXT: v_div_fixup_f16 v0, v2, v0, -1.0 +; GFX11-NEXT: v_div_fixup_f16 v1, v3, v1, -1.0 ; GFX11-NEXT: v_pack_b32_f16 v0, v0, v1 ; GFX11-NEXT: s_setpc_b64 s[30:31] %sqrt = call contract <2 x half> @llvm.sqrt.v2f16(<2 x half> %a) diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/flat-scratch.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/flat-scratch.ll index e6a8baceee02..2356dad5275c 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/flat-scratch.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/flat-scratch.ll @@ -1,15 +1,15 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -global-isel -mattr=-promote-alloca -mattr=+enable-flat-scratch < %s | FileCheck -check-prefix=GFX9 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -global-isel -mattr=-promote-alloca -mattr=+enable-flat-scratch < %s | FileCheck -check-prefix=GFX10 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -global-isel -mattr=-promote-alloca < %s | FileCheck -check-prefix=GFX942 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -global-isel -mattr=-promote-alloca < %s | FileCheck -check-prefix=GFX11 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -global-isel -mattr=-promote-alloca < %s | FileCheck -check-prefix=GFX12 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -global-isel -new-reg-bank-select -mattr=-promote-alloca -mattr=+enable-flat-scratch < %s | FileCheck -check-prefix=GFX9 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -global-isel -new-reg-bank-select -mattr=-promote-alloca -mattr=+enable-flat-scratch < %s | FileCheck -check-prefix=GFX10 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -global-isel -new-reg-bank-select -mattr=-promote-alloca < %s | FileCheck -check-prefix=GFX942 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -global-isel -new-reg-bank-select -mattr=-promote-alloca < %s | FileCheck -check-prefix=GFX11 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -global-isel -new-reg-bank-select -mattr=-promote-alloca < %s | FileCheck -check-prefix=GFX12 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -global-isel -mattr=-unaligned-access-mode -mattr=-promote-alloca -mattr=+enable-flat-scratch < %s | FileCheck -check-prefixes=UNALIGNED_GFX9 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -global-isel -mattr=-unaligned-access-mode -mattr=-promote-alloca -mattr=+enable-flat-scratch < %s | FileCheck -check-prefixes=UNALIGNED_GFX10 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -global-isel -mattr=-unaligned-access-mode -mattr=-promote-alloca < %s | FileCheck -check-prefixes=UNALIGNED_GFX942 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -global-isel -mattr=-unaligned-access-mode -mattr=-promote-alloca < %s | FileCheck -check-prefixes=UNALIGNED_GFX11 %s -; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -global-isel -mattr=-unaligned-access-mode -mattr=-promote-alloca < %s | FileCheck -check-prefixes=UNALIGNED_GFX12 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -global-isel -new-reg-bank-select -mattr=-unaligned-access-mode -mattr=-promote-alloca -mattr=+enable-flat-scratch < %s | FileCheck -check-prefixes=UNALIGNED_GFX9 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1030 -global-isel -new-reg-bank-select -mattr=-unaligned-access-mode -mattr=-promote-alloca -mattr=+enable-flat-scratch < %s | FileCheck -check-prefixes=UNALIGNED_GFX10 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx942 -global-isel -new-reg-bank-select -mattr=-unaligned-access-mode -mattr=-promote-alloca < %s | FileCheck -check-prefixes=UNALIGNED_GFX942 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -global-isel -new-reg-bank-select -mattr=-unaligned-access-mode -mattr=-promote-alloca < %s | FileCheck -check-prefixes=UNALIGNED_GFX11 %s +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -global-isel -new-reg-bank-select -mattr=-unaligned-access-mode -mattr=-promote-alloca < %s | FileCheck -check-prefixes=UNALIGNED_GFX12 %s define amdgpu_kernel void @store_load_sindex_kernel(i32 %idx) { ; GFX9-LABEL: store_load_sindex_kernel: diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/frem.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/frem.ll index 302b2395642d..a43bfb5d4567 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/frem.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/frem.ll @@ -88,11 +88,10 @@ define amdgpu_kernel void @frem_f16(ptr addrspace(1) %out, ptr addrspace(1) %in1 ; CI-NEXT: v_or_b32_e32 v1, s4, v0 ; CI-NEXT: .LBB0_8: ; %Flow19 ; CI-NEXT: v_cvt_f32_f16_e32 v0, s3 -; CI-NEXT: v_cvt_f32_f16_e32 v2, 0 ; CI-NEXT: s_and_b32 s2, s2, 0x7fff ; CI-NEXT: s_cmpk_lg_i32 s2, 0x7c00 ; CI-NEXT: s_cselect_b32 s2, 1, 0 -; CI-NEXT: v_cmp_nlg_f32_e32 vcc, v0, v2 +; CI-NEXT: v_cmp_nlg_f32_e32 vcc, 0, v0 ; CI-NEXT: v_mov_b32_e32 v0, 0x7e00 ; CI-NEXT: s_and_b32 s2, 1, s2 ; CI-NEXT: v_cndmask_b32_e32 v1, v1, v0, vcc @@ -1048,7 +1047,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cvt_f32_f16_e64 v1, |s1| ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v2, v1 ; CI-NEXT: s_cbranch_vccz .LBB9_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else20 ; CI-NEXT: s_and_b32 s2, s0, 0x8000 ; CI-NEXT: v_cmp_eq_f32_e32 vcc, v2, v1 ; CI-NEXT: v_mov_b32_e32 v0, s2 @@ -1059,7 +1058,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s2, s2, 1 ; CI-NEXT: s_cmp_lg_u32 s2, 0 ; CI-NEXT: s_cbranch_scc1 .LBB9_8 -; CI-NEXT: ; %bb.3: ; %frem.compute +; CI-NEXT: ; %bb.3: ; %frem.compute19 ; CI-NEXT: v_frexp_mant_f32_e32 v3, v1 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v6, v1 ; CI-NEXT: v_ldexp_f32_e64 v1, v3, 1 @@ -1084,10 +1083,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v2 ; CI-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB9_6 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; CI-NEXT: v_add_i32_e32 v2, vcc, 11, v5 ; CI-NEXT: v_sub_i32_e32 v2, vcc, v2, v6 -; CI-NEXT: .LBB9_5: ; %frem.loop_body +; CI-NEXT: .LBB9_5: ; %frem.loop_body27 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v5, v4 ; CI-NEXT: v_mul_f32_e32 v4, v5, v3 @@ -1103,7 +1102,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB9_7 ; CI-NEXT: .LBB9_6: ; CI-NEXT: v_mov_b32_e32 v5, v4 -; CI-NEXT: .LBB9_7: ; %frem.loop_exit +; CI-NEXT: .LBB9_7: ; %frem.loop_exit28 ; CI-NEXT: v_add_i32_e32 v2, vcc, -10, v2 ; CI-NEXT: v_ldexp_f32_e32 v2, v5, v2 ; CI-NEXT: v_mul_f32_e32 v3, v2, v3 @@ -1126,7 +1125,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: ; implicit-def: $vgpr1 ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v3, v2 ; CI-NEXT: s_cbranch_vccz .LBB9_10 -; CI-NEXT: ; %bb.9: ; %frem.else20 +; CI-NEXT: ; %bb.9: ; %frem.else ; CI-NEXT: s_and_b32 s4, s2, 0x8000 ; CI-NEXT: v_cmp_eq_f32_e32 vcc, v3, v2 ; CI-NEXT: v_mov_b32_e32 v1, s4 @@ -1137,7 +1136,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s4, s4, 1 ; CI-NEXT: s_cmp_lg_u32 s4, 0 ; CI-NEXT: s_cbranch_scc1 .LBB9_16 -; CI-NEXT: ; %bb.11: ; %frem.compute19 +; CI-NEXT: ; %bb.11: ; %frem.compute ; CI-NEXT: v_frexp_mant_f32_e32 v4, v2 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v7, v2 ; CI-NEXT: v_ldexp_f32_e64 v2, v4, 1 @@ -1162,10 +1161,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v3 ; CI-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB9_14 -; CI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; CI-NEXT: v_add_i32_e32 v3, vcc, 11, v6 ; CI-NEXT: v_sub_i32_e32 v3, vcc, v3, v7 -; CI-NEXT: .LBB9_13: ; %frem.loop_body27 +; CI-NEXT: .LBB9_13: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v6, v5 ; CI-NEXT: v_mul_f32_e32 v5, v6, v4 @@ -1181,7 +1180,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB9_15 ; CI-NEXT: .LBB9_14: ; CI-NEXT: v_mov_b32_e32 v6, v5 -; CI-NEXT: .LBB9_15: ; %frem.loop_exit28 +; CI-NEXT: .LBB9_15: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v3, vcc, -10, v3 ; CI-NEXT: v_ldexp_f32_e32 v3, v6, v3 ; CI-NEXT: v_mul_f32_e32 v4, v3, v4 @@ -1197,16 +1196,15 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_or_b32_e32 v1, s4, v1 ; CI-NEXT: .LBB9_16: ; %Flow54 ; CI-NEXT: v_cvt_f32_f16_e32 v2, s1 -; CI-NEXT: v_cvt_f32_f16_e32 v3, 0 ; CI-NEXT: s_and_b32 s0, s0, 0x7fff ; CI-NEXT: s_cmpk_lg_i32 s0, 0x7c00 ; CI-NEXT: s_cselect_b32 s4, 1, 0 -; CI-NEXT: v_cmp_nlg_f32_e32 vcc, v2, v3 +; CI-NEXT: v_cmp_nlg_f32_e32 vcc, 0, v2 ; CI-NEXT: v_cvt_f32_f16_e32 v2, s3 ; CI-NEXT: s_and_b32 s2, s2, 0x7fff ; CI-NEXT: s_cmpk_lg_i32 s2, 0x7c00 ; CI-NEXT: s_cselect_b32 s2, 1, 0 -; CI-NEXT: v_cmp_nlg_f32_e64 s[0:1], v2, v3 +; CI-NEXT: v_cmp_nlg_f32_e64 s[0:1], 0, v2 ; CI-NEXT: v_and_b32_e32 v0, 0xffff, v0 ; CI-NEXT: v_mov_b32_e32 v2, 0x7e00 ; CI-NEXT: s_and_b32 s3, 1, s4 @@ -1239,7 +1237,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cvt_f32_f16_e64 v1, |s1| ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v2, v1 ; VI-NEXT: s_cbranch_vccz .LBB9_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else20 ; VI-NEXT: s_and_b32 s2, s0, 0x8000 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v2, v1 ; VI-NEXT: v_mov_b32_e32 v0, s2 @@ -1250,7 +1248,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s2, s2, 1 ; VI-NEXT: s_cmp_lg_u32 s2, 0 ; VI-NEXT: s_cbranch_scc1 .LBB9_8 -; VI-NEXT: ; %bb.3: ; %frem.compute +; VI-NEXT: ; %bb.3: ; %frem.compute19 ; VI-NEXT: v_frexp_mant_f32_e32 v3, v1 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v6, v1 ; VI-NEXT: v_ldexp_f32 v1, v3, 1 @@ -1275,10 +1273,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v2 ; VI-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB9_6 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; VI-NEXT: v_add_u32_e32 v2, vcc, 11, v5 ; VI-NEXT: v_sub_u32_e32 v2, vcc, v2, v6 -; VI-NEXT: .LBB9_5: ; %frem.loop_body +; VI-NEXT: .LBB9_5: ; %frem.loop_body27 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v5, v4 ; VI-NEXT: v_mul_f32_e32 v4, v5, v3 @@ -1294,7 +1292,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB9_7 ; VI-NEXT: .LBB9_6: ; VI-NEXT: v_mov_b32_e32 v5, v4 -; VI-NEXT: .LBB9_7: ; %frem.loop_exit +; VI-NEXT: .LBB9_7: ; %frem.loop_exit28 ; VI-NEXT: v_add_u32_e32 v2, vcc, -10, v2 ; VI-NEXT: v_ldexp_f32 v2, v5, v2 ; VI-NEXT: v_mul_f32_e32 v3, v2, v3 @@ -1317,7 +1315,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: ; implicit-def: $vgpr1 ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v3, v2 ; VI-NEXT: s_cbranch_vccz .LBB9_10 -; VI-NEXT: ; %bb.9: ; %frem.else20 +; VI-NEXT: ; %bb.9: ; %frem.else ; VI-NEXT: s_and_b32 s3, s4, 0x8000 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v3, v2 ; VI-NEXT: v_mov_b32_e32 v1, s3 @@ -1328,7 +1326,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s3, s3, 1 ; VI-NEXT: s_cmp_lg_u32 s3, 0 ; VI-NEXT: s_cbranch_scc1 .LBB9_16 -; VI-NEXT: ; %bb.11: ; %frem.compute19 +; VI-NEXT: ; %bb.11: ; %frem.compute ; VI-NEXT: v_frexp_mant_f32_e32 v4, v2 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v7, v2 ; VI-NEXT: v_ldexp_f32 v2, v4, 1 @@ -1353,10 +1351,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v3 ; VI-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB9_14 -; VI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; VI-NEXT: v_add_u32_e32 v3, vcc, 11, v6 ; VI-NEXT: v_sub_u32_e32 v3, vcc, v3, v7 -; VI-NEXT: .LBB9_13: ; %frem.loop_body27 +; VI-NEXT: .LBB9_13: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v6, v5 ; VI-NEXT: v_mul_f32_e32 v5, v6, v4 @@ -1372,7 +1370,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB9_15 ; VI-NEXT: .LBB9_14: ; VI-NEXT: v_mov_b32_e32 v6, v5 -; VI-NEXT: .LBB9_15: ; %frem.loop_exit28 +; VI-NEXT: .LBB9_15: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v3, vcc, -10, v3 ; VI-NEXT: v_ldexp_f32 v3, v6, v3 ; VI-NEXT: v_mul_f32_e32 v4, v3, v4 @@ -1427,7 +1425,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cvt_f32_f16_e64 v1, |s2| ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v2, v1 ; CI-NEXT: s_cbranch_vccz .LBB10_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else86 ; CI-NEXT: s_and_b32 s0, s4, 0x8000 ; CI-NEXT: v_cmp_eq_f32_e32 vcc, v2, v1 ; CI-NEXT: v_mov_b32_e32 v0, s0 @@ -1438,7 +1436,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s0, s0, 1 ; CI-NEXT: s_cmp_lg_u32 s0, 0 ; CI-NEXT: s_cbranch_scc1 .LBB10_8 -; CI-NEXT: ; %bb.3: ; %frem.compute +; CI-NEXT: ; %bb.3: ; %frem.compute85 ; CI-NEXT: v_frexp_mant_f32_e32 v3, v1 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v6, v1 ; CI-NEXT: v_ldexp_f32_e64 v1, v3, 1 @@ -1463,10 +1461,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v2 ; CI-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB10_6 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; CI-NEXT: v_add_i32_e32 v2, vcc, 11, v5 ; CI-NEXT: v_sub_i32_e32 v2, vcc, v2, v6 -; CI-NEXT: .LBB10_5: ; %frem.loop_body +; CI-NEXT: .LBB10_5: ; %frem.loop_body93 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v5, v4 ; CI-NEXT: v_mul_f32_e32 v4, v5, v3 @@ -1482,7 +1480,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB10_7 ; CI-NEXT: .LBB10_6: ; CI-NEXT: v_mov_b32_e32 v5, v4 -; CI-NEXT: .LBB10_7: ; %frem.loop_exit +; CI-NEXT: .LBB10_7: ; %frem.loop_exit94 ; CI-NEXT: v_add_i32_e32 v2, vcc, -10, v2 ; CI-NEXT: v_ldexp_f32_e32 v2, v5, v2 ; CI-NEXT: v_mul_f32_e32 v3, v2, v3 @@ -1505,7 +1503,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: ; implicit-def: $vgpr1 ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v3, v2 ; CI-NEXT: s_cbranch_vccz .LBB10_10 -; CI-NEXT: ; %bb.9: ; %frem.else20 +; CI-NEXT: ; %bb.9: ; %frem.else53 ; CI-NEXT: s_and_b32 s1, s6, 0x8000 ; CI-NEXT: v_cmp_eq_f32_e32 vcc, v3, v2 ; CI-NEXT: v_mov_b32_e32 v1, s1 @@ -1516,7 +1514,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s1, s1, 1 ; CI-NEXT: s_cmp_lg_u32 s1, 0 ; CI-NEXT: s_cbranch_scc1 .LBB10_16 -; CI-NEXT: ; %bb.11: ; %frem.compute19 +; CI-NEXT: ; %bb.11: ; %frem.compute52 ; CI-NEXT: v_frexp_mant_f32_e32 v4, v2 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v7, v2 ; CI-NEXT: v_ldexp_f32_e64 v2, v4, 1 @@ -1541,10 +1539,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v3 ; CI-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB10_14 -; CI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; CI-NEXT: v_add_i32_e32 v3, vcc, 11, v6 ; CI-NEXT: v_sub_i32_e32 v3, vcc, v3, v7 -; CI-NEXT: .LBB10_13: ; %frem.loop_body27 +; CI-NEXT: .LBB10_13: ; %frem.loop_body60 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v6, v5 ; CI-NEXT: v_mul_f32_e32 v5, v6, v4 @@ -1560,7 +1558,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB10_15 ; CI-NEXT: .LBB10_14: ; CI-NEXT: v_mov_b32_e32 v6, v5 -; CI-NEXT: .LBB10_15: ; %frem.loop_exit28 +; CI-NEXT: .LBB10_15: ; %frem.loop_exit61 ; CI-NEXT: v_add_i32_e32 v3, vcc, -10, v3 ; CI-NEXT: v_ldexp_f32_e32 v3, v6, v3 ; CI-NEXT: v_mul_f32_e32 v4, v3, v4 @@ -1581,7 +1579,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: ; implicit-def: $vgpr2 ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v4, v3 ; CI-NEXT: s_cbranch_vccz .LBB10_18 -; CI-NEXT: ; %bb.17: ; %frem.else53 +; CI-NEXT: ; %bb.17: ; %frem.else20 ; CI-NEXT: s_and_b32 s1, s5, 0x8000 ; CI-NEXT: v_cmp_eq_f32_e32 vcc, v4, v3 ; CI-NEXT: v_mov_b32_e32 v2, s1 @@ -1592,7 +1590,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s1, s1, 1 ; CI-NEXT: s_cmp_lg_u32 s1, 0 ; CI-NEXT: s_cbranch_scc1 .LBB10_24 -; CI-NEXT: ; %bb.19: ; %frem.compute52 +; CI-NEXT: ; %bb.19: ; %frem.compute19 ; CI-NEXT: v_frexp_mant_f32_e32 v5, v3 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v8, v3 ; CI-NEXT: v_ldexp_f32_e64 v3, v5, 1 @@ -1617,10 +1615,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v4 ; CI-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB10_22 -; CI-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; CI-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; CI-NEXT: v_add_i32_e32 v4, vcc, 11, v7 ; CI-NEXT: v_sub_i32_e32 v4, vcc, v4, v8 -; CI-NEXT: .LBB10_21: ; %frem.loop_body60 +; CI-NEXT: .LBB10_21: ; %frem.loop_body27 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v7, v6 ; CI-NEXT: v_mul_f32_e32 v6, v7, v5 @@ -1636,7 +1634,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB10_23 ; CI-NEXT: .LBB10_22: ; CI-NEXT: v_mov_b32_e32 v7, v6 -; CI-NEXT: .LBB10_23: ; %frem.loop_exit61 +; CI-NEXT: .LBB10_23: ; %frem.loop_exit28 ; CI-NEXT: v_add_i32_e32 v4, vcc, -10, v4 ; CI-NEXT: v_ldexp_f32_e32 v4, v7, v4 ; CI-NEXT: v_mul_f32_e32 v5, v4, v5 @@ -1659,7 +1657,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: ; implicit-def: $vgpr3 ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v5, v4 ; CI-NEXT: s_cbranch_vccz .LBB10_26 -; CI-NEXT: ; %bb.25: ; %frem.else86 +; CI-NEXT: ; %bb.25: ; %frem.else ; CI-NEXT: s_and_b32 s1, s7, 0x8000 ; CI-NEXT: v_cmp_eq_f32_e32 vcc, v5, v4 ; CI-NEXT: v_mov_b32_e32 v3, s1 @@ -1670,7 +1668,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s1, s1, 1 ; CI-NEXT: s_cmp_lg_u32 s1, 0 ; CI-NEXT: s_cbranch_scc1 .LBB10_32 -; CI-NEXT: ; %bb.27: ; %frem.compute85 +; CI-NEXT: ; %bb.27: ; %frem.compute ; CI-NEXT: v_frexp_mant_f32_e32 v6, v4 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v9, v4 ; CI-NEXT: v_ldexp_f32_e64 v4, v6, 1 @@ -1695,10 +1693,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v5 ; CI-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB10_30 -; CI-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; CI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; CI-NEXT: v_add_i32_e32 v5, vcc, 11, v8 ; CI-NEXT: v_sub_i32_e32 v5, vcc, v5, v9 -; CI-NEXT: .LBB10_29: ; %frem.loop_body93 +; CI-NEXT: .LBB10_29: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v8, v7 ; CI-NEXT: v_mul_f32_e32 v7, v8, v6 @@ -1714,7 +1712,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB10_31 ; CI-NEXT: .LBB10_30: ; CI-NEXT: v_mov_b32_e32 v8, v7 -; CI-NEXT: .LBB10_31: ; %frem.loop_exit94 +; CI-NEXT: .LBB10_31: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v5, vcc, -10, v5 ; CI-NEXT: v_ldexp_f32_e32 v5, v8, v5 ; CI-NEXT: v_mul_f32_e32 v6, v5, v6 @@ -1730,26 +1728,25 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_or_b32_e32 v3, s1, v3 ; CI-NEXT: .LBB10_32: ; %Flow124 ; CI-NEXT: v_cvt_f32_f16_e32 v4, s2 -; CI-NEXT: v_cvt_f32_f16_e32 v5, 0 ; CI-NEXT: s_and_b32 s1, s4, 0x7fff ; CI-NEXT: s_cmpk_lg_i32 s1, 0x7c00 ; CI-NEXT: s_cselect_b32 s11, 1, 0 -; CI-NEXT: v_cmp_nlg_f32_e32 vcc, v4, v5 +; CI-NEXT: v_cmp_nlg_f32_e32 vcc, 0, v4 ; CI-NEXT: v_cvt_f32_f16_e32 v4, s0 ; CI-NEXT: s_and_b32 s2, s6, 0x7fff ; CI-NEXT: s_cmpk_lg_i32 s2, 0x7c00 ; CI-NEXT: s_cselect_b32 s6, 1, 0 -; CI-NEXT: v_cmp_nlg_f32_e64 s[0:1], v4, v5 +; CI-NEXT: v_cmp_nlg_f32_e64 s[0:1], 0, v4 ; CI-NEXT: v_cvt_f32_f16_e32 v4, s3 ; CI-NEXT: s_and_b32 s4, s5, 0x7fff ; CI-NEXT: s_cmpk_lg_i32 s4, 0x7c00 ; CI-NEXT: s_cselect_b32 s12, 1, 0 -; CI-NEXT: v_cmp_nlg_f32_e64 s[2:3], v4, v5 +; CI-NEXT: v_cmp_nlg_f32_e64 s[2:3], 0, v4 ; CI-NEXT: v_cvt_f32_f16_e32 v4, s10 ; CI-NEXT: s_and_b32 s7, s7, 0x7fff ; CI-NEXT: s_cmpk_lg_i32 s7, 0x7c00 ; CI-NEXT: s_cselect_b32 s7, 1, 0 -; CI-NEXT: v_cmp_nlg_f32_e64 s[4:5], v4, v5 +; CI-NEXT: v_cmp_nlg_f32_e64 s[4:5], 0, v4 ; CI-NEXT: v_and_b32_e32 v0, 0xffff, v0 ; CI-NEXT: v_mov_b32_e32 v4, 0x7e00 ; CI-NEXT: s_and_b32 s10, 1, s11 @@ -1794,7 +1791,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cvt_f32_f16_e64 v1, |s6| ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v2, v1 ; VI-NEXT: s_cbranch_vccz .LBB10_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else86 ; VI-NEXT: s_and_b32 s0, s8, 0x8000 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v2, v1 ; VI-NEXT: v_mov_b32_e32 v0, s0 @@ -1805,7 +1802,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s0, s0, 1 ; VI-NEXT: s_cmp_lg_u32 s0, 0 ; VI-NEXT: s_cbranch_scc1 .LBB10_8 -; VI-NEXT: ; %bb.3: ; %frem.compute +; VI-NEXT: ; %bb.3: ; %frem.compute85 ; VI-NEXT: v_frexp_mant_f32_e32 v3, v1 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v6, v1 ; VI-NEXT: v_ldexp_f32 v1, v3, 1 @@ -1830,10 +1827,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v2 ; VI-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB10_6 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; VI-NEXT: v_add_u32_e32 v2, vcc, 11, v5 ; VI-NEXT: v_sub_u32_e32 v2, vcc, v2, v6 -; VI-NEXT: .LBB10_5: ; %frem.loop_body +; VI-NEXT: .LBB10_5: ; %frem.loop_body93 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v5, v4 ; VI-NEXT: v_mul_f32_e32 v4, v5, v3 @@ -1849,7 +1846,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB10_7 ; VI-NEXT: .LBB10_6: ; VI-NEXT: v_mov_b32_e32 v5, v4 -; VI-NEXT: .LBB10_7: ; %frem.loop_exit +; VI-NEXT: .LBB10_7: ; %frem.loop_exit94 ; VI-NEXT: v_add_u32_e32 v2, vcc, -10, v2 ; VI-NEXT: v_ldexp_f32 v2, v5, v2 ; VI-NEXT: v_mul_f32_e32 v3, v2, v3 @@ -1872,7 +1869,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: ; implicit-def: $vgpr1 ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v3, v2 ; VI-NEXT: s_cbranch_vccz .LBB10_10 -; VI-NEXT: ; %bb.9: ; %frem.else20 +; VI-NEXT: ; %bb.9: ; %frem.else53 ; VI-NEXT: s_and_b32 s0, s4, 0x8000 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v3, v2 ; VI-NEXT: v_mov_b32_e32 v1, s0 @@ -1883,7 +1880,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s0, s0, 1 ; VI-NEXT: s_cmp_lg_u32 s0, 0 ; VI-NEXT: s_cbranch_scc1 .LBB10_16 -; VI-NEXT: ; %bb.11: ; %frem.compute19 +; VI-NEXT: ; %bb.11: ; %frem.compute52 ; VI-NEXT: v_frexp_mant_f32_e32 v4, v2 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v7, v2 ; VI-NEXT: v_ldexp_f32 v2, v4, 1 @@ -1908,10 +1905,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v3 ; VI-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB10_14 -; VI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; VI-NEXT: v_add_u32_e32 v3, vcc, 11, v6 ; VI-NEXT: v_sub_u32_e32 v3, vcc, v3, v7 -; VI-NEXT: .LBB10_13: ; %frem.loop_body27 +; VI-NEXT: .LBB10_13: ; %frem.loop_body60 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v6, v5 ; VI-NEXT: v_mul_f32_e32 v5, v6, v4 @@ -1927,7 +1924,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB10_15 ; VI-NEXT: .LBB10_14: ; VI-NEXT: v_mov_b32_e32 v6, v5 -; VI-NEXT: .LBB10_15: ; %frem.loop_exit28 +; VI-NEXT: .LBB10_15: ; %frem.loop_exit61 ; VI-NEXT: v_add_u32_e32 v3, vcc, -10, v3 ; VI-NEXT: v_ldexp_f32 v3, v6, v3 ; VI-NEXT: v_mul_f32_e32 v4, v3, v4 @@ -1948,7 +1945,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: ; implicit-def: $vgpr2 ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v4, v3 ; VI-NEXT: s_cbranch_vccz .LBB10_18 -; VI-NEXT: ; %bb.17: ; %frem.else53 +; VI-NEXT: ; %bb.17: ; %frem.else20 ; VI-NEXT: s_and_b32 s0, s9, 0x8000 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v4, v3 ; VI-NEXT: v_mov_b32_e32 v2, s0 @@ -1959,7 +1956,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s0, s0, 1 ; VI-NEXT: s_cmp_lg_u32 s0, 0 ; VI-NEXT: s_cbranch_scc1 .LBB10_24 -; VI-NEXT: ; %bb.19: ; %frem.compute52 +; VI-NEXT: ; %bb.19: ; %frem.compute19 ; VI-NEXT: v_frexp_mant_f32_e32 v5, v3 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v8, v3 ; VI-NEXT: v_ldexp_f32 v3, v5, 1 @@ -1984,10 +1981,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v4 ; VI-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB10_22 -; VI-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; VI-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; VI-NEXT: v_add_u32_e32 v4, vcc, 11, v7 ; VI-NEXT: v_sub_u32_e32 v4, vcc, v4, v8 -; VI-NEXT: .LBB10_21: ; %frem.loop_body60 +; VI-NEXT: .LBB10_21: ; %frem.loop_body27 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v7, v6 ; VI-NEXT: v_mul_f32_e32 v6, v7, v5 @@ -2003,7 +2000,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB10_23 ; VI-NEXT: .LBB10_22: ; VI-NEXT: v_mov_b32_e32 v7, v6 -; VI-NEXT: .LBB10_23: ; %frem.loop_exit61 +; VI-NEXT: .LBB10_23: ; %frem.loop_exit28 ; VI-NEXT: v_add_u32_e32 v4, vcc, -10, v4 ; VI-NEXT: v_ldexp_f32 v4, v7, v4 ; VI-NEXT: v_mul_f32_e32 v5, v4, v5 @@ -2026,7 +2023,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: ; implicit-def: $vgpr3 ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v5, v4 ; VI-NEXT: s_cbranch_vccz .LBB10_26 -; VI-NEXT: ; %bb.25: ; %frem.else86 +; VI-NEXT: ; %bb.25: ; %frem.else ; VI-NEXT: s_and_b32 s0, s12, 0x8000 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v5, v4 ; VI-NEXT: v_mov_b32_e32 v3, s0 @@ -2037,7 +2034,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s0, s0, 1 ; VI-NEXT: s_cmp_lg_u32 s0, 0 ; VI-NEXT: s_cbranch_scc1 .LBB10_32 -; VI-NEXT: ; %bb.27: ; %frem.compute85 +; VI-NEXT: ; %bb.27: ; %frem.compute ; VI-NEXT: v_frexp_mant_f32_e32 v6, v4 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v9, v4 ; VI-NEXT: v_ldexp_f32 v4, v6, 1 @@ -2062,10 +2059,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 11, v5 ; VI-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB10_30 -; VI-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; VI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; VI-NEXT: v_add_u32_e32 v5, vcc, 11, v8 ; VI-NEXT: v_sub_u32_e32 v5, vcc, v5, v9 -; VI-NEXT: .LBB10_29: ; %frem.loop_body93 +; VI-NEXT: .LBB10_29: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v8, v7 ; VI-NEXT: v_mul_f32_e32 v7, v8, v6 @@ -2081,7 +2078,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB10_31 ; VI-NEXT: .LBB10_30: ; VI-NEXT: v_mov_b32_e32 v8, v7 -; VI-NEXT: .LBB10_31: ; %frem.loop_exit94 +; VI-NEXT: .LBB10_31: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v5, vcc, -10, v5 ; VI-NEXT: v_ldexp_f32 v5, v8, v5 ; VI-NEXT: v_mul_f32_e32 v6, v5, v6 @@ -2147,7 +2144,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f32_e64 vcc, |s2|, |v0| ; CI-NEXT: ; implicit-def: $vgpr0 ; CI-NEXT: s_cbranch_vccz .LBB11_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else16 ; CI-NEXT: s_and_b32 s6, s2, 0x80000000 ; CI-NEXT: v_mov_b32_e32 v1, s4 ; CI-NEXT: v_mov_b32_e32 v0, s2 @@ -2159,7 +2156,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s6, s6, 1 ; CI-NEXT: s_cmp_lg_u32 s6, 0 ; CI-NEXT: s_cbranch_scc1 .LBB11_8 -; CI-NEXT: ; %bb.3: ; %frem.compute +; CI-NEXT: ; %bb.3: ; %frem.compute15 ; CI-NEXT: v_frexp_mant_f32_e64 v1, |s4| ; CI-NEXT: v_ldexp_f32_e64 v1, v1, 1 ; CI-NEXT: v_div_scale_f32 v3, s[6:7], v1, v1, 1.0 @@ -2184,10 +2181,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v2 ; CI-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB11_6 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; CI-NEXT: v_add_i32_e32 v2, vcc, 12, v5 ; CI-NEXT: v_sub_i32_e32 v2, vcc, v2, v6 -; CI-NEXT: .LBB11_5: ; %frem.loop_body +; CI-NEXT: .LBB11_5: ; %frem.loop_body23 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v5, v4 ; CI-NEXT: v_mul_f32_e32 v4, v5, v3 @@ -2203,7 +2200,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB11_7 ; CI-NEXT: .LBB11_6: ; CI-NEXT: v_mov_b32_e32 v5, v4 -; CI-NEXT: .LBB11_7: ; %frem.loop_exit +; CI-NEXT: .LBB11_7: ; %frem.loop_exit24 ; CI-NEXT: v_add_i32_e32 v2, vcc, -11, v2 ; CI-NEXT: v_ldexp_f32_e32 v2, v5, v2 ; CI-NEXT: v_mul_f32_e32 v3, v2, v3 @@ -2222,7 +2219,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_mov_b32 s6, 1 ; CI-NEXT: ; implicit-def: $vgpr1 ; CI-NEXT: s_cbranch_vccz .LBB11_10 -; CI-NEXT: ; %bb.9: ; %frem.else16 +; CI-NEXT: ; %bb.9: ; %frem.else ; CI-NEXT: s_and_b32 s6, s3, 0x80000000 ; CI-NEXT: v_mov_b32_e32 v2, s5 ; CI-NEXT: v_mov_b32_e32 v1, s3 @@ -2234,7 +2231,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s6, s6, 1 ; CI-NEXT: s_cmp_lg_u32 s6, 0 ; CI-NEXT: s_cbranch_scc1 .LBB11_16 -; CI-NEXT: ; %bb.11: ; %frem.compute15 +; CI-NEXT: ; %bb.11: ; %frem.compute ; CI-NEXT: v_frexp_mant_f32_e64 v2, |s5| ; CI-NEXT: v_ldexp_f32_e64 v2, v2, 1 ; CI-NEXT: v_div_scale_f32 v4, s[6:7], v2, v2, 1.0 @@ -2259,10 +2256,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v3 ; CI-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB11_14 -; CI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; CI-NEXT: v_add_i32_e32 v3, vcc, 12, v6 ; CI-NEXT: v_sub_i32_e32 v3, vcc, v3, v7 -; CI-NEXT: .LBB11_13: ; %frem.loop_body23 +; CI-NEXT: .LBB11_13: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v6, v5 ; CI-NEXT: v_mul_f32_e32 v5, v6, v4 @@ -2278,7 +2275,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB11_15 ; CI-NEXT: .LBB11_14: ; CI-NEXT: v_mov_b32_e32 v6, v5 -; CI-NEXT: .LBB11_15: ; %frem.loop_exit24 +; CI-NEXT: .LBB11_15: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v3, vcc, -11, v3 ; CI-NEXT: v_ldexp_f32_e32 v3, v6, v3 ; CI-NEXT: v_mul_f32_e32 v4, v3, v4 @@ -2320,7 +2317,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f32_e64 vcc, |s2|, |v0| ; VI-NEXT: ; implicit-def: $vgpr0 ; VI-NEXT: s_cbranch_vccz .LBB11_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else16 ; VI-NEXT: s_and_b32 s6, s2, 0x80000000 ; VI-NEXT: v_mov_b32_e32 v1, s4 ; VI-NEXT: v_mov_b32_e32 v0, s2 @@ -2332,7 +2329,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s6, s6, 1 ; VI-NEXT: s_cmp_lg_u32 s6, 0 ; VI-NEXT: s_cbranch_scc1 .LBB11_8 -; VI-NEXT: ; %bb.3: ; %frem.compute +; VI-NEXT: ; %bb.3: ; %frem.compute15 ; VI-NEXT: v_frexp_mant_f32_e64 v1, |s4| ; VI-NEXT: v_ldexp_f32 v1, v1, 1 ; VI-NEXT: v_div_scale_f32 v3, s[6:7], v1, v1, 1.0 @@ -2357,10 +2354,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v2 ; VI-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB11_6 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v5 ; VI-NEXT: v_sub_u32_e32 v2, vcc, v2, v6 -; VI-NEXT: .LBB11_5: ; %frem.loop_body +; VI-NEXT: .LBB11_5: ; %frem.loop_body23 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v5, v4 ; VI-NEXT: v_mul_f32_e32 v4, v5, v3 @@ -2376,7 +2373,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB11_7 ; VI-NEXT: .LBB11_6: ; VI-NEXT: v_mov_b32_e32 v5, v4 -; VI-NEXT: .LBB11_7: ; %frem.loop_exit +; VI-NEXT: .LBB11_7: ; %frem.loop_exit24 ; VI-NEXT: v_add_u32_e32 v2, vcc, -11, v2 ; VI-NEXT: v_ldexp_f32 v2, v5, v2 ; VI-NEXT: v_mul_f32_e32 v3, v2, v3 @@ -2395,7 +2392,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_mov_b32 s6, 1 ; VI-NEXT: ; implicit-def: $vgpr1 ; VI-NEXT: s_cbranch_vccz .LBB11_10 -; VI-NEXT: ; %bb.9: ; %frem.else16 +; VI-NEXT: ; %bb.9: ; %frem.else ; VI-NEXT: s_and_b32 s6, s3, 0x80000000 ; VI-NEXT: v_mov_b32_e32 v2, s5 ; VI-NEXT: v_mov_b32_e32 v1, s3 @@ -2407,7 +2404,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s6, s6, 1 ; VI-NEXT: s_cmp_lg_u32 s6, 0 ; VI-NEXT: s_cbranch_scc1 .LBB11_16 -; VI-NEXT: ; %bb.11: ; %frem.compute15 +; VI-NEXT: ; %bb.11: ; %frem.compute ; VI-NEXT: v_frexp_mant_f32_e64 v2, |s5| ; VI-NEXT: v_ldexp_f32 v2, v2, 1 ; VI-NEXT: v_div_scale_f32 v4, s[6:7], v2, v2, 1.0 @@ -2432,10 +2429,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v3 ; VI-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB11_14 -; VI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; VI-NEXT: v_add_u32_e32 v3, vcc, 12, v6 ; VI-NEXT: v_sub_u32_e32 v3, vcc, v3, v7 -; VI-NEXT: .LBB11_13: ; %frem.loop_body23 +; VI-NEXT: .LBB11_13: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v6, v5 ; VI-NEXT: v_mul_f32_e32 v5, v6, v4 @@ -2451,7 +2448,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB11_15 ; VI-NEXT: .LBB11_14: ; VI-NEXT: v_mov_b32_e32 v6, v5 -; VI-NEXT: .LBB11_15: ; %frem.loop_exit24 +; VI-NEXT: .LBB11_15: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v3, vcc, -11, v3 ; VI-NEXT: v_ldexp_f32 v3, v6, v3 ; VI-NEXT: v_mul_f32_e32 v4, v3, v4 @@ -2501,7 +2498,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f32_e64 vcc, |s4|, |v0| ; CI-NEXT: ; implicit-def: $vgpr0 ; CI-NEXT: s_cbranch_vccz .LBB12_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else78 ; CI-NEXT: s_and_b32 s2, s4, 0x80000000 ; CI-NEXT: v_mov_b32_e32 v1, s8 ; CI-NEXT: v_mov_b32_e32 v0, s4 @@ -2513,7 +2510,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s2, s2, 1 ; CI-NEXT: s_cmp_lg_u32 s2, 0 ; CI-NEXT: s_cbranch_scc1 .LBB12_8 -; CI-NEXT: ; %bb.3: ; %frem.compute +; CI-NEXT: ; %bb.3: ; %frem.compute77 ; CI-NEXT: v_frexp_mant_f32_e64 v1, |s8| ; CI-NEXT: v_ldexp_f32_e64 v1, v1, 1 ; CI-NEXT: v_div_scale_f32 v3, s[2:3], v1, v1, 1.0 @@ -2538,10 +2535,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v2 ; CI-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB12_6 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; CI-NEXT: v_add_i32_e32 v2, vcc, 12, v5 ; CI-NEXT: v_sub_i32_e32 v2, vcc, v2, v6 -; CI-NEXT: .LBB12_5: ; %frem.loop_body +; CI-NEXT: .LBB12_5: ; %frem.loop_body85 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v5, v4 ; CI-NEXT: v_mul_f32_e32 v4, v5, v3 @@ -2557,7 +2554,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB12_7 ; CI-NEXT: .LBB12_6: ; CI-NEXT: v_mov_b32_e32 v5, v4 -; CI-NEXT: .LBB12_7: ; %frem.loop_exit +; CI-NEXT: .LBB12_7: ; %frem.loop_exit86 ; CI-NEXT: v_add_i32_e32 v2, vcc, -11, v2 ; CI-NEXT: v_ldexp_f32_e32 v2, v5, v2 ; CI-NEXT: v_mul_f32_e32 v3, v2, v3 @@ -2576,7 +2573,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_mov_b32 s2, 1 ; CI-NEXT: ; implicit-def: $vgpr1 ; CI-NEXT: s_cbranch_vccz .LBB12_10 -; CI-NEXT: ; %bb.9: ; %frem.else16 +; CI-NEXT: ; %bb.9: ; %frem.else47 ; CI-NEXT: s_and_b32 s2, s5, 0x80000000 ; CI-NEXT: v_mov_b32_e32 v2, s9 ; CI-NEXT: v_mov_b32_e32 v1, s5 @@ -2588,7 +2585,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s2, s2, 1 ; CI-NEXT: s_cmp_lg_u32 s2, 0 ; CI-NEXT: s_cbranch_scc1 .LBB12_16 -; CI-NEXT: ; %bb.11: ; %frem.compute15 +; CI-NEXT: ; %bb.11: ; %frem.compute46 ; CI-NEXT: v_frexp_mant_f32_e64 v2, |s9| ; CI-NEXT: v_ldexp_f32_e64 v2, v2, 1 ; CI-NEXT: v_div_scale_f32 v4, s[2:3], v2, v2, 1.0 @@ -2613,10 +2610,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v3 ; CI-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB12_14 -; CI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; CI-NEXT: v_add_i32_e32 v3, vcc, 12, v6 ; CI-NEXT: v_sub_i32_e32 v3, vcc, v3, v7 -; CI-NEXT: .LBB12_13: ; %frem.loop_body23 +; CI-NEXT: .LBB12_13: ; %frem.loop_body54 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v6, v5 ; CI-NEXT: v_mul_f32_e32 v5, v6, v4 @@ -2632,7 +2629,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB12_15 ; CI-NEXT: .LBB12_14: ; CI-NEXT: v_mov_b32_e32 v6, v5 -; CI-NEXT: .LBB12_15: ; %frem.loop_exit24 +; CI-NEXT: .LBB12_15: ; %frem.loop_exit55 ; CI-NEXT: v_add_i32_e32 v3, vcc, -11, v3 ; CI-NEXT: v_ldexp_f32_e32 v3, v6, v3 ; CI-NEXT: v_mul_f32_e32 v4, v3, v4 @@ -2651,7 +2648,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_mov_b32 s2, 1 ; CI-NEXT: ; implicit-def: $vgpr2 ; CI-NEXT: s_cbranch_vccz .LBB12_18 -; CI-NEXT: ; %bb.17: ; %frem.else47 +; CI-NEXT: ; %bb.17: ; %frem.else16 ; CI-NEXT: s_and_b32 s2, s6, 0x80000000 ; CI-NEXT: v_mov_b32_e32 v3, s10 ; CI-NEXT: v_mov_b32_e32 v2, s6 @@ -2663,7 +2660,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s2, s2, 1 ; CI-NEXT: s_cmp_lg_u32 s2, 0 ; CI-NEXT: s_cbranch_scc1 .LBB12_24 -; CI-NEXT: ; %bb.19: ; %frem.compute46 +; CI-NEXT: ; %bb.19: ; %frem.compute15 ; CI-NEXT: v_frexp_mant_f32_e64 v3, |s10| ; CI-NEXT: v_ldexp_f32_e64 v3, v3, 1 ; CI-NEXT: v_div_scale_f32 v5, s[2:3], v3, v3, 1.0 @@ -2688,10 +2685,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v4 ; CI-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB12_22 -; CI-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; CI-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; CI-NEXT: v_add_i32_e32 v4, vcc, 12, v7 ; CI-NEXT: v_sub_i32_e32 v4, vcc, v4, v8 -; CI-NEXT: .LBB12_21: ; %frem.loop_body54 +; CI-NEXT: .LBB12_21: ; %frem.loop_body23 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v7, v6 ; CI-NEXT: v_mul_f32_e32 v6, v7, v5 @@ -2707,7 +2704,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB12_23 ; CI-NEXT: .LBB12_22: ; CI-NEXT: v_mov_b32_e32 v7, v6 -; CI-NEXT: .LBB12_23: ; %frem.loop_exit55 +; CI-NEXT: .LBB12_23: ; %frem.loop_exit24 ; CI-NEXT: v_add_i32_e32 v4, vcc, -11, v4 ; CI-NEXT: v_ldexp_f32_e32 v4, v7, v4 ; CI-NEXT: v_mul_f32_e32 v5, v4, v5 @@ -2726,7 +2723,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_mov_b32 s2, 1 ; CI-NEXT: ; implicit-def: $vgpr3 ; CI-NEXT: s_cbranch_vccz .LBB12_26 -; CI-NEXT: ; %bb.25: ; %frem.else78 +; CI-NEXT: ; %bb.25: ; %frem.else ; CI-NEXT: s_and_b32 s2, s7, 0x80000000 ; CI-NEXT: v_mov_b32_e32 v4, s11 ; CI-NEXT: v_mov_b32_e32 v3, s7 @@ -2738,7 +2735,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s2, s2, 1 ; CI-NEXT: s_cmp_lg_u32 s2, 0 ; CI-NEXT: s_cbranch_scc1 .LBB12_32 -; CI-NEXT: ; %bb.27: ; %frem.compute77 +; CI-NEXT: ; %bb.27: ; %frem.compute ; CI-NEXT: v_frexp_mant_f32_e64 v4, |s11| ; CI-NEXT: v_ldexp_f32_e64 v4, v4, 1 ; CI-NEXT: v_div_scale_f32 v6, s[2:3], v4, v4, 1.0 @@ -2763,10 +2760,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v5 ; CI-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB12_30 -; CI-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; CI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; CI-NEXT: v_add_i32_e32 v5, vcc, 12, v8 ; CI-NEXT: v_sub_i32_e32 v5, vcc, v5, v9 -; CI-NEXT: .LBB12_29: ; %frem.loop_body85 +; CI-NEXT: .LBB12_29: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v8, v7 ; CI-NEXT: v_mul_f32_e32 v7, v8, v6 @@ -2782,7 +2779,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB12_31 ; CI-NEXT: .LBB12_30: ; CI-NEXT: v_mov_b32_e32 v8, v7 -; CI-NEXT: .LBB12_31: ; %frem.loop_exit86 +; CI-NEXT: .LBB12_31: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v5, vcc, -11, v5 ; CI-NEXT: v_ldexp_f32_e32 v5, v8, v5 ; CI-NEXT: v_mul_f32_e32 v6, v5, v6 @@ -2832,7 +2829,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f32_e64 vcc, |s4|, |v0| ; VI-NEXT: ; implicit-def: $vgpr0 ; VI-NEXT: s_cbranch_vccz .LBB12_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else78 ; VI-NEXT: s_and_b32 s2, s4, 0x80000000 ; VI-NEXT: v_mov_b32_e32 v1, s8 ; VI-NEXT: v_mov_b32_e32 v0, s4 @@ -2844,7 +2841,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s2, s2, 1 ; VI-NEXT: s_cmp_lg_u32 s2, 0 ; VI-NEXT: s_cbranch_scc1 .LBB12_8 -; VI-NEXT: ; %bb.3: ; %frem.compute +; VI-NEXT: ; %bb.3: ; %frem.compute77 ; VI-NEXT: v_frexp_mant_f32_e64 v1, |s8| ; VI-NEXT: v_ldexp_f32 v1, v1, 1 ; VI-NEXT: v_div_scale_f32 v3, s[2:3], v1, v1, 1.0 @@ -2869,10 +2866,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v2 ; VI-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB12_6 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; VI-NEXT: v_add_u32_e32 v2, vcc, 12, v5 ; VI-NEXT: v_sub_u32_e32 v2, vcc, v2, v6 -; VI-NEXT: .LBB12_5: ; %frem.loop_body +; VI-NEXT: .LBB12_5: ; %frem.loop_body85 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v5, v4 ; VI-NEXT: v_mul_f32_e32 v4, v5, v3 @@ -2888,7 +2885,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB12_7 ; VI-NEXT: .LBB12_6: ; VI-NEXT: v_mov_b32_e32 v5, v4 -; VI-NEXT: .LBB12_7: ; %frem.loop_exit +; VI-NEXT: .LBB12_7: ; %frem.loop_exit86 ; VI-NEXT: v_add_u32_e32 v2, vcc, -11, v2 ; VI-NEXT: v_ldexp_f32 v2, v5, v2 ; VI-NEXT: v_mul_f32_e32 v3, v2, v3 @@ -2907,7 +2904,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_mov_b32 s2, 1 ; VI-NEXT: ; implicit-def: $vgpr1 ; VI-NEXT: s_cbranch_vccz .LBB12_10 -; VI-NEXT: ; %bb.9: ; %frem.else16 +; VI-NEXT: ; %bb.9: ; %frem.else47 ; VI-NEXT: s_and_b32 s2, s5, 0x80000000 ; VI-NEXT: v_mov_b32_e32 v2, s9 ; VI-NEXT: v_mov_b32_e32 v1, s5 @@ -2919,7 +2916,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s2, s2, 1 ; VI-NEXT: s_cmp_lg_u32 s2, 0 ; VI-NEXT: s_cbranch_scc1 .LBB12_16 -; VI-NEXT: ; %bb.11: ; %frem.compute15 +; VI-NEXT: ; %bb.11: ; %frem.compute46 ; VI-NEXT: v_frexp_mant_f32_e64 v2, |s9| ; VI-NEXT: v_ldexp_f32 v2, v2, 1 ; VI-NEXT: v_div_scale_f32 v4, s[2:3], v2, v2, 1.0 @@ -2944,10 +2941,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v3 ; VI-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB12_14 -; VI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; VI-NEXT: v_add_u32_e32 v3, vcc, 12, v6 ; VI-NEXT: v_sub_u32_e32 v3, vcc, v3, v7 -; VI-NEXT: .LBB12_13: ; %frem.loop_body23 +; VI-NEXT: .LBB12_13: ; %frem.loop_body54 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v6, v5 ; VI-NEXT: v_mul_f32_e32 v5, v6, v4 @@ -2963,7 +2960,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB12_15 ; VI-NEXT: .LBB12_14: ; VI-NEXT: v_mov_b32_e32 v6, v5 -; VI-NEXT: .LBB12_15: ; %frem.loop_exit24 +; VI-NEXT: .LBB12_15: ; %frem.loop_exit55 ; VI-NEXT: v_add_u32_e32 v3, vcc, -11, v3 ; VI-NEXT: v_ldexp_f32 v3, v6, v3 ; VI-NEXT: v_mul_f32_e32 v4, v3, v4 @@ -2982,7 +2979,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_mov_b32 s2, 1 ; VI-NEXT: ; implicit-def: $vgpr2 ; VI-NEXT: s_cbranch_vccz .LBB12_18 -; VI-NEXT: ; %bb.17: ; %frem.else47 +; VI-NEXT: ; %bb.17: ; %frem.else16 ; VI-NEXT: s_and_b32 s2, s6, 0x80000000 ; VI-NEXT: v_mov_b32_e32 v3, s10 ; VI-NEXT: v_mov_b32_e32 v2, s6 @@ -2994,7 +2991,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s2, s2, 1 ; VI-NEXT: s_cmp_lg_u32 s2, 0 ; VI-NEXT: s_cbranch_scc1 .LBB12_24 -; VI-NEXT: ; %bb.19: ; %frem.compute46 +; VI-NEXT: ; %bb.19: ; %frem.compute15 ; VI-NEXT: v_frexp_mant_f32_e64 v3, |s10| ; VI-NEXT: v_ldexp_f32 v3, v3, 1 ; VI-NEXT: v_div_scale_f32 v5, s[2:3], v3, v3, 1.0 @@ -3019,10 +3016,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v4 ; VI-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB12_22 -; VI-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; VI-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; VI-NEXT: v_add_u32_e32 v4, vcc, 12, v7 ; VI-NEXT: v_sub_u32_e32 v4, vcc, v4, v8 -; VI-NEXT: .LBB12_21: ; %frem.loop_body54 +; VI-NEXT: .LBB12_21: ; %frem.loop_body23 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v7, v6 ; VI-NEXT: v_mul_f32_e32 v6, v7, v5 @@ -3038,7 +3035,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB12_23 ; VI-NEXT: .LBB12_22: ; VI-NEXT: v_mov_b32_e32 v7, v6 -; VI-NEXT: .LBB12_23: ; %frem.loop_exit55 +; VI-NEXT: .LBB12_23: ; %frem.loop_exit24 ; VI-NEXT: v_add_u32_e32 v4, vcc, -11, v4 ; VI-NEXT: v_ldexp_f32 v4, v7, v4 ; VI-NEXT: v_mul_f32_e32 v5, v4, v5 @@ -3057,7 +3054,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_mov_b32 s2, 1 ; VI-NEXT: ; implicit-def: $vgpr3 ; VI-NEXT: s_cbranch_vccz .LBB12_26 -; VI-NEXT: ; %bb.25: ; %frem.else78 +; VI-NEXT: ; %bb.25: ; %frem.else ; VI-NEXT: s_and_b32 s2, s7, 0x80000000 ; VI-NEXT: v_mov_b32_e32 v4, s11 ; VI-NEXT: v_mov_b32_e32 v3, s7 @@ -3069,7 +3066,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s2, s2, 1 ; VI-NEXT: s_cmp_lg_u32 s2, 0 ; VI-NEXT: s_cbranch_scc1 .LBB12_32 -; VI-NEXT: ; %bb.27: ; %frem.compute77 +; VI-NEXT: ; %bb.27: ; %frem.compute ; VI-NEXT: v_frexp_mant_f32_e64 v4, |s11| ; VI-NEXT: v_ldexp_f32 v4, v4, 1 ; VI-NEXT: v_div_scale_f32 v6, s[2:3], v4, v4, 1.0 @@ -3094,10 +3091,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 12, v5 ; VI-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB12_30 -; VI-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; VI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; VI-NEXT: v_add_u32_e32 v5, vcc, 12, v8 ; VI-NEXT: v_sub_u32_e32 v5, vcc, v5, v9 -; VI-NEXT: .LBB12_29: ; %frem.loop_body85 +; VI-NEXT: .LBB12_29: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v8, v7 ; VI-NEXT: v_mul_f32_e32 v7, v8, v6 @@ -3113,7 +3110,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB12_31 ; VI-NEXT: .LBB12_30: ; VI-NEXT: v_mov_b32_e32 v8, v7 -; VI-NEXT: .LBB12_31: ; %frem.loop_exit86 +; VI-NEXT: .LBB12_31: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v5, vcc, -11, v5 ; VI-NEXT: v_ldexp_f32 v5, v8, v5 ; VI-NEXT: v_mul_f32_e32 v6, v5, v6 @@ -3172,7 +3169,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f64_e64 vcc, |s[4:5]|, |v[0:1]| ; CI-NEXT: ; implicit-def: $vgpr0_vgpr1 ; CI-NEXT: s_cbranch_vccz .LBB13_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else16 ; CI-NEXT: v_mov_b32_e32 v0, s8 ; CI-NEXT: v_mov_b32_e32 v1, s9 ; CI-NEXT: v_cmp_eq_f64_e64 vcc, |s[4:5]|, |v[0:1]| @@ -3190,7 +3187,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s2, s2, 1 ; CI-NEXT: s_cmp_lg_u32 s2, 0 ; CI-NEXT: s_cbranch_scc1 .LBB13_8 -; CI-NEXT: ; %bb.3: ; %frem.compute +; CI-NEXT: ; %bb.3: ; %frem.compute15 ; CI-NEXT: v_frexp_mant_f64_e64 v[0:1], |s[4:5]| ; CI-NEXT: v_frexp_exp_i32_f64_e64 v6, |s[4:5]| ; CI-NEXT: v_frexp_exp_i32_f64_e64 v7, |s[8:9]| @@ -3213,10 +3210,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 26, v9 ; CI-NEXT: v_div_fixup_f64 v[2:3], v[2:3], v[0:1], 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB13_6 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; CI-NEXT: v_add_i32_e32 v6, vcc, 26, v6 ; CI-NEXT: v_sub_i32_e32 v9, vcc, v6, v7 -; CI-NEXT: .LBB13_5: ; %frem.loop_body +; CI-NEXT: .LBB13_5: ; %frem.loop_body23 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v7, v5 ; CI-NEXT: v_mov_b32_e32 v6, v4 @@ -3235,7 +3232,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: .LBB13_6: ; CI-NEXT: v_mov_b32_e32 v7, v5 ; CI-NEXT: v_mov_b32_e32 v6, v4 -; CI-NEXT: .LBB13_7: ; %frem.loop_exit +; CI-NEXT: .LBB13_7: ; %frem.loop_exit24 ; CI-NEXT: v_add_i32_e32 v4, vcc, 0xffffffe7, v9 ; CI-NEXT: v_ldexp_f64 v[4:5], v[6:7], v4 ; CI-NEXT: s_mov_b32 s2, 0 @@ -3259,7 +3256,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_mov_b32 s2, 1 ; CI-NEXT: ; implicit-def: $vgpr2_vgpr3 ; CI-NEXT: s_cbranch_vccz .LBB13_10 -; CI-NEXT: ; %bb.9: ; %frem.else16 +; CI-NEXT: ; %bb.9: ; %frem.else ; CI-NEXT: v_mov_b32_e32 v2, s10 ; CI-NEXT: v_mov_b32_e32 v3, s11 ; CI-NEXT: v_cmp_eq_f64_e64 vcc, |s[6:7]|, |v[2:3]| @@ -3277,7 +3274,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_xor_b32 s2, s2, 1 ; CI-NEXT: s_cmp_lg_u32 s2, 0 ; CI-NEXT: s_cbranch_scc1 .LBB13_16 -; CI-NEXT: ; %bb.11: ; %frem.compute15 +; CI-NEXT: ; %bb.11: ; %frem.compute ; CI-NEXT: v_frexp_mant_f64_e64 v[2:3], |s[6:7]| ; CI-NEXT: v_frexp_exp_i32_f64_e64 v8, |s[6:7]| ; CI-NEXT: v_frexp_exp_i32_f64_e64 v9, |s[10:11]| @@ -3300,10 +3297,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ge_i32_e32 vcc, 26, v11 ; CI-NEXT: v_div_fixup_f64 v[4:5], v[4:5], v[2:3], 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB13_14 -; CI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; CI-NEXT: v_add_i32_e32 v8, vcc, 26, v8 ; CI-NEXT: v_sub_i32_e32 v11, vcc, v8, v9 -; CI-NEXT: .LBB13_13: ; %frem.loop_body23 +; CI-NEXT: .LBB13_13: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v9, v7 ; CI-NEXT: v_mov_b32_e32 v8, v6 @@ -3322,7 +3319,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: .LBB13_14: ; CI-NEXT: v_mov_b32_e32 v9, v7 ; CI-NEXT: v_mov_b32_e32 v8, v6 -; CI-NEXT: .LBB13_15: ; %frem.loop_exit24 +; CI-NEXT: .LBB13_15: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v6, vcc, 0xffffffe7, v11 ; CI-NEXT: v_ldexp_f64 v[6:7], v[8:9], v6 ; CI-NEXT: s_mov_b32 s2, 0 @@ -3374,7 +3371,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f64_e64 vcc, |s[4:5]|, |v[0:1]| ; VI-NEXT: ; implicit-def: $vgpr0_vgpr1 ; VI-NEXT: s_cbranch_vccz .LBB13_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else16 ; VI-NEXT: v_mov_b32_e32 v0, s8 ; VI-NEXT: v_mov_b32_e32 v1, s9 ; VI-NEXT: v_cmp_eq_f64_e64 vcc, |s[4:5]|, |v[0:1]| @@ -3392,7 +3389,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s2, s2, 1 ; VI-NEXT: s_cmp_lg_u32 s2, 0 ; VI-NEXT: s_cbranch_scc1 .LBB13_8 -; VI-NEXT: ; %bb.3: ; %frem.compute +; VI-NEXT: ; %bb.3: ; %frem.compute15 ; VI-NEXT: v_frexp_mant_f64_e64 v[0:1], |s[4:5]| ; VI-NEXT: v_frexp_exp_i32_f64_e64 v6, |s[4:5]| ; VI-NEXT: v_frexp_exp_i32_f64_e64 v7, |s[8:9]| @@ -3415,10 +3412,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 26, v9 ; VI-NEXT: v_div_fixup_f64 v[2:3], v[2:3], v[0:1], 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB13_6 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; VI-NEXT: v_add_u32_e32 v6, vcc, 26, v6 ; VI-NEXT: v_sub_u32_e32 v9, vcc, v6, v7 -; VI-NEXT: .LBB13_5: ; %frem.loop_body +; VI-NEXT: .LBB13_5: ; %frem.loop_body23 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v7, v5 ; VI-NEXT: v_mov_b32_e32 v6, v4 @@ -3437,7 +3434,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: .LBB13_6: ; VI-NEXT: v_mov_b32_e32 v7, v5 ; VI-NEXT: v_mov_b32_e32 v6, v4 -; VI-NEXT: .LBB13_7: ; %frem.loop_exit +; VI-NEXT: .LBB13_7: ; %frem.loop_exit24 ; VI-NEXT: v_add_u32_e32 v4, vcc, 0xffffffe7, v9 ; VI-NEXT: v_ldexp_f64 v[4:5], v[6:7], v4 ; VI-NEXT: s_mov_b32 s2, 0 @@ -3461,7 +3458,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_mov_b32 s2, 1 ; VI-NEXT: ; implicit-def: $vgpr2_vgpr3 ; VI-NEXT: s_cbranch_vccz .LBB13_10 -; VI-NEXT: ; %bb.9: ; %frem.else16 +; VI-NEXT: ; %bb.9: ; %frem.else ; VI-NEXT: v_mov_b32_e32 v2, s10 ; VI-NEXT: v_mov_b32_e32 v3, s11 ; VI-NEXT: v_cmp_eq_f64_e64 vcc, |s[6:7]|, |v[2:3]| @@ -3479,7 +3476,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_xor_b32 s2, s2, 1 ; VI-NEXT: s_cmp_lg_u32 s2, 0 ; VI-NEXT: s_cbranch_scc1 .LBB13_16 -; VI-NEXT: ; %bb.11: ; %frem.compute15 +; VI-NEXT: ; %bb.11: ; %frem.compute ; VI-NEXT: v_frexp_mant_f64_e64 v[2:3], |s[6:7]| ; VI-NEXT: v_frexp_exp_i32_f64_e64 v8, |s[6:7]| ; VI-NEXT: v_frexp_exp_i32_f64_e64 v9, |s[10:11]| @@ -3502,10 +3499,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ge_i32_e32 vcc, 26, v11 ; VI-NEXT: v_div_fixup_f64 v[4:5], v[4:5], v[2:3], 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB13_14 -; VI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; VI-NEXT: v_add_u32_e32 v8, vcc, 26, v8 ; VI-NEXT: v_sub_u32_e32 v11, vcc, v8, v9 -; VI-NEXT: .LBB13_13: ; %frem.loop_body23 +; VI-NEXT: .LBB13_13: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v9, v7 ; VI-NEXT: v_mov_b32_e32 v8, v6 @@ -3524,7 +3521,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: .LBB13_14: ; VI-NEXT: v_mov_b32_e32 v9, v7 ; VI-NEXT: v_mov_b32_e32 v8, v6 -; VI-NEXT: .LBB13_15: ; %frem.loop_exit24 +; VI-NEXT: .LBB13_15: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v6, vcc, 0xffffffe7, v11 ; VI-NEXT: v_ldexp_f64 v[6:7], v[8:9], v6 ; VI-NEXT: s_mov_b32 s2, 0 diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.large.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.large.ll index 920d8fa85a44..ae7f6ec5d1c0 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.large.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/insertelement.large.ll @@ -1,7 +1,7 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefix=GCN %s -; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s -; RUN: llc -global-isel -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefix=GCN %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefix=GFX11 %s define amdgpu_kernel void @v_insert_v64i32_37(ptr addrspace(1) %ptr.in, ptr addrspace(1) %ptr.out) #0 { ; GCN-LABEL: v_insert_v64i32_37: diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/lds-misaligned-bug.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/lds-misaligned-bug.ll index cfbb429420cf..aabf256a1196 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/lds-misaligned-bug.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/lds-misaligned-bug.ll @@ -1,11 +1,11 @@ -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-WGP %s -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1011 < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-WGP %s -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1012 < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-WGP %s -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1010 -mattr=+cumode < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-CU %s -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1010 -mattr=+cumode,+unaligned-access-mode < %s | FileCheck -check-prefixes=GCN,UNALIGNED %s -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-CU %s -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 -mattr=+cumode < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-CU %s -; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1100 -mattr=+cumode,+unaligned-access-mode < %s | FileCheck -check-prefixes=GCN,UNALIGNED %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-WGP %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1011 < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-WGP %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1012 < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-WGP %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1010 -mattr=+cumode < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-CU %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1010 -mattr=+cumode,+unaligned-access-mode < %s | FileCheck -check-prefixes=GCN,UNALIGNED %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-CU %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=+cumode < %s | FileCheck -check-prefixes=GCN,ALIGNED,ALIGNED-CU %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=+cumode,+unaligned-access-mode < %s | FileCheck -check-prefixes=GCN,UNALIGNED %s ; GCN-LABEL: test_local_misaligned_v2: ; GCN-DAG: ds_{{read2|load_2addr}}_b32 diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.workitem.id.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.workitem.id.ll index 66cdfc263d7f..7b923f4c9328 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.workitem.id.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.workitem.id.ll @@ -1,14 +1,14 @@ ; RUN: sed 's/CODE_OBJECT_VERSION/400/g' %s | opt -S -mtriple=amdgcn-amd-amdhsa -passes=amdgpu-attributor -o %t.v4.ll ; RUN: sed 's/CODE_OBJECT_VERSION/600/g' %s | opt -S -mtriple=amdgcn-amd-amdhsa -passes=amdgpu-attributor -o %t.v6.ll -; RUN: llc -global-isel -mtriple=amdgcn-unknown-amdhsa < %t.v4.ll | FileCheck --check-prefixes=ALL,HSA,UNPACKED %s -; RUN: llc -global-isel -mtriple=amdgcn-unknown-amdhsa < %t.v4.ll | FileCheck --check-prefixes=ALL,HSA,UNPACKED %s -; RUN: llc -global-isel -mtriple=amdgcn-- -mcpu=hawaii -mattr=+flat-for-global < %t.v4.ll | FileCheck --check-prefixes=ALL,MESA,UNPACKED %s -; RUN: llc -global-isel -mtriple=amdgcn-- -mcpu=tonga -mattr=+flat-for-global < %t.v4.ll | FileCheck --check-prefixes=ALL,MESA,UNPACKED %s -; RUN: llc -global-isel -mtriple=amdgcn-unknown-mesa3d -mattr=+flat-for-global -mcpu=hawaii < %t.v4.ll | FileCheck -check-prefixes=ALL,MESA3D,UNPACKED %s -; RUN: llc -global-isel -mtriple=amdgcn-unknown-mesa3d -mcpu=tonga < %t.v4.ll | FileCheck -check-prefixes=ALL,MESA3D,UNPACKED %s -; RUN: llc -global-isel -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx90a < %t.v4.ll | FileCheck -check-prefixes=ALL,PACKED-TID %s -; RUN: llc -global-isel -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx1100 -amdgpu-enable-vopd=0 < %t.v4.ll | FileCheck -check-prefixes=ALL,PACKED-TID %s -; RUN: llc -global-isel -mtriple=amdgcn-unknown-amdhsa --amdhsa-code-object-version=6 -mcpu=gfx11-generic -amdgpu-enable-vopd=0 < %t.v6.ll | FileCheck -check-prefixes=ALL,PACKED-TID %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-unknown-amdhsa < %t.v4.ll | FileCheck --check-prefixes=ALL,HSA,UNPACKED %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-unknown-amdhsa < %t.v4.ll | FileCheck --check-prefixes=ALL,HSA,UNPACKED %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-- -mcpu=hawaii -mattr=+flat-for-global < %t.v4.ll | FileCheck --check-prefixes=ALL,MESA,UNPACKED %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-- -mcpu=tonga -mattr=+flat-for-global < %t.v4.ll | FileCheck --check-prefixes=ALL,MESA,UNPACKED %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-unknown-mesa3d -mattr=+flat-for-global -mcpu=hawaii < %t.v4.ll | FileCheck -check-prefixes=ALL,MESA3D,UNPACKED %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-unknown-mesa3d -mcpu=tonga < %t.v4.ll | FileCheck -check-prefixes=ALL,MESA3D,UNPACKED %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx90a < %t.v4.ll | FileCheck -check-prefixes=ALL,PACKED-TID %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx1100 -amdgpu-enable-vopd=0 < %t.v4.ll | FileCheck -check-prefixes=ALL,PACKED-TID %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-unknown-amdhsa --amdhsa-code-object-version=6 -mcpu=gfx11-generic -amdgpu-enable-vopd=0 < %t.v6.ll | FileCheck -check-prefixes=ALL,PACKED-TID %s declare i32 @llvm.amdgcn.workitem.id.x() #0 declare i32 @llvm.amdgcn.workitem.id.y() #0 diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-assert-sext.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-assert-sext.mir new file mode 100644 index 000000000000..87836e2aed5a --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-assert-sext.mir @@ -0,0 +1,170 @@ +# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +# RUN: llc -mtriple=amdgcn -mcpu=gfx90a -run-pass=amdgpu-regbankselect,amdgpu-regbanklegalize %s -verify-machineinstrs -o - | FileCheck %s + +--- +name: assert_sext_vgpr +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $vgpr0 + + ; CHECK-LABEL: name: assert_sext_vgpr + ; CHECK: liveins: $vgpr0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:vgpr(s32) = COPY $vgpr0 + ; CHECK-NEXT: %assert_sext:vgpr(s32) = G_ASSERT_SEXT %copy, 4 + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_sext(s32) + %copy:_(s32) = COPY $vgpr0 + %assert_sext:_(s32) = G_ASSERT_SEXT %copy, 4 + S_ENDPGM 0, implicit %assert_sext +... + +--- +name: assert_sext_sgpr +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $sgpr8 + + ; CHECK-LABEL: name: assert_sext_sgpr + ; CHECK: liveins: $sgpr8 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:sgpr(s32) = COPY $sgpr8 + ; CHECK-NEXT: %assert_sext:sgpr(s32) = G_ASSERT_SEXT %copy, 4 + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_sext(s32) + %copy:_(s32) = COPY $sgpr8 + %assert_sext:_(s32) = G_ASSERT_SEXT %copy, 4 + S_ENDPGM 0, implicit %assert_sext +... + +--- +name: assert_sext_agpr +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $agpr0 + + ; CHECK-LABEL: name: assert_sext_agpr + ; CHECK: liveins: $agpr0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:vgpr(s32) = COPY $agpr0 + ; CHECK-NEXT: %assert_sext:vgpr(s32) = G_ASSERT_SEXT %copy, 4 + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_sext(s32) + %copy:_(s32) = COPY $agpr0 + %assert_sext:_(s32) = G_ASSERT_SEXT %copy, 4 + S_ENDPGM 0, implicit %assert_sext +... + +--- +name: assert_sext_vgpr_regclass +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $vgpr0 + + ; CHECK-LABEL: name: assert_sext_vgpr_regclass + ; CHECK: liveins: $vgpr0 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:vgpr_32(s32) = COPY $vgpr0 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY %copy(s32) + ; CHECK-NEXT: %assert_sext:vgpr(s32) = G_ASSERT_SEXT [[COPY]], 4 + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_sext(s32) + %copy:vgpr_32(s32) = COPY $vgpr0 + %assert_sext:_(s32) = G_ASSERT_SEXT %copy, 4 + S_ENDPGM 0, implicit %assert_sext +... + +--- +name: assert_sext_sgpr_regcllass +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $sgpr8 + + ; CHECK-LABEL: name: assert_sext_sgpr_regcllass + ; CHECK: liveins: $sgpr8 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:sgpr_32(s32) = COPY $sgpr8 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY %copy(s32) + ; CHECK-NEXT: %assert_sext:sgpr(s32) = G_ASSERT_SEXT [[COPY]], 4 + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_sext(s32) + %copy:sgpr_32(s32) = COPY $sgpr8 + %assert_sext:_(s32) = G_ASSERT_SEXT %copy, 4 + S_ENDPGM 0, implicit %assert_sext +... + +--- +name: assert_sext_vgpr_64 +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $vgpr0_vgpr1 + + ; CHECK-LABEL: name: assert_sext_vgpr_64 + ; CHECK: liveins: $vgpr0_vgpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:vreg_64(s64) = COPY $vgpr0_vgpr1 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s64) = COPY %copy(s64) + ; CHECK-NEXT: [[ASSERT_SEXT:%[0-9]+]]:vgpr(s64) = G_ASSERT_SEXT [[COPY]], 4 + ; CHECK-NEXT: %assert_sext:vreg_64(s64) = COPY [[ASSERT_SEXT]](s64) + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_sext(s64) + %copy:vreg_64(s64) = COPY $vgpr0_vgpr1 + %assert_sext:vreg_64(s64) = G_ASSERT_SEXT %copy, 4 + S_ENDPGM 0, implicit %assert_sext +... + +--- +name: assert_sext_sgpr_64 +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $sgpr0_sgpr1 + + ; CHECK-LABEL: name: assert_sext_sgpr_64 + ; CHECK: liveins: $sgpr0_sgpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:sreg_64(s64) = COPY $sgpr0_sgpr1 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s64) = COPY %copy(s64) + ; CHECK-NEXT: [[ASSERT_SEXT:%[0-9]+]]:sgpr(s64) = G_ASSERT_SEXT [[COPY]], 4 + ; CHECK-NEXT: %assert_sext:sreg_64(s64) = COPY [[ASSERT_SEXT]](s64) + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_sext(s64) + %copy:sreg_64(s64) = COPY $sgpr0_sgpr1 + %assert_sext:sreg_64(s64) = G_ASSERT_SEXT %copy, 4 + S_ENDPGM 0, implicit %assert_sext +... + +--- +name: assert_sext_agpr_64 +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $agpr0_agpr1 + + ; CHECK-LABEL: name: assert_sext_agpr_64 + ; CHECK: liveins: $agpr0_agpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:areg_64(s64) = COPY $agpr0_agpr1 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s64) = COPY %copy(s64) + ; CHECK-NEXT: [[ASSERT_SEXT:%[0-9]+]]:vgpr(s64) = G_ASSERT_SEXT [[COPY]], 4 + ; CHECK-NEXT: %assert_sext:areg_64(s64) = COPY [[ASSERT_SEXT]](s64) + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_sext(s64) + %copy:areg_64(s64) = COPY $agpr0_agpr1 + %assert_sext:areg_64(s64) = G_ASSERT_SEXT %copy, 4 + S_ENDPGM 0, implicit %assert_sext +... diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-assert-zext.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-assert-zext.mir index 0bce908969c1..c64a8ec8ae6d 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-assert-zext.mir +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-assert-zext.mir @@ -1,5 +1,5 @@ # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py -# RUN: llc -mtriple=amdgcn -mcpu=gfx90a -run-pass=regbankselect %s -verify-machineinstrs -o - | FileCheck %s +# RUN: llc -mtriple=amdgcn -mcpu=gfx90a -run-pass=amdgpu-regbankselect,amdgpu-regbanklegalize %s -verify-machineinstrs -o - | FileCheck %s --- name: assert_zext_vgpr @@ -53,8 +53,8 @@ body: | ; CHECK-LABEL: name: assert_zext_agpr ; CHECK: liveins: $agpr0 ; CHECK-NEXT: {{ $}} - ; CHECK-NEXT: %copy:agpr(s32) = COPY $agpr0 - ; CHECK-NEXT: %assert_zext:agpr(s32) = G_ASSERT_ZEXT %copy, 4 + ; CHECK-NEXT: %copy:vgpr(s32) = COPY $agpr0 + ; CHECK-NEXT: %assert_zext:vgpr(s32) = G_ASSERT_ZEXT %copy, 4 ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_zext(s32) %copy:_(s32) = COPY $agpr0 %assert_zext:_(s32) = G_ASSERT_ZEXT %copy, 4 @@ -74,7 +74,8 @@ body: | ; CHECK: liveins: $vgpr0 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: %copy:vgpr_32(s32) = COPY $vgpr0 - ; CHECK-NEXT: %assert_zext:vgpr(s32) = G_ASSERT_ZEXT %copy, 4 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s32) = COPY %copy(s32) + ; CHECK-NEXT: %assert_zext:vgpr(s32) = G_ASSERT_ZEXT [[COPY]], 4 ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_zext(s32) %copy:vgpr_32(s32) = COPY $vgpr0 %assert_zext:_(s32) = G_ASSERT_ZEXT %copy, 4 @@ -94,9 +95,76 @@ body: | ; CHECK: liveins: $sgpr8 ; CHECK-NEXT: {{ $}} ; CHECK-NEXT: %copy:sgpr_32(s32) = COPY $sgpr8 - ; CHECK-NEXT: %assert_zext:sgpr(s32) = G_ASSERT_ZEXT %copy, 4 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY %copy(s32) + ; CHECK-NEXT: %assert_zext:sgpr(s32) = G_ASSERT_ZEXT [[COPY]], 4 ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_zext(s32) %copy:sgpr_32(s32) = COPY $sgpr8 %assert_zext:_(s32) = G_ASSERT_ZEXT %copy, 4 S_ENDPGM 0, implicit %assert_zext ... + +--- +name: assert_zext_vgpr_64 +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $vgpr0_vgpr1 + + ; CHECK-LABEL: name: assert_zext_vgpr_64 + ; CHECK: liveins: $vgpr0_vgpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:vreg_64(s64) = COPY $vgpr0_vgpr1 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s64) = COPY %copy(s64) + ; CHECK-NEXT: [[ASSERT_ZEXT:%[0-9]+]]:vgpr(s64) = G_ASSERT_ZEXT [[COPY]], 4 + ; CHECK-NEXT: %assert_zext:vreg_64(s64) = COPY [[ASSERT_ZEXT]](s64) + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_zext(s64) + %copy:vreg_64(s64) = COPY $vgpr0_vgpr1 + %assert_zext:vreg_64(s64) = G_ASSERT_ZEXT %copy, 4 + S_ENDPGM 0, implicit %assert_zext +... + +--- +name: assert_zext_sgpr_64 +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $sgpr0_sgpr1 + + ; CHECK-LABEL: name: assert_zext_sgpr_64 + ; CHECK: liveins: $sgpr0_sgpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:sreg_64(s64) = COPY $sgpr0_sgpr1 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s64) = COPY %copy(s64) + ; CHECK-NEXT: [[ASSERT_ZEXT:%[0-9]+]]:sgpr(s64) = G_ASSERT_ZEXT [[COPY]], 4 + ; CHECK-NEXT: %assert_zext:sreg_64(s64) = COPY [[ASSERT_ZEXT]](s64) + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_zext(s64) + %copy:sreg_64(s64) = COPY $sgpr0_sgpr1 + %assert_zext:sreg_64(s64) = G_ASSERT_ZEXT %copy, 4 + S_ENDPGM 0, implicit %assert_zext +... + +--- +name: assert_zext_agpr_64 +alignment: 4 +legalized: true +tracksRegLiveness: true +body: | + bb.0: + liveins: $agpr0_agpr1 + + ; CHECK-LABEL: name: assert_zext_agpr_64 + ; CHECK: liveins: $agpr0_agpr1 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: %copy:areg_64(s64) = COPY $agpr0_agpr1 + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vgpr(s64) = COPY %copy(s64) + ; CHECK-NEXT: [[ASSERT_ZEXT:%[0-9]+]]:vgpr(s64) = G_ASSERT_ZEXT [[COPY]], 4 + ; CHECK-NEXT: %assert_zext:areg_64(s64) = COPY [[ASSERT_ZEXT]](s64) + ; CHECK-NEXT: S_ENDPGM 0, implicit %assert_zext(s64) + %copy:areg_64(s64) = COPY $agpr0_agpr1 + %assert_zext:areg_64(s64) = G_ASSERT_ZEXT %copy, 4 + S_ENDPGM 0, implicit %assert_zext +... diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-smax.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-smax.mir index eee553e4e872..a7023d0ecb50 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-smax.mir +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-smax.mir @@ -1,6 +1,5 @@ # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py -# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=regbankselect -verify-machineinstrs -regbankselect-fast -o - %s | FileCheck %s -# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=regbankselect -verify-machineinstrs -regbankselect-greedy -o - %s | FileCheck %s +# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass='amdgpu-regbankselect,amdgpu-regbanklegalize' -o - %s | FileCheck %s --- name: smax_s32_ss @@ -188,8 +187,7 @@ body: | ; CHECK-NEXT: [[ASHR:%[0-9]+]]:sgpr(s32) = G_ASHR [[BITCAST]], [[C]](s32) ; CHECK-NEXT: [[BITCAST1:%[0-9]+]]:sgpr(s32) = G_BITCAST [[COPY1]](<2 x s16>) ; CHECK-NEXT: [[SEXT_INREG1:%[0-9]+]]:sgpr(s32) = G_SEXT_INREG [[BITCAST1]], 16 - ; CHECK-NEXT: [[C1:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 16 - ; CHECK-NEXT: [[ASHR1:%[0-9]+]]:sgpr(s32) = G_ASHR [[BITCAST1]], [[C1]](s32) + ; CHECK-NEXT: [[ASHR1:%[0-9]+]]:sgpr(s32) = G_ASHR [[BITCAST1]], [[C]](s32) ; CHECK-NEXT: [[SMAX:%[0-9]+]]:sgpr(s32) = G_SMAX [[SEXT_INREG]], [[SEXT_INREG1]] ; CHECK-NEXT: [[SMAX1:%[0-9]+]]:sgpr(s32) = G_SMAX [[ASHR]], [[ASHR1]] ; CHECK-NEXT: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[SMAX]](s32), [[SMAX1]](s32) diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-smin.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-smin.mir index ef60aa81e492..9dd5f45d2607 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-smin.mir +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-smin.mir @@ -1,6 +1,5 @@ # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py -# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=regbankselect -verify-machineinstrs -regbankselect-fast -o - %s | FileCheck %s -# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=regbankselect -verify-machineinstrs -regbankselect-greedy -o - %s | FileCheck %s +# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass='amdgpu-regbankselect,amdgpu-regbanklegalize' -o - %s | FileCheck %s --- name: smin_s32_ss @@ -191,8 +190,7 @@ body: | ; CHECK-NEXT: [[ASHR:%[0-9]+]]:sgpr(s32) = G_ASHR [[BITCAST]], [[C]](s32) ; CHECK-NEXT: [[BITCAST1:%[0-9]+]]:sgpr(s32) = G_BITCAST [[COPY1]](<2 x s16>) ; CHECK-NEXT: [[SEXT_INREG1:%[0-9]+]]:sgpr(s32) = G_SEXT_INREG [[BITCAST1]], 16 - ; CHECK-NEXT: [[C1:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 16 - ; CHECK-NEXT: [[ASHR1:%[0-9]+]]:sgpr(s32) = G_ASHR [[BITCAST1]], [[C1]](s32) + ; CHECK-NEXT: [[ASHR1:%[0-9]+]]:sgpr(s32) = G_ASHR [[BITCAST1]], [[C]](s32) ; CHECK-NEXT: [[SMIN:%[0-9]+]]:sgpr(s32) = G_SMIN [[SEXT_INREG]], [[SEXT_INREG1]] ; CHECK-NEXT: [[SMIN1:%[0-9]+]]:sgpr(s32) = G_SMIN [[ASHR]], [[ASHR1]] ; CHECK-NEXT: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[SMIN]](s32), [[SMIN1]](s32) diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-umax.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-umax.mir index 36a38aac1cca..59d7dcea9f2d 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-umax.mir +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-umax.mir @@ -1,6 +1,5 @@ # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py -# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=regbankselect -verify-machineinstrs -regbankselect-fast -o - %s | FileCheck %s -# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=regbankselect -verify-machineinstrs -regbankselect-greedy -o - %s | FileCheck %s +# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass='amdgpu-regbankselect,amdgpu-regbanklegalize' -o - %s | FileCheck %s --- name: umax_s32_ss @@ -186,15 +185,13 @@ body: | ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr1 ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:sgpr(s32) = G_BITCAST [[COPY]](<2 x s16>) - ; CHECK-NEXT: [[C:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 16 - ; CHECK-NEXT: [[LSHR:%[0-9]+]]:sgpr(s32) = G_LSHR [[BITCAST]], [[C]](s32) - ; CHECK-NEXT: [[C1:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 65535 - ; CHECK-NEXT: [[AND:%[0-9]+]]:sgpr(s32) = G_AND [[BITCAST]], [[C1]] + ; CHECK-NEXT: [[C:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 65535 + ; CHECK-NEXT: [[AND:%[0-9]+]]:sgpr(s32) = G_AND [[BITCAST]], [[C]] + ; CHECK-NEXT: [[C1:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 16 + ; CHECK-NEXT: [[LSHR:%[0-9]+]]:sgpr(s32) = G_LSHR [[BITCAST]], [[C1]](s32) ; CHECK-NEXT: [[BITCAST1:%[0-9]+]]:sgpr(s32) = G_BITCAST [[COPY1]](<2 x s16>) - ; CHECK-NEXT: [[C2:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 16 - ; CHECK-NEXT: [[LSHR1:%[0-9]+]]:sgpr(s32) = G_LSHR [[BITCAST1]], [[C2]](s32) - ; CHECK-NEXT: [[C3:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 65535 - ; CHECK-NEXT: [[AND1:%[0-9]+]]:sgpr(s32) = G_AND [[BITCAST1]], [[C3]] + ; CHECK-NEXT: [[AND1:%[0-9]+]]:sgpr(s32) = G_AND [[BITCAST1]], [[C]] + ; CHECK-NEXT: [[LSHR1:%[0-9]+]]:sgpr(s32) = G_LSHR [[BITCAST1]], [[C1]](s32) ; CHECK-NEXT: [[UMAX:%[0-9]+]]:sgpr(s32) = G_UMAX [[AND]], [[AND1]] ; CHECK-NEXT: [[UMAX1:%[0-9]+]]:sgpr(s32) = G_UMAX [[LSHR]], [[LSHR1]] ; CHECK-NEXT: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[UMAX]](s32), [[UMAX1]](s32) diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-umin.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-umin.mir index bb232b5e0765..fdb05f60e48e 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-umin.mir +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-umin.mir @@ -1,6 +1,5 @@ # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py -# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=regbankselect -verify-machineinstrs -regbankselect-fast -o - %s | FileCheck %s -# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=regbankselect -verify-machineinstrs -regbankselect-greedy -o - %s | FileCheck %s +# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass='amdgpu-regbankselect,amdgpu-regbanklegalize' -o - %s | FileCheck %s --- name: umin_s32_ss @@ -190,15 +189,13 @@ body: | ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr0 ; CHECK-NEXT: [[COPY1:%[0-9]+]]:sgpr(<2 x s16>) = COPY $sgpr1 ; CHECK-NEXT: [[BITCAST:%[0-9]+]]:sgpr(s32) = G_BITCAST [[COPY]](<2 x s16>) - ; CHECK-NEXT: [[C:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 16 - ; CHECK-NEXT: [[LSHR:%[0-9]+]]:sgpr(s32) = G_LSHR [[BITCAST]], [[C]](s32) - ; CHECK-NEXT: [[C1:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 65535 - ; CHECK-NEXT: [[AND:%[0-9]+]]:sgpr(s32) = G_AND [[BITCAST]], [[C1]] + ; CHECK-NEXT: [[C:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 65535 + ; CHECK-NEXT: [[AND:%[0-9]+]]:sgpr(s32) = G_AND [[BITCAST]], [[C]] + ; CHECK-NEXT: [[C1:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 16 + ; CHECK-NEXT: [[LSHR:%[0-9]+]]:sgpr(s32) = G_LSHR [[BITCAST]], [[C1]](s32) ; CHECK-NEXT: [[BITCAST1:%[0-9]+]]:sgpr(s32) = G_BITCAST [[COPY1]](<2 x s16>) - ; CHECK-NEXT: [[C2:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 16 - ; CHECK-NEXT: [[LSHR1:%[0-9]+]]:sgpr(s32) = G_LSHR [[BITCAST1]], [[C2]](s32) - ; CHECK-NEXT: [[C3:%[0-9]+]]:sgpr(s32) = G_CONSTANT i32 65535 - ; CHECK-NEXT: [[AND1:%[0-9]+]]:sgpr(s32) = G_AND [[BITCAST1]], [[C3]] + ; CHECK-NEXT: [[AND1:%[0-9]+]]:sgpr(s32) = G_AND [[BITCAST1]], [[C]] + ; CHECK-NEXT: [[LSHR1:%[0-9]+]]:sgpr(s32) = G_LSHR [[BITCAST1]], [[C1]](s32) ; CHECK-NEXT: [[UMIN:%[0-9]+]]:sgpr(s32) = G_UMIN [[AND]], [[AND1]] ; CHECK-NEXT: [[UMIN1:%[0-9]+]]:sgpr(s32) = G_UMIN [[LSHR]], [[LSHR1]] ; CHECK-NEXT: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:sgpr(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[UMIN]](s32), [[UMIN1]](s32) diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/sdiv.i64.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/sdiv.i64.ll index 9ffc565d9d47..4f2c454e1335 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/sdiv.i64.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/sdiv.i64.ll @@ -2537,202 +2537,195 @@ define <2 x i64> @v_sdiv_v2i64_24bit(<2 x i64> %num, <2 x i64> %den) { ; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GISEL-NEXT: v_and_b32_e32 v1, 0xffffff, v4 ; GISEL-NEXT: v_cvt_f32_u32_e32 v3, v1 -; GISEL-NEXT: v_cvt_f32_ubyte0_e32 v5, 0 -; GISEL-NEXT: v_sub_i32_e32 v10, vcc, 0, v1 -; GISEL-NEXT: v_mac_f32_e32 v3, 0x4f800000, v5 +; GISEL-NEXT: v_cvt_f32_ubyte0_e32 v9, 0 +; GISEL-NEXT: v_sub_i32_e32 v8, vcc, 0, v1 +; GISEL-NEXT: v_mac_f32_e32 v3, 0x4f800000, v9 ; GISEL-NEXT: v_rcp_iflag_f32_e32 v3, v3 -; GISEL-NEXT: v_subb_u32_e64 v11, s[4:5], 0, 0, vcc +; GISEL-NEXT: v_subb_u32_e64 v10, s[4:5], 0, 0, vcc ; GISEL-NEXT: v_mul_f32_e32 v3, 0x5f7ffffc, v3 ; GISEL-NEXT: v_mul_f32_e32 v4, 0x2f800000, v3 -; GISEL-NEXT: v_trunc_f32_e32 v7, v4 -; GISEL-NEXT: v_mac_f32_e32 v3, 0xcf800000, v7 -; GISEL-NEXT: v_cvt_u32_f32_e32 v9, v3 -; GISEL-NEXT: v_cvt_u32_f32_e32 v12, v7 -; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v10, v9, 0 -; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v10, v12, v[4:5] -; GISEL-NEXT: v_mul_lo_u32 v4, v12, v3 -; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v11, v9, v[7:8] -; GISEL-NEXT: v_mul_hi_u32 v8, v9, v3 -; GISEL-NEXT: v_mul_hi_u32 v3, v12, v3 -; GISEL-NEXT: v_mul_lo_u32 v13, v9, v7 -; GISEL-NEXT: v_mul_lo_u32 v14, v12, v7 -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v13 +; GISEL-NEXT: v_trunc_f32_e32 v5, v4 +; GISEL-NEXT: v_mac_f32_e32 v3, 0xcf800000, v5 +; GISEL-NEXT: v_cvt_u32_f32_e32 v7, v3 +; GISEL-NEXT: v_cvt_u32_f32_e32 v11, v5 +; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v8, v7, 0 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v8, v11, v[4:5] +; GISEL-NEXT: v_mul_hi_u32 v12, v7, v3 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v10, v7, v[4:5] +; GISEL-NEXT: v_mul_lo_u32 v5, v11, v3 +; GISEL-NEXT: v_mul_hi_u32 v3, v11, v3 +; GISEL-NEXT: v_mul_lo_u32 v13, v7, v4 +; GISEL-NEXT: v_mul_lo_u32 v14, v11, v4 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v13 ; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v8 -; GISEL-NEXT: v_mul_hi_u32 v8, v9, v7 -; GISEL-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v13, v4 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v12 +; GISEL-NEXT: v_mul_hi_u32 v12, v7, v4 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v13, v5 ; GISEL-NEXT: v_add_i32_e32 v3, vcc, v14, v3 ; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v8 -; GISEL-NEXT: v_cndmask_b32_e64 v8, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v8, vcc, v13, v8 -; GISEL-NEXT: v_mul_hi_u32 v7, v12, v7 -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v4 -; GISEL-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v8, v4 -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v7, v4 -; GISEL-NEXT: v_add_i32_e32 v9, vcc, v9, v3 -; GISEL-NEXT: v_addc_u32_e32 v12, vcc, v12, v4, vcc -; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v10, v9, 0 -; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v10, v12, v[4:5] -; GISEL-NEXT: v_mul_lo_u32 v4, v12, v3 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v12 +; GISEL-NEXT: v_cndmask_b32_e64 v12, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v12, vcc, v13, v12 +; GISEL-NEXT: v_mul_hi_u32 v4, v11, v4 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v12, v5 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v5 +; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v3 +; GISEL-NEXT: v_addc_u32_e32 v11, vcc, v11, v4, vcc +; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v8, v7, 0 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v8, v11, v[4:5] +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v10, v7, v[4:5] +; GISEL-NEXT: v_mul_lo_u32 v5, v11, v3 ; GISEL-NEXT: v_and_b32_e32 v10, 0xffffff, v0 -; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v11, v9, v[7:8] -; GISEL-NEXT: v_mul_hi_u32 v0, v9, v3 -; GISEL-NEXT: v_mul_hi_u32 v3, v12, v3 -; GISEL-NEXT: v_mul_lo_u32 v8, v9, v7 -; GISEL-NEXT: v_and_b32_e32 v11, 0xffffff, v2 -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v8 +; GISEL-NEXT: v_mul_lo_u32 v8, v7, v4 +; GISEL-NEXT: v_mul_hi_u32 v0, v7, v3 +; GISEL-NEXT: v_mul_hi_u32 v3, v11, v3 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v8 ; GISEL-NEXT: v_cndmask_b32_e64 v8, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0 +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v5, v0 ; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc -; GISEL-NEXT: v_mul_lo_u32 v4, v12, v7 +; GISEL-NEXT: v_mul_lo_u32 v5, v11, v4 ; GISEL-NEXT: v_add_i32_e32 v0, vcc, v8, v0 -; GISEL-NEXT: v_mul_hi_u32 v8, v9, v7 -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v4, v3 -; GISEL-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc +; GISEL-NEXT: v_mul_hi_u32 v8, v7, v4 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v5, v3 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc ; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v8 ; GISEL-NEXT: v_cndmask_b32_e64 v8, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v8 -; GISEL-NEXT: v_mul_hi_u32 v7, v12, v7 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v8 +; GISEL-NEXT: v_mul_hi_u32 v4, v11, v4 ; GISEL-NEXT: v_add_i32_e32 v0, vcc, v3, v0 ; GISEL-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v5, v3 ; GISEL-NEXT: v_add_i32_e32 v3, vcc, v4, v3 -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v7, v3 -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v9, v0 -; GISEL-NEXT: v_addc_u32_e32 v4, vcc, v12, v3, vcc -; GISEL-NEXT: v_mul_lo_u32 v7, 0, v0 -; GISEL-NEXT: v_mul_lo_u32 v8, v10, v4 -; GISEL-NEXT: v_and_b32_e32 v3, 0xffffff, v6 -; GISEL-NEXT: v_mul_hi_u32 v6, v10, v0 +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v7, v0 +; GISEL-NEXT: v_addc_u32_e32 v3, vcc, v11, v3, vcc +; GISEL-NEXT: v_mul_lo_u32 v4, 0, v0 +; GISEL-NEXT: v_mul_lo_u32 v5, v10, v3 +; GISEL-NEXT: v_mul_hi_u32 v7, v10, v0 ; GISEL-NEXT: v_mul_hi_u32 v0, 0, v0 -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v8 -; GISEL-NEXT: v_mul_lo_u32 v8, 0, v4 -; GISEL-NEXT: v_add_i32_e32 v6, vcc, v7, v6 -; GISEL-NEXT: v_mul_hi_u32 v7, v10, v4 -; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v8, v0 +; GISEL-NEXT: v_and_b32_e32 v11, 0xffffff, v2 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v5 +; GISEL-NEXT: v_mul_lo_u32 v5, 0, v3 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v7 +; GISEL-NEXT: v_mul_hi_u32 v7, v10, v3 +; GISEL-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v5, v0 ; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v7 -; GISEL-NEXT: v_cndmask_b32_e64 v8, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v6 -; GISEL-NEXT: v_mul_hi_u32 v4, 0, v4 -; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, vcc -; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v1, v0, 0 -; GISEL-NEXT: v_add_i32_e32 v8, vcc, v8, v9 -; GISEL-NEXT: v_cvt_f32_u32_e32 v9, v3 -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v8 -; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v1, v4, v[7:8] -; GISEL-NEXT: v_mac_f32_e32 v9, 0x4f800000, v5 -; GISEL-NEXT: v_rcp_iflag_f32_e32 v2, v9 -; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], 0, v0, v[7:8] -; GISEL-NEXT: v_sub_i32_e32 v8, vcc, v10, v6 +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v4 +; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v1, v0, 0 +; GISEL-NEXT: v_mul_hi_u32 v4, 0, v3 +; GISEL-NEXT: v_and_b32_e32 v3, 0xffffff, v6 +; GISEL-NEXT: v_mov_b32_e32 v5, v8 +; GISEL-NEXT: v_cvt_f32_u32_e32 v8, v3 +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v1, v4, v[5:6] +; GISEL-NEXT: v_mac_f32_e32 v8, 0x4f800000, v9 +; GISEL-NEXT: v_rcp_iflag_f32_e32 v2, v8 +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], 0, v0, v[5:6] +; GISEL-NEXT: v_sub_i32_e32 v9, vcc, v10, v7 ; GISEL-NEXT: v_mul_f32_e32 v2, 0x5f7ffffc, v2 -; GISEL-NEXT: v_mul_f32_e32 v5, 0x2f800000, v2 -; GISEL-NEXT: v_subb_u32_e64 v9, s[4:5], 0, v7, vcc -; GISEL-NEXT: v_sub_i32_e64 v10, s[4:5], 0, v7 -; GISEL-NEXT: v_trunc_f32_e32 v7, v5 -; GISEL-NEXT: v_mac_f32_e32 v2, 0xcf800000, v7 +; GISEL-NEXT: v_mul_f32_e32 v6, 0x2f800000, v2 +; GISEL-NEXT: v_trunc_f32_e32 v8, v6 +; GISEL-NEXT: v_mac_f32_e32 v2, 0xcf800000, v8 ; GISEL-NEXT: v_cvt_u32_f32_e32 v12, v2 +; GISEL-NEXT: v_subb_u32_e64 v10, s[4:5], 0, v5, vcc ; GISEL-NEXT: v_sub_i32_e64 v13, s[4:5], 0, v3 ; GISEL-NEXT: v_subb_u32_e64 v14, s[4:5], 0, 0, s[4:5] -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v13, v12, 0 -; GISEL-NEXT: v_cvt_u32_f32_e32 v15, v7 -; GISEL-NEXT: v_cmp_ge_u32_e64 s[4:5], v8, v1 -; GISEL-NEXT: v_mov_b32_e32 v2, v6 -; GISEL-NEXT: v_cndmask_b32_e64 v16, 0, -1, s[4:5] -; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v13, v15, v[2:3] -; GISEL-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v9 -; GISEL-NEXT: v_cndmask_b32_e64 v9, -1, v16, s[4:5] -; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v14, v12, v[6:7] -; GISEL-NEXT: v_subbrev_u32_e32 v2, vcc, 0, v10, vcc -; GISEL-NEXT: v_mul_lo_u32 v7, v15, v5 -; GISEL-NEXT: v_mul_lo_u32 v10, v12, v6 -; GISEL-NEXT: v_sub_i32_e32 v8, vcc, v8, v1 -; GISEL-NEXT: v_subbrev_u32_e32 v16, vcc, 0, v2, vcc -; GISEL-NEXT: v_mul_hi_u32 v2, v12, v5 -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v10 +; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v13, v12, 0 +; GISEL-NEXT: v_cvt_u32_f32_e32 v15, v8 +; GISEL-NEXT: v_sub_i32_e64 v5, s[4:5], 0, v5 +; GISEL-NEXT: v_mov_b32_e32 v2, v7 +; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v13, v15, v[2:3] +; GISEL-NEXT: v_cmp_ge_u32_e64 s[4:5], v9, v1 +; GISEL-NEXT: v_cndmask_b32_e64 v2, 0, -1, s[4:5] +; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v14, v12, v[7:8] +; GISEL-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v10 +; GISEL-NEXT: v_cndmask_b32_e64 v8, -1, v2, s[4:5] +; GISEL-NEXT: v_mul_lo_u32 v2, v15, v6 +; GISEL-NEXT: v_mul_lo_u32 v10, v12, v7 +; GISEL-NEXT: v_subbrev_u32_e32 v16, vcc, 0, v5, vcc +; GISEL-NEXT: v_mul_hi_u32 v5, v12, v6 +; GISEL-NEXT: v_add_i32_e32 v2, vcc, v2, v10 ; GISEL-NEXT: v_cndmask_b32_e64 v10, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v2, vcc, v7, v2 +; GISEL-NEXT: v_add_i32_e32 v2, vcc, v2, v5 ; GISEL-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc -; GISEL-NEXT: v_mul_lo_u32 v7, v15, v6 -; GISEL-NEXT: v_mul_hi_u32 v5, v15, v5 +; GISEL-NEXT: v_mul_lo_u32 v5, v15, v7 +; GISEL-NEXT: v_mul_hi_u32 v6, v15, v6 ; GISEL-NEXT: v_add_i32_e32 v2, vcc, v10, v2 -; GISEL-NEXT: v_mul_hi_u32 v10, v12, v6 -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v7, v5 -; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc +; GISEL-NEXT: v_mul_hi_u32 v10, v12, v7 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v6 +; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc ; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v10 ; GISEL-NEXT: v_cndmask_b32_e64 v10, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v10 -; GISEL-NEXT: v_mul_hi_u32 v6, v15, v6 +; GISEL-NEXT: v_add_i32_e32 v6, vcc, v6, v10 +; GISEL-NEXT: v_mul_hi_u32 v7, v15, v7 ; GISEL-NEXT: v_add_i32_e32 v2, vcc, v5, v2 ; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v7, v5 ; GISEL-NEXT: v_add_i32_e32 v5, vcc, v6, v5 -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v12, v2 -; GISEL-NEXT: v_addc_u32_e32 v10, vcc, v15, v5, vcc -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v13, v7, 0 -; GISEL-NEXT: v_add_i32_e32 v12, vcc, 1, v0 -; GISEL-NEXT: v_addc_u32_e32 v15, vcc, 0, v4, vcc -; GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v8, v1 -; GISEL-NEXT: v_mov_b32_e32 v1, v6 -; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v13, v10, v[1:2] -; GISEL-NEXT: v_cndmask_b32_e64 v8, 0, -1, vcc -; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v16 -; GISEL-NEXT: v_mad_u64_u32 v[1:2], s[4:5], v14, v7, v[1:2] -; GISEL-NEXT: v_cndmask_b32_e32 v6, -1, v8, vcc -; GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v12 -; GISEL-NEXT: v_addc_u32_e32 v8, vcc, 0, v15, vcc -; GISEL-NEXT: v_mul_lo_u32 v13, v10, v5 -; GISEL-NEXT: v_mul_lo_u32 v14, v7, v1 -; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6 -; GISEL-NEXT: v_mul_hi_u32 v6, v7, v5 -; GISEL-NEXT: v_cndmask_b32_e32 v2, v12, v2, vcc -; GISEL-NEXT: v_add_i32_e64 v12, s[4:5], v13, v14 -; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, 1, s[4:5] -; GISEL-NEXT: v_add_i32_e64 v6, s[4:5], v12, v6 -; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, s[4:5] -; GISEL-NEXT: v_mul_lo_u32 v12, v10, v1 -; GISEL-NEXT: v_mul_hi_u32 v5, v10, v5 -; GISEL-NEXT: v_add_i32_e64 v6, s[4:5], v13, v6 -; GISEL-NEXT: v_mul_hi_u32 v13, v7, v1 -; GISEL-NEXT: v_add_i32_e64 v5, s[4:5], v12, v5 -; GISEL-NEXT: v_cndmask_b32_e64 v12, 0, 1, s[4:5] -; GISEL-NEXT: v_add_i32_e64 v5, s[4:5], v5, v13 -; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, 1, s[4:5] -; GISEL-NEXT: v_add_i32_e64 v12, s[4:5], v12, v13 -; GISEL-NEXT: v_mul_hi_u32 v1, v10, v1 -; GISEL-NEXT: v_add_i32_e64 v5, s[4:5], v5, v6 -; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, s[4:5] -; GISEL-NEXT: v_add_i32_e64 v6, s[4:5], v12, v6 -; GISEL-NEXT: v_add_i32_e64 v1, s[4:5], v1, v6 -; GISEL-NEXT: v_add_i32_e64 v5, s[4:5], v7, v5 -; GISEL-NEXT: v_addc_u32_e64 v1, s[4:5], v10, v1, s[4:5] +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v7, v5 +; GISEL-NEXT: v_add_i32_e32 v10, vcc, v12, v2 +; GISEL-NEXT: v_addc_u32_e32 v12, vcc, v15, v5, vcc +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v13, v10, 0 +; GISEL-NEXT: v_sub_i32_e32 v9, vcc, v9, v1 +; GISEL-NEXT: v_mov_b32_e32 v2, v6 +; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v13, v12, v[2:3] +; GISEL-NEXT: v_subbrev_u32_e32 v15, vcc, 0, v16, vcc +; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v14, v10, v[6:7] +; GISEL-NEXT: v_add_i32_e32 v2, vcc, 1, v0 +; GISEL-NEXT: v_addc_u32_e32 v13, vcc, 0, v4, vcc +; GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v9, v1 +; GISEL-NEXT: v_mul_lo_u32 v7, v12, v5 +; GISEL-NEXT: v_mul_lo_u32 v9, v10, v6 +; GISEL-NEXT: v_mul_hi_u32 v14, v10, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v1, 0, -1, vcc +; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v15 +; GISEL-NEXT: v_cndmask_b32_e32 v1, -1, v1, vcc +; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v9 +; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v14 +; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc +; GISEL-NEXT: v_mul_lo_u32 v14, v12, v6 +; GISEL-NEXT: v_mul_hi_u32 v5, v12, v5 +; GISEL-NEXT: v_add_i32_e32 v7, vcc, v9, v7 +; GISEL-NEXT: v_mul_hi_u32 v9, v10, v6 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v14, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v14, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v9 +; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v9, vcc, v14, v9 +; GISEL-NEXT: v_mul_hi_u32 v6, v12, v6 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v7 +; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v7, vcc, v9, v7 +; GISEL-NEXT: v_add_i32_e32 v6, vcc, v6, v7 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v10, v5 +; GISEL-NEXT: v_addc_u32_e32 v7, vcc, v12, v6, vcc ; GISEL-NEXT: v_mul_lo_u32 v6, 0, v5 -; GISEL-NEXT: v_mul_lo_u32 v7, v11, v1 -; GISEL-NEXT: v_mul_hi_u32 v10, v11, v5 -; GISEL-NEXT: v_cndmask_b32_e32 v8, v15, v8, vcc +; GISEL-NEXT: v_mul_lo_u32 v9, v11, v7 +; GISEL-NEXT: v_mul_hi_u32 v14, v11, v5 +; GISEL-NEXT: v_add_i32_e32 v10, vcc, 1, v2 +; GISEL-NEXT: v_addc_u32_e32 v12, vcc, 0, v13, vcc +; GISEL-NEXT: v_add_i32_e32 v6, vcc, v6, v9 +; GISEL-NEXT: v_mul_lo_u32 v9, 0, v7 ; GISEL-NEXT: v_mul_hi_u32 v5, 0, v5 -; GISEL-NEXT: v_add_i32_e32 v6, vcc, v6, v7 -; GISEL-NEXT: v_mul_lo_u32 v7, 0, v1 -; GISEL-NEXT: v_add_i32_e32 v6, vcc, v6, v10 -; GISEL-NEXT: v_mul_hi_u32 v10, v11, v1 +; GISEL-NEXT: v_add_i32_e32 v6, vcc, v6, v14 +; GISEL-NEXT: v_mul_hi_u32 v14, v11, v7 ; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v7, v5 -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v10 -; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v10, vcc, v5, v6 -; GISEL-NEXT: v_mul_hi_u32 v1, 0, v1 -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v3, v10, 0 -; GISEL-NEXT: v_cndmask_b32_e64 v12, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v12 -; GISEL-NEXT: v_add_i32_e32 v12, vcc, v1, v7 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v9, v5 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v14 +; GISEL-NEXT: v_add_i32_e32 v9, vcc, v5, v6 +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v3, v9, 0 +; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1 +; GISEL-NEXT: v_cndmask_b32_e32 v2, v2, v10, vcc +; GISEL-NEXT: v_mul_hi_u32 v10, 0, v7 ; GISEL-NEXT: v_mov_b32_e32 v1, v6 -; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v3, v12, v[1:2] -; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9 +; GISEL-NEXT: v_cndmask_b32_e32 v12, v13, v12, vcc +; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v8 +; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v3, v10, v[1:2] ; GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc -; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], 0, v10, v[6:7] -; GISEL-NEXT: v_cndmask_b32_e32 v1, v4, v8, vcc +; GISEL-NEXT: v_cndmask_b32_e32 v1, v4, v12, vcc +; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], 0, v9, v[6:7] ; GISEL-NEXT: v_sub_i32_e32 v2, vcc, v11, v5 ; GISEL-NEXT: v_sub_i32_e64 v5, s[4:5], 0, v6 ; GISEL-NEXT: v_subb_u32_e64 v4, s[4:5], 0, v6, vcc @@ -2743,8 +2736,8 @@ define <2 x i64> @v_sdiv_v2i64_24bit(<2 x i64> %num, <2 x i64> %den) { ; GISEL-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v4 ; GISEL-NEXT: v_subbrev_u32_e32 v5, vcc, 0, v5, vcc ; GISEL-NEXT: v_cndmask_b32_e64 v4, -1, v6, s[4:5] -; GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v10 -; GISEL-NEXT: v_addc_u32_e32 v7, vcc, 0, v12, vcc +; GISEL-NEXT: v_add_i32_e32 v6, vcc, 1, v9 +; GISEL-NEXT: v_addc_u32_e32 v7, vcc, 0, v10, vcc ; GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v2, v3 ; GISEL-NEXT: v_cndmask_b32_e64 v2, 0, -1, vcc ; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v5 @@ -2755,8 +2748,8 @@ define <2 x i64> @v_sdiv_v2i64_24bit(<2 x i64> %num, <2 x i64> %den) { ; GISEL-NEXT: v_cndmask_b32_e32 v2, v6, v3, vcc ; GISEL-NEXT: v_cndmask_b32_e32 v3, v7, v5, vcc ; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v4 -; GISEL-NEXT: v_cndmask_b32_e32 v2, v10, v2, vcc -; GISEL-NEXT: v_cndmask_b32_e32 v3, v12, v3, vcc +; GISEL-NEXT: v_cndmask_b32_e32 v2, v9, v2, vcc +; GISEL-NEXT: v_cndmask_b32_e32 v3, v10, v3, vcc ; GISEL-NEXT: s_setpc_b64 s[30:31] ; ; CGP-LABEL: v_sdiv_v2i64_24bit: diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/smed3.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/smed3.ll index ac1e11b0f3cd..dfa613c69733 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/smed3.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/smed3.ll @@ -1,7 +1,7 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc -global-isel -mtriple=amdgcn-amd-mesa3d -mcpu=gfx802 < %s | FileCheck -check-prefixes=GFX89,GFX8 %s -; RUN: llc -global-isel -mtriple=amdgcn-amd-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX89,GFX9 %s -; RUN: llc -global-isel -mtriple=amdgcn-amd-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-mesa3d -mcpu=gfx802 < %s | FileCheck -check-prefixes=GFX89,GFX8 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX89,GFX9 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s define i32 @test_min_max_ValK0_K1_i32(i32 %a) { ; GFX89-LABEL: test_min_max_ValK0_K1_i32: diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/srem.i64.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/srem.i64.ll index 82279e641ed6..40b5db0a1544 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/srem.i64.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/srem.i64.ll @@ -3035,203 +3035,193 @@ define <2 x i64> @v_srem_v2i64_24bit(<2 x i64> %num, <2 x i64> %den) { ; GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GISEL-NEXT: v_and_b32_e32 v1, 0xffffff, v4 ; GISEL-NEXT: v_cvt_f32_u32_e32 v3, v1 -; GISEL-NEXT: v_cvt_f32_ubyte0_e32 v4, 0 -; GISEL-NEXT: v_sub_i32_e32 v11, vcc, 0, v1 -; GISEL-NEXT: v_mac_f32_e32 v3, 0x4f800000, v4 +; GISEL-NEXT: v_cvt_f32_ubyte0_e32 v7, 0 +; GISEL-NEXT: v_sub_i32_e32 v9, vcc, 0, v1 +; GISEL-NEXT: v_mac_f32_e32 v3, 0x4f800000, v7 ; GISEL-NEXT: v_rcp_iflag_f32_e32 v3, v3 -; GISEL-NEXT: v_subb_u32_e64 v12, s[4:5], 0, 0, vcc +; GISEL-NEXT: v_subb_u32_e64 v10, s[4:5], 0, 0, vcc ; GISEL-NEXT: v_and_b32_e32 v2, 0xffffff, v2 ; GISEL-NEXT: v_mul_f32_e32 v3, 0x5f7ffffc, v3 -; GISEL-NEXT: v_mul_f32_e32 v5, 0x2f800000, v3 -; GISEL-NEXT: v_trunc_f32_e32 v5, v5 +; GISEL-NEXT: v_mul_f32_e32 v4, 0x2f800000, v3 +; GISEL-NEXT: v_trunc_f32_e32 v5, v4 ; GISEL-NEXT: v_mac_f32_e32 v3, 0xcf800000, v5 -; GISEL-NEXT: v_cvt_u32_f32_e32 v10, v3 -; GISEL-NEXT: v_cvt_u32_f32_e32 v5, v5 -; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v11, v10, 0 -; GISEL-NEXT: v_mov_b32_e32 v3, v8 -; GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v11, v5, v[3:4] -; GISEL-NEXT: v_mul_lo_u32 v3, v5, v7 -; GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v10, v[8:9] -; GISEL-NEXT: v_mul_hi_u32 v9, v10, v7 -; GISEL-NEXT: v_mul_hi_u32 v7, v5, v7 -; GISEL-NEXT: v_mul_lo_u32 v13, v10, v8 -; GISEL-NEXT: v_mul_lo_u32 v14, v5, v8 -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v13 +; GISEL-NEXT: v_cvt_u32_f32_e32 v8, v3 +; GISEL-NEXT: v_cvt_u32_f32_e32 v11, v5 +; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v8, 0 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v9, v11, v[4:5] +; GISEL-NEXT: v_mul_hi_u32 v12, v8, v3 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v10, v8, v[4:5] +; GISEL-NEXT: v_mul_lo_u32 v5, v11, v3 +; GISEL-NEXT: v_mul_hi_u32 v3, v11, v3 +; GISEL-NEXT: v_mul_lo_u32 v13, v8, v4 +; GISEL-NEXT: v_mul_lo_u32 v14, v11, v4 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v13 ; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v9 -; GISEL-NEXT: v_mul_hi_u32 v9, v10, v8 -; GISEL-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v13, v3 -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v14, v7 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v12 +; GISEL-NEXT: v_mul_hi_u32 v12, v8, v4 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v13, v5 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v14, v3 ; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v9 -; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v9, vcc, v13, v9 -; GISEL-NEXT: v_mul_hi_u32 v8, v5, v8 -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v7, v3 -; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v9, v7 -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v8, v7 -; GISEL-NEXT: v_add_i32_e32 v10, vcc, v10, v3 -; GISEL-NEXT: v_addc_u32_e32 v5, vcc, v5, v7, vcc -; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v11, v10, 0 -; GISEL-NEXT: v_mov_b32_e32 v3, v8 -; GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v11, v5, v[3:4] -; GISEL-NEXT: v_mul_lo_u32 v3, v5, v7 -; GISEL-NEXT: v_and_b32_e32 v11, 0xffffff, v0 -; GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v12, v10, v[8:9] -; GISEL-NEXT: v_mul_hi_u32 v0, v10, v7 -; GISEL-NEXT: v_mul_hi_u32 v7, v5, v7 -; GISEL-NEXT: v_mul_lo_u32 v9, v10, v8 -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v9 -; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v3, v0 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v12 +; GISEL-NEXT: v_cndmask_b32_e64 v12, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v12, vcc, v13, v12 +; GISEL-NEXT: v_mul_hi_u32 v4, v11, v4 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v12, v5 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v5 +; GISEL-NEXT: v_add_i32_e32 v8, vcc, v8, v3 +; GISEL-NEXT: v_addc_u32_e32 v11, vcc, v11, v4, vcc +; GISEL-NEXT: v_mad_u64_u32 v[3:4], s[4:5], v9, v8, 0 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v9, v11, v[4:5] +; GISEL-NEXT: v_mul_hi_u32 v9, v8, v3 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v10, v8, v[4:5] +; GISEL-NEXT: v_and_b32_e32 v10, 0xffffff, v0 +; GISEL-NEXT: v_mul_lo_u32 v0, v11, v3 +; GISEL-NEXT: v_mul_lo_u32 v5, v8, v4 +; GISEL-NEXT: v_mul_hi_u32 v3, v11, v3 +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v9 ; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc -; GISEL-NEXT: v_mul_lo_u32 v3, v5, v8 -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v9, v0 -; GISEL-NEXT: v_mul_hi_u32 v9, v10, v8 -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v7 -; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v9 +; GISEL-NEXT: v_mul_lo_u32 v9, v11, v4 +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v5, v0 +; GISEL-NEXT: v_mul_hi_u32 v5, v8, v4 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v9, v3 ; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v9 -; GISEL-NEXT: v_mul_hi_u32 v8, v5, v8 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v3, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v9, v5 +; GISEL-NEXT: v_mul_hi_u32 v4, v11, v4 ; GISEL-NEXT: v_add_i32_e32 v0, vcc, v3, v0 ; GISEL-NEXT: v_cndmask_b32_e64 v3, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v7, v3 -; GISEL-NEXT: v_add_i32_e32 v3, vcc, v8, v3 -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v10, v0 -; GISEL-NEXT: v_addc_u32_e32 v5, vcc, v5, v3, vcc -; GISEL-NEXT: v_mul_lo_u32 v7, 0, v0 -; GISEL-NEXT: v_mul_lo_u32 v8, v11, v5 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v5, v3 +; GISEL-NEXT: v_add_i32_e32 v3, vcc, v4, v3 +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v8, v0 +; GISEL-NEXT: v_addc_u32_e32 v8, vcc, v11, v3, vcc +; GISEL-NEXT: v_mul_lo_u32 v4, 0, v0 +; GISEL-NEXT: v_mul_lo_u32 v5, v10, v8 ; GISEL-NEXT: v_and_b32_e32 v3, 0xffffff, v6 -; GISEL-NEXT: v_mul_hi_u32 v6, v11, v0 +; GISEL-NEXT: v_mul_hi_u32 v6, v10, v0 ; GISEL-NEXT: v_mul_hi_u32 v0, 0, v0 -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v8 -; GISEL-NEXT: v_mul_lo_u32 v8, 0, v5 -; GISEL-NEXT: v_add_i32_e32 v6, vcc, v7, v6 -; GISEL-NEXT: v_mul_hi_u32 v7, v11, v5 -; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v8, v0 -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v7 -; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v8, vcc, v0, v6 -; GISEL-NEXT: v_mul_hi_u32 v9, 0, v5 -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v1, v8, 0 -; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v7, v0 -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v9, v0 -; GISEL-NEXT: v_mov_b32_e32 v0, v6 -; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], v1, v7, v[0:1] -; GISEL-NEXT: v_cvt_f32_u32_e32 v0, v3 -; GISEL-NEXT: v_mad_u64_u32 v[6:7], s[4:5], 0, v8, v[6:7] -; GISEL-NEXT: v_mac_f32_e32 v0, 0x4f800000, v4 -; GISEL-NEXT: v_rcp_iflag_f32_e32 v0, v0 -; GISEL-NEXT: v_sub_i32_e32 v7, vcc, v11, v5 -; GISEL-NEXT: v_subb_u32_e64 v8, s[4:5], 0, v6, vcc -; GISEL-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v0 -; GISEL-NEXT: v_mul_f32_e32 v4, 0x2f800000, v0 -; GISEL-NEXT: v_trunc_f32_e32 v9, v4 -; GISEL-NEXT: v_mac_f32_e32 v0, 0xcf800000, v9 -; GISEL-NEXT: v_cvt_u32_f32_e32 v10, v0 -; GISEL-NEXT: v_sub_i32_e64 v11, s[4:5], 0, v3 -; GISEL-NEXT: v_subb_u32_e64 v12, s[4:5], 0, 0, s[4:5] -; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v11, v10, 0 -; GISEL-NEXT: v_cvt_u32_f32_e32 v9, v9 -; GISEL-NEXT: v_sub_i32_e64 v13, s[4:5], 0, v6 -; GISEL-NEXT: v_mov_b32_e32 v0, v5 -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v11, v9, v[0:1] -; GISEL-NEXT: v_cmp_ge_u32_e64 s[4:5], v7, v1 -; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, -1, s[4:5] -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v12, v10, v[5:6] -; GISEL-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v8 -; GISEL-NEXT: v_cndmask_b32_e64 v14, -1, v0, s[4:5] -; GISEL-NEXT: v_mul_lo_u32 v0, v9, v4 -; GISEL-NEXT: v_mul_lo_u32 v6, v10, v5 -; GISEL-NEXT: v_mul_hi_u32 v15, v10, v4 -; GISEL-NEXT: v_subbrev_u32_e32 v13, vcc, 0, v13, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v6 -; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v15 -; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc -; GISEL-NEXT: v_mul_lo_u32 v15, v9, v5 -; GISEL-NEXT: v_mul_hi_u32 v4, v9, v4 -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v6, v0 -; GISEL-NEXT: v_mul_hi_u32 v6, v10, v5 -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v15, v4 -; GISEL-NEXT: v_cndmask_b32_e64 v15, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v5 +; GISEL-NEXT: v_mul_lo_u32 v5, 0, v8 ; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v6 -; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v6, vcc, v15, v6 -; GISEL-NEXT: v_mul_hi_u32 v5, v9, v5 -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v4, v0 +; GISEL-NEXT: v_mul_hi_u32 v6, v10, v8 ; GISEL-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v6, v4 -; GISEL-NEXT: v_add_i32_e32 v4, vcc, v5, v4 -; GISEL-NEXT: v_add_i32_e32 v10, vcc, v10, v0 -; GISEL-NEXT: v_addc_u32_e32 v9, vcc, v9, v4, vcc -; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v11, v10, 0 -; GISEL-NEXT: v_sub_i32_e32 v15, vcc, v7, v1 +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v5, v0 +; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v6 +; GISEL-NEXT: v_add_i32_e32 v9, vcc, v0, v4 +; GISEL-NEXT: v_cvt_f32_u32_e32 v0, v3 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v1, v9, 0 +; GISEL-NEXT: v_mul_hi_u32 v6, 0, v8 +; GISEL-NEXT: v_mac_f32_e32 v0, 0x4f800000, v7 +; GISEL-NEXT: v_rcp_iflag_f32_e32 v7, v0 ; GISEL-NEXT: v_mov_b32_e32 v0, v5 -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v11, v9, v[0:1] -; GISEL-NEXT: v_subbrev_u32_e32 v13, vcc, 0, v13, vcc -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v12, v10, v[5:6] -; GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v15, v1 -; GISEL-NEXT: v_cndmask_b32_e64 v16, 0, -1, vcc -; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v13 -; GISEL-NEXT: v_cndmask_b32_e32 v0, -1, v16, vcc -; GISEL-NEXT: v_sub_i32_e32 v1, vcc, v15, v1 -; GISEL-NEXT: v_subbrev_u32_e32 v6, vcc, 0, v13, vcc -; GISEL-NEXT: v_mul_lo_u32 v11, v9, v4 -; GISEL-NEXT: v_mul_lo_u32 v12, v10, v5 -; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v0 -; GISEL-NEXT: v_mul_hi_u32 v0, v10, v4 -; GISEL-NEXT: v_mul_hi_u32 v4, v9, v4 -; GISEL-NEXT: v_add_i32_e64 v11, s[4:5], v11, v12 -; GISEL-NEXT: v_cndmask_b32_e64 v12, 0, 1, s[4:5] -; GISEL-NEXT: v_add_i32_e64 v0, s[4:5], v11, v0 -; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[4:5] -; GISEL-NEXT: v_mul_lo_u32 v11, v9, v5 -; GISEL-NEXT: v_add_i32_e64 v0, s[4:5], v12, v0 -; GISEL-NEXT: v_mul_hi_u32 v12, v10, v5 -; GISEL-NEXT: v_add_i32_e64 v4, s[4:5], v11, v4 -; GISEL-NEXT: v_cndmask_b32_e64 v11, 0, 1, s[4:5] -; GISEL-NEXT: v_add_i32_e64 v4, s[4:5], v4, v12 -; GISEL-NEXT: v_cndmask_b32_e64 v12, 0, 1, s[4:5] -; GISEL-NEXT: v_add_i32_e64 v11, s[4:5], v11, v12 -; GISEL-NEXT: v_mul_hi_u32 v5, v9, v5 -; GISEL-NEXT: v_add_i32_e64 v0, s[4:5], v4, v0 +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v1, v6, v[0:1] +; GISEL-NEXT: v_mul_f32_e32 v0, 0x5f7ffffc, v7 +; GISEL-NEXT: v_mul_f32_e32 v7, 0x2f800000, v0 +; GISEL-NEXT: v_trunc_f32_e32 v11, v7 +; GISEL-NEXT: v_mac_f32_e32 v0, 0xcf800000, v11 +; GISEL-NEXT: v_cvt_u32_f32_e32 v12, v0 +; GISEL-NEXT: v_sub_i32_e32 v13, vcc, 0, v3 +; GISEL-NEXT: v_cvt_u32_f32_e32 v11, v11 +; GISEL-NEXT: v_mad_u64_u32 v[7:8], s[4:5], v13, v12, 0 +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], 0, v9, v[5:6] +; GISEL-NEXT: v_mov_b32_e32 v0, v8 +; GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v13, v11, v[0:1] +; GISEL-NEXT: v_subb_u32_e64 v14, s[4:5], 0, 0, vcc +; GISEL-NEXT: v_mad_u64_u32 v[8:9], s[4:5], v14, v12, v[8:9] +; GISEL-NEXT: v_sub_i32_e32 v10, vcc, v10, v4 +; GISEL-NEXT: v_subb_u32_e64 v15, s[4:5], 0, v5, vcc +; GISEL-NEXT: v_sub_i32_e64 v0, s[4:5], 0, v5 +; GISEL-NEXT: v_mul_lo_u32 v4, v11, v7 +; GISEL-NEXT: v_mul_lo_u32 v5, v12, v8 +; GISEL-NEXT: v_mul_hi_u32 v9, v12, v7 +; GISEL-NEXT: v_cmp_ge_u32_e64 s[4:5], v10, v1 +; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, -1, s[4:5] +; GISEL-NEXT: v_add_i32_e64 v4, s[4:5], v4, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, s[4:5] +; GISEL-NEXT: v_add_i32_e64 v4, s[4:5], v4, v9 ; GISEL-NEXT: v_cndmask_b32_e64 v4, 0, 1, s[4:5] -; GISEL-NEXT: v_add_i32_e64 v4, s[4:5], v11, v4 +; GISEL-NEXT: v_mul_lo_u32 v9, v11, v8 +; GISEL-NEXT: v_mul_hi_u32 v7, v11, v7 ; GISEL-NEXT: v_add_i32_e64 v4, s[4:5], v5, v4 -; GISEL-NEXT: v_add_i32_e64 v0, s[4:5], v10, v0 -; GISEL-NEXT: v_addc_u32_e64 v4, s[4:5], v9, v4, s[4:5] -; GISEL-NEXT: v_mul_lo_u32 v5, 0, v0 -; GISEL-NEXT: v_mul_lo_u32 v9, v2, v4 -; GISEL-NEXT: v_cndmask_b32_e32 v10, v13, v6, vcc -; GISEL-NEXT: v_mul_hi_u32 v6, v2, v0 -; GISEL-NEXT: v_cndmask_b32_e32 v1, v15, v1, vcc -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v9 -; GISEL-NEXT: v_mul_lo_u32 v9, 0, v4 -; GISEL-NEXT: v_mul_hi_u32 v0, 0, v0 -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v6 -; GISEL-NEXT: v_mul_hi_u32 v6, v2, v4 +; GISEL-NEXT: v_mul_hi_u32 v5, v12, v8 +; GISEL-NEXT: v_add_i32_e64 v7, s[4:5], v9, v7 +; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, s[4:5] +; GISEL-NEXT: v_add_i32_e64 v5, s[4:5], v7, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, 1, s[4:5] +; GISEL-NEXT: v_add_i32_e64 v7, s[4:5], v9, v7 +; GISEL-NEXT: v_mul_hi_u32 v8, v11, v8 +; GISEL-NEXT: v_add_i32_e64 v4, s[4:5], v5, v4 +; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, s[4:5] +; GISEL-NEXT: v_add_i32_e64 v5, s[4:5], v7, v5 +; GISEL-NEXT: v_add_i32_e64 v5, s[4:5], v8, v5 +; GISEL-NEXT: v_add_i32_e64 v7, s[4:5], v12, v4 +; GISEL-NEXT: v_addc_u32_e64 v8, s[4:5], v11, v5, s[4:5] +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v13, v7, 0 +; GISEL-NEXT: v_cmp_eq_u32_e64 s[4:5], 0, v15 +; GISEL-NEXT: v_subbrev_u32_e32 v11, vcc, 0, v0, vcc +; GISEL-NEXT: v_mov_b32_e32 v0, v5 +; GISEL-NEXT: v_cndmask_b32_e64 v9, -1, v6, s[4:5] +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v13, v8, v[0:1] +; GISEL-NEXT: v_sub_i32_e32 v0, vcc, v10, v1 +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v14, v7, v[5:6] +; GISEL-NEXT: v_mul_lo_u32 v12, v8, v4 +; GISEL-NEXT: v_subbrev_u32_e32 v11, vcc, 0, v11, vcc +; GISEL-NEXT: v_mul_lo_u32 v13, v7, v5 +; GISEL-NEXT: v_mul_hi_u32 v14, v7, v4 +; GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v0, v1 +; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, -1, vcc +; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v11 +; GISEL-NEXT: v_cndmask_b32_e32 v6, -1, v6, vcc +; GISEL-NEXT: v_add_i32_e32 v12, vcc, v12, v13 +; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v12, vcc, v12, v14 +; GISEL-NEXT: v_cndmask_b32_e64 v12, 0, 1, vcc +; GISEL-NEXT: v_mul_lo_u32 v14, v8, v5 +; GISEL-NEXT: v_mul_hi_u32 v4, v8, v4 +; GISEL-NEXT: v_add_i32_e32 v12, vcc, v13, v12 +; GISEL-NEXT: v_mul_hi_u32 v13, v7, v5 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v14, v4 +; GISEL-NEXT: v_cndmask_b32_e64 v14, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v13 +; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v13, vcc, v14, v13 +; GISEL-NEXT: v_mul_hi_u32 v5, v8, v5 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v12 +; GISEL-NEXT: v_cndmask_b32_e64 v12, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v12, vcc, v13, v12 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v12 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v7, v4 +; GISEL-NEXT: v_addc_u32_e32 v7, vcc, v8, v5, vcc +; GISEL-NEXT: v_mul_lo_u32 v5, 0, v4 +; GISEL-NEXT: v_mul_lo_u32 v8, v2, v7 +; GISEL-NEXT: v_mul_hi_u32 v13, v2, v4 +; GISEL-NEXT: v_sub_i32_e32 v1, vcc, v0, v1 +; GISEL-NEXT: v_subbrev_u32_e32 v12, vcc, 0, v11, vcc +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v8 +; GISEL-NEXT: v_mul_lo_u32 v8, 0, v7 +; GISEL-NEXT: v_mul_hi_u32 v4, 0, v4 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v13 +; GISEL-NEXT: v_mul_hi_u32 v13, v2, v7 ; GISEL-NEXT: v_cndmask_b32_e64 v5, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v9, v0 -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v0, v6 -; GISEL-NEXT: v_cndmask_b32_e64 v6, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v9, vcc, v0, v5 -; GISEL-NEXT: v_mul_hi_u32 v11, 0, v4 -; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v3, v9, 0 -; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v0, vcc, v6, v0 -; GISEL-NEXT: v_add_i32_e32 v6, vcc, v11, v0 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v8, v4 +; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v13 +; GISEL-NEXT: v_add_i32_e32 v8, vcc, v4, v5 +; GISEL-NEXT: v_mad_u64_u32 v[4:5], s[4:5], v3, v8, 0 +; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v6 +; GISEL-NEXT: v_mul_hi_u32 v6, 0, v7 +; GISEL-NEXT: v_cndmask_b32_e32 v1, v0, v1, vcc ; GISEL-NEXT: v_mov_b32_e32 v0, v5 +; GISEL-NEXT: v_cndmask_b32_e32 v7, v11, v12, vcc ; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], v3, v6, v[0:1] -; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v14 -; GISEL-NEXT: v_cndmask_b32_e32 v0, v7, v1, vcc -; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], 0, v9, v[5:6] -; GISEL-NEXT: v_cndmask_b32_e32 v1, v8, v10, vcc +; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v9 +; GISEL-NEXT: v_cndmask_b32_e32 v0, v10, v1, vcc +; GISEL-NEXT: v_mad_u64_u32 v[5:6], s[4:5], 0, v8, v[5:6] +; GISEL-NEXT: v_cndmask_b32_e32 v1, v15, v7, vcc ; GISEL-NEXT: v_sub_i32_e32 v2, vcc, v2, v4 ; GISEL-NEXT: v_subb_u32_e64 v4, s[4:5], 0, v5, vcc ; GISEL-NEXT: v_sub_i32_e64 v5, s[4:5], 0, v5 diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/udiv.i64.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/udiv.i64.ll index 4de10788a6bd..ded985ec3a1e 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/udiv.i64.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/udiv.i64.ll @@ -2053,90 +2053,82 @@ define <2 x i64> @v_udiv_v2i64_24bit(<2 x i64> %num, <2 x i64> %den) { ; GISEL-NEXT: v_mul_lo_u32 v12, v6, v2 ; GISEL-NEXT: v_mul_lo_u32 v13, 0, v2 ; GISEL-NEXT: v_mul_hi_u32 v14, v6, v2 -; GISEL-NEXT: v_mul_hi_u32 v2, 0, v2 -; GISEL-NEXT: v_mul_lo_u32 v15, v0, v5 +; GISEL-NEXT: v_mul_hi_u32 v15, 0, v2 +; GISEL-NEXT: v_mul_lo_u32 v2, v0, v5 ; GISEL-NEXT: v_mul_lo_u32 v16, 0, v5 ; GISEL-NEXT: v_mul_hi_u32 v17, v0, v5 ; GISEL-NEXT: v_mul_hi_u32 v5, 0, v5 ; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v12 ; GISEL-NEXT: v_add_i32_e32 v8, vcc, v13, v8 -; GISEL-NEXT: v_add_i32_e32 v10, vcc, v10, v15 +; GISEL-NEXT: v_mul_lo_u32 v12, v3, v15 +; GISEL-NEXT: v_add_i32_e32 v2, vcc, v10, v2 ; GISEL-NEXT: v_add_i32_e32 v7, vcc, v16, v7 +; GISEL-NEXT: v_mul_lo_u32 v10, v1, v5 ; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v9 ; GISEL-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc ; GISEL-NEXT: v_add_i32_e32 v8, vcc, v8, v14 -; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v10, vcc, v10, v11 -; GISEL-NEXT: v_cndmask_b32_e64 v10, 0, 1, vcc +; GISEL-NEXT: v_add_i32_e32 v2, vcc, v2, v11 +; GISEL-NEXT: v_cndmask_b32_e64 v2, 0, 1, vcc ; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v17 -; GISEL-NEXT: v_cndmask_b32_e64 v11, 0, 1, vcc ; GISEL-NEXT: v_add_i32_e32 v4, vcc, v8, v4 -; GISEL-NEXT: v_cndmask_b32_e64 v8, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v10 -; GISEL-NEXT: v_cndmask_b32_e64 v10, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v8, vcc, v9, v8 -; GISEL-NEXT: v_mul_lo_u32 v9, v3, v4 -; GISEL-NEXT: v_mul_lo_u32 v12, 0, v4 -; GISEL-NEXT: v_mul_hi_u32 v13, v3, v4 -; GISEL-NEXT: v_add_i32_e32 v10, vcc, v11, v10 -; GISEL-NEXT: v_mul_lo_u32 v11, v1, v7 -; GISEL-NEXT: v_mul_lo_u32 v14, 0, v7 -; GISEL-NEXT: v_mul_hi_u32 v15, v1, v7 -; GISEL-NEXT: v_add_i32_e32 v8, vcc, v2, v8 -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v10 -; GISEL-NEXT: v_mul_lo_u32 v2, v3, v8 -; GISEL-NEXT: v_add_i32_e32 v10, vcc, 1, v4 -; GISEL-NEXT: v_addc_u32_e32 v16, vcc, 0, v8, vcc -; GISEL-NEXT: v_mul_lo_u32 v17, v1, v5 -; GISEL-NEXT: v_add_i32_e32 v18, vcc, 1, v7 +; GISEL-NEXT: v_add_i32_e32 v2, vcc, v7, v2 +; GISEL-NEXT: v_mul_lo_u32 v7, v3, v4 +; GISEL-NEXT: v_mul_lo_u32 v8, 0, v4 +; GISEL-NEXT: v_mul_hi_u32 v9, v3, v4 +; GISEL-NEXT: v_add_i32_e32 v11, vcc, 1, v4 +; GISEL-NEXT: v_addc_u32_e32 v13, vcc, 0, v15, vcc +; GISEL-NEXT: v_mul_lo_u32 v14, v1, v2 +; GISEL-NEXT: v_mul_lo_u32 v16, 0, v2 +; GISEL-NEXT: v_mul_hi_u32 v17, v1, v2 +; GISEL-NEXT: v_add_i32_e32 v18, vcc, 1, v2 ; GISEL-NEXT: v_addc_u32_e32 v19, vcc, 0, v5, vcc -; GISEL-NEXT: v_add_i32_e32 v2, vcc, v12, v2 -; GISEL-NEXT: v_add_i32_e32 v12, vcc, v14, v17 -; GISEL-NEXT: v_add_i32_e32 v14, vcc, 1, v10 -; GISEL-NEXT: v_addc_u32_e32 v17, vcc, 0, v16, vcc -; GISEL-NEXT: v_add_i32_e32 v2, vcc, v2, v13 -; GISEL-NEXT: v_add_i32_e32 v12, vcc, v12, v15 -; GISEL-NEXT: v_sub_i32_e32 v6, vcc, v6, v9 -; GISEL-NEXT: v_subb_u32_e64 v9, s[4:5], 0, v2, vcc +; GISEL-NEXT: v_add_i32_e32 v8, vcc, v8, v12 +; GISEL-NEXT: v_add_i32_e32 v10, vcc, v16, v10 +; GISEL-NEXT: v_add_i32_e32 v12, vcc, 1, v11 +; GISEL-NEXT: v_addc_u32_e32 v16, vcc, 0, v13, vcc +; GISEL-NEXT: v_add_i32_e32 v8, vcc, v8, v9 +; GISEL-NEXT: v_add_i32_e32 v9, vcc, v10, v17 +; GISEL-NEXT: v_sub_i32_e32 v6, vcc, v6, v7 +; GISEL-NEXT: v_subb_u32_e64 v7, s[4:5], 0, v8, vcc ; GISEL-NEXT: v_cmp_ge_u32_e64 s[4:5], v6, v3 -; GISEL-NEXT: v_cndmask_b32_e64 v13, 0, -1, s[4:5] -; GISEL-NEXT: v_sub_i32_e64 v0, s[4:5], v0, v11 -; GISEL-NEXT: v_subb_u32_e64 v11, s[6:7], 0, v12, s[4:5] +; GISEL-NEXT: v_cndmask_b32_e64 v10, 0, -1, s[4:5] +; GISEL-NEXT: v_sub_i32_e64 v0, s[4:5], v0, v14 +; GISEL-NEXT: v_subb_u32_e64 v14, s[6:7], 0, v9, s[4:5] ; GISEL-NEXT: v_cmp_ge_u32_e64 s[6:7], v0, v1 -; GISEL-NEXT: v_cndmask_b32_e64 v15, 0, -1, s[6:7] -; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v9 -; GISEL-NEXT: v_cmp_eq_u32_e64 s[8:9], 0, v11 -; GISEL-NEXT: v_add_i32_e64 v9, s[10:11], 1, v18 -; GISEL-NEXT: v_addc_u32_e64 v11, s[10:11], 0, v19, s[10:11] -; GISEL-NEXT: v_sub_i32_e64 v2, s[10:11], 0, v2 -; GISEL-NEXT: v_sub_i32_e64 v12, s[10:11], 0, v12 -; GISEL-NEXT: v_cndmask_b32_e64 v13, -1, v13, s[6:7] -; GISEL-NEXT: v_subbrev_u32_e32 v2, vcc, 0, v2, vcc -; GISEL-NEXT: v_cndmask_b32_e64 v15, -1, v15, s[8:9] -; GISEL-NEXT: v_subbrev_u32_e64 v12, vcc, 0, v12, s[4:5] +; GISEL-NEXT: v_cndmask_b32_e64 v17, 0, -1, s[6:7] +; GISEL-NEXT: v_cmp_eq_u32_e64 s[6:7], 0, v7 +; GISEL-NEXT: v_cmp_eq_u32_e64 s[8:9], 0, v14 +; GISEL-NEXT: v_add_i32_e64 v7, s[10:11], 1, v18 +; GISEL-NEXT: v_addc_u32_e64 v14, s[10:11], 0, v19, s[10:11] +; GISEL-NEXT: v_sub_i32_e64 v8, s[10:11], 0, v8 +; GISEL-NEXT: v_sub_i32_e64 v9, s[10:11], 0, v9 +; GISEL-NEXT: v_cndmask_b32_e64 v10, -1, v10, s[6:7] +; GISEL-NEXT: v_subbrev_u32_e32 v8, vcc, 0, v8, vcc +; GISEL-NEXT: v_cndmask_b32_e64 v17, -1, v17, s[8:9] +; GISEL-NEXT: v_subbrev_u32_e64 v9, vcc, 0, v9, s[4:5] ; GISEL-NEXT: v_sub_i32_e32 v6, vcc, v6, v3 -; GISEL-NEXT: v_subbrev_u32_e32 v2, vcc, 0, v2, vcc +; GISEL-NEXT: v_subbrev_u32_e32 v8, vcc, 0, v8, vcc ; GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v6, v3 ; GISEL-NEXT: v_cndmask_b32_e64 v3, 0, -1, vcc ; GISEL-NEXT: v_sub_i32_e32 v0, vcc, v0, v1 -; GISEL-NEXT: v_subbrev_u32_e32 v6, vcc, 0, v12, vcc +; GISEL-NEXT: v_subbrev_u32_e32 v6, vcc, 0, v9, vcc ; GISEL-NEXT: v_cmp_ge_u32_e32 vcc, v0, v1 ; GISEL-NEXT: v_cndmask_b32_e64 v0, 0, -1, vcc -; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v2 +; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v8 ; GISEL-NEXT: v_cndmask_b32_e32 v1, -1, v3, vcc ; GISEL-NEXT: v_cmp_eq_u32_e32 vcc, 0, v6 ; GISEL-NEXT: v_cndmask_b32_e32 v0, -1, v0, vcc ; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v1 -; GISEL-NEXT: v_cndmask_b32_e32 v1, v10, v14, vcc +; GISEL-NEXT: v_cndmask_b32_e32 v1, v11, v12, vcc ; GISEL-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v0 -; GISEL-NEXT: v_cndmask_b32_e64 v2, v18, v9, s[4:5] -; GISEL-NEXT: v_cndmask_b32_e32 v3, v16, v17, vcc -; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v13 +; GISEL-NEXT: v_cndmask_b32_e64 v3, v18, v7, s[4:5] +; GISEL-NEXT: v_cndmask_b32_e32 v6, v13, v16, vcc +; GISEL-NEXT: v_cmp_ne_u32_e32 vcc, 0, v10 ; GISEL-NEXT: v_cndmask_b32_e32 v0, v4, v1, vcc -; GISEL-NEXT: v_cndmask_b32_e64 v4, v19, v11, s[4:5] -; GISEL-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v15 -; GISEL-NEXT: v_cndmask_b32_e64 v2, v7, v2, s[4:5] -; GISEL-NEXT: v_cndmask_b32_e32 v1, v8, v3, vcc +; GISEL-NEXT: v_cndmask_b32_e64 v4, v19, v14, s[4:5] +; GISEL-NEXT: v_cmp_ne_u32_e64 s[4:5], 0, v17 +; GISEL-NEXT: v_cndmask_b32_e64 v2, v2, v3, s[4:5] +; GISEL-NEXT: v_cndmask_b32_e32 v1, v15, v6, vcc ; GISEL-NEXT: v_cndmask_b32_e64 v3, v5, v4, s[4:5] ; GISEL-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/umed3.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/umed3.ll index 2b54123f01ad..f5068f547a60 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/umed3.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/umed3.ll @@ -1,7 +1,7 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py -; RUN: llc -global-isel -mtriple=amdgcn-amd-mesa3d -mcpu=gfx802 < %s | FileCheck -check-prefixes=GFX89,GFX8 %s -; RUN: llc -global-isel -mtriple=amdgcn-amd-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX89,GFX9 %s -; RUN: llc -global-isel -mtriple=amdgcn-amd-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-mesa3d -mcpu=gfx802 < %s | FileCheck -check-prefixes=GFX89,GFX8 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-mesa3d -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX89,GFX9 %s +; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-mesa3d -mcpu=gfx1010 < %s | FileCheck -check-prefix=GFX10 %s define i32 @test_min_max_ValK0_K1_u32(i32 %a) { ; GFX89-LABEL: test_min_max_ValK0_K1_u32: diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/urem.i64.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/urem.i64.ll index a41ec8e7ce3e..be5543b9b5b7 100644 --- a/llvm/test/CodeGen/AMDGPU/GlobalISel/urem.i64.ll +++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/urem.i64.ll @@ -2058,42 +2058,34 @@ define <2 x i64> @v_urem_v2i64_24bit(<2 x i64> %num, <2 x i64> %den) { ; GISEL-NEXT: v_mul_hi_u32 v5, 0, v5 ; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v12 ; GISEL-NEXT: v_add_i32_e32 v8, vcc, v13, v8 +; GISEL-NEXT: v_mul_lo_u32 v2, v3, v2 ; GISEL-NEXT: v_add_i32_e32 v10, vcc, v10, v15 ; GISEL-NEXT: v_add_i32_e32 v7, vcc, v16, v7 +; GISEL-NEXT: v_mul_lo_u32 v5, v1, v5 ; GISEL-NEXT: v_add_i32_e32 v4, vcc, v4, v9 ; GISEL-NEXT: v_cndmask_b32_e64 v4, 0, 1, vcc ; GISEL-NEXT: v_add_i32_e32 v8, vcc, v8, v14 +; GISEL-NEXT: v_add_i32_e32 v9, vcc, v10, v11 ; GISEL-NEXT: v_cndmask_b32_e64 v9, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v10, vcc, v10, v11 -; GISEL-NEXT: v_cndmask_b32_e64 v10, 0, 1, vcc ; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v17 -; GISEL-NEXT: v_cndmask_b32_e64 v11, 0, 1, vcc ; GISEL-NEXT: v_add_i32_e32 v4, vcc, v8, v4 -; GISEL-NEXT: v_cndmask_b32_e64 v8, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v10 -; GISEL-NEXT: v_cndmask_b32_e64 v10, 0, 1, vcc -; GISEL-NEXT: v_add_i32_e32 v8, vcc, v9, v8 -; GISEL-NEXT: v_mul_lo_u32 v9, v3, v4 -; GISEL-NEXT: v_mul_lo_u32 v12, 0, v4 +; GISEL-NEXT: v_add_i32_e32 v7, vcc, v7, v9 +; GISEL-NEXT: v_mul_lo_u32 v8, v3, v4 +; GISEL-NEXT: v_mul_lo_u32 v9, 0, v4 ; GISEL-NEXT: v_mul_hi_u32 v4, v3, v4 -; GISEL-NEXT: v_add_i32_e32 v10, vcc, v11, v10 -; GISEL-NEXT: v_mul_lo_u32 v11, v1, v7 -; GISEL-NEXT: v_mul_lo_u32 v13, 0, v7 +; GISEL-NEXT: v_mul_lo_u32 v10, v1, v7 +; GISEL-NEXT: v_mul_lo_u32 v11, 0, v7 ; GISEL-NEXT: v_mul_hi_u32 v7, v1, v7 -; GISEL-NEXT: v_add_i32_e32 v2, vcc, v2, v8 -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v5, v10 -; GISEL-NEXT: v_mul_lo_u32 v2, v3, v2 -; GISEL-NEXT: v_mul_lo_u32 v5, v1, v5 -; GISEL-NEXT: v_add_i32_e32 v2, vcc, v12, v2 -; GISEL-NEXT: v_add_i32_e32 v5, vcc, v13, v5 +; GISEL-NEXT: v_add_i32_e32 v2, vcc, v9, v2 +; GISEL-NEXT: v_add_i32_e32 v5, vcc, v11, v5 ; GISEL-NEXT: v_add_i32_e32 v2, vcc, v2, v4 ; GISEL-NEXT: v_add_i32_e32 v4, vcc, v5, v7 -; GISEL-NEXT: v_sub_i32_e32 v5, vcc, v6, v9 +; GISEL-NEXT: v_sub_i32_e32 v5, vcc, v6, v8 ; GISEL-NEXT: v_subb_u32_e64 v6, s[4:5], 0, v2, vcc ; GISEL-NEXT: v_sub_i32_e64 v2, s[4:5], 0, v2 ; GISEL-NEXT: v_cmp_ge_u32_e64 s[4:5], v5, v3 ; GISEL-NEXT: v_cndmask_b32_e64 v7, 0, -1, s[4:5] -; GISEL-NEXT: v_sub_i32_e64 v8, s[4:5], v0, v11 +; GISEL-NEXT: v_sub_i32_e64 v8, s[4:5], v0, v10 ; GISEL-NEXT: v_subb_u32_e64 v9, s[6:7], 0, v4, s[4:5] ; GISEL-NEXT: v_sub_i32_e64 v0, s[6:7], 0, v4 ; GISEL-NEXT: v_cmp_ge_u32_e64 s[6:7], v8, v1 diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn-ieee.ll b/llvm/test/CodeGen/AMDGPU/amdgcn-ieee.ll index f96a6f74e3c0..b239c4694c03 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn-ieee.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn-ieee.ll @@ -1,13 +1,89 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 ; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti < %s | FileCheck -check-prefix=GCN %s -; GCN-LABEL: {{^}}kernel_ieee_mode_default: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]] -; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]] -; GCN-NOT: v_mul_f32 define amdgpu_kernel void @kernel_ieee_mode_default() #0 { +; GCN-LABEL: kernel_ieee_mode_default: +; GCN: .amd_kernel_code_t +; GCN-NEXT: amd_code_version_major = 1 +; GCN-NEXT: amd_code_version_minor = 2 +; GCN-NEXT: amd_machine_kind = 1 +; GCN-NEXT: amd_machine_version_major = 6 +; GCN-NEXT: amd_machine_version_minor = 0 +; GCN-NEXT: amd_machine_version_stepping = 0 +; GCN-NEXT: kernel_code_entry_byte_offset = 256 +; GCN-NEXT: kernel_code_prefetch_byte_size = 0 +; GCN-NEXT: granulated_workitem_vgpr_count = 0 +; GCN-NEXT: granulated_wavefront_sgpr_count = 0 +; GCN-NEXT: priority = 0 +; GCN-NEXT: float_mode = 240 +; GCN-NEXT: priv = 0 +; GCN-NEXT: enable_dx10_clamp = 1 +; GCN-NEXT: debug_mode = 0 +; GCN-NEXT: enable_ieee_mode = 1 +; GCN-NEXT: enable_wgp_mode = 0 +; GCN-NEXT: enable_mem_ordered = 0 +; GCN-NEXT: enable_fwd_progress = 0 +; GCN-NEXT: enable_sgpr_private_segment_wave_byte_offset = 0 +; GCN-NEXT: user_sgpr_count = 12 +; GCN-NEXT: enable_trap_handler = 0 +; GCN-NEXT: enable_sgpr_workgroup_id_x = 1 +; GCN-NEXT: enable_sgpr_workgroup_id_y = 1 +; GCN-NEXT: enable_sgpr_workgroup_id_z = 1 +; GCN-NEXT: enable_sgpr_workgroup_info = 0 +; GCN-NEXT: enable_vgpr_workitem_id = 2 +; GCN-NEXT: enable_exception_msb = 0 +; GCN-NEXT: granulated_lds_size = 0 +; GCN-NEXT: enable_exception = 0 +; GCN-NEXT: enable_sgpr_private_segment_buffer = 1 +; GCN-NEXT: enable_sgpr_dispatch_ptr = 1 +; GCN-NEXT: enable_sgpr_queue_ptr = 1 +; GCN-NEXT: enable_sgpr_kernarg_segment_ptr = 1 +; GCN-NEXT: enable_sgpr_dispatch_id = 1 +; GCN-NEXT: enable_sgpr_flat_scratch_init = 0 +; GCN-NEXT: enable_sgpr_private_segment_size = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_x = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_y = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_z = 0 +; GCN-NEXT: enable_wavefront_size32 = 0 +; GCN-NEXT: enable_ordered_append_gds = 0 +; GCN-NEXT: private_element_size = 1 +; GCN-NEXT: is_ptr64 = 1 +; GCN-NEXT: is_dynamic_callstack = 0 +; GCN-NEXT: is_debug_enabled = 0 +; GCN-NEXT: is_xnack_enabled = 0 +; GCN-NEXT: workitem_private_segment_byte_size = 0 +; GCN-NEXT: workgroup_group_segment_byte_size = 0 +; GCN-NEXT: gds_segment_byte_size = 0 +; GCN-NEXT: kernarg_segment_byte_size = 16 +; GCN-NEXT: workgroup_fbarrier_count = 0 +; GCN-NEXT: wavefront_sgpr_count = 4 +; GCN-NEXT: workitem_vgpr_count = 2 +; GCN-NEXT: reserved_vgpr_first = 0 +; GCN-NEXT: reserved_vgpr_count = 0 +; GCN-NEXT: reserved_sgpr_first = 0 +; GCN-NEXT: reserved_sgpr_count = 0 +; GCN-NEXT: debug_wavefront_private_segment_offset_sgpr = 0 +; GCN-NEXT: debug_private_segment_buffer_sgpr = 0 +; GCN-NEXT: kernarg_segment_alignment = 4 +; GCN-NEXT: group_segment_alignment = 4 +; GCN-NEXT: private_segment_alignment = 4 +; GCN-NEXT: wavefront_size = 6 +; GCN-NEXT: call_convention = -1 +; GCN-NEXT: runtime_loader_kernel_symbol = 0 +; GCN-NEXT: .end_amd_kernel_code_t +; GCN-NEXT: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0 +; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1 +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -15,14 +91,89 @@ define amdgpu_kernel void @kernel_ieee_mode_default() #0 { ret void } -; GCN-LABEL: {{^}}kernel_ieee_mode_on: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]] -; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]] -; GCN-NOT: v_mul_f32 define amdgpu_kernel void @kernel_ieee_mode_on() #1 { +; GCN-LABEL: kernel_ieee_mode_on: +; GCN: .amd_kernel_code_t +; GCN-NEXT: amd_code_version_major = 1 +; GCN-NEXT: amd_code_version_minor = 2 +; GCN-NEXT: amd_machine_kind = 1 +; GCN-NEXT: amd_machine_version_major = 6 +; GCN-NEXT: amd_machine_version_minor = 0 +; GCN-NEXT: amd_machine_version_stepping = 0 +; GCN-NEXT: kernel_code_entry_byte_offset = 256 +; GCN-NEXT: kernel_code_prefetch_byte_size = 0 +; GCN-NEXT: granulated_workitem_vgpr_count = 0 +; GCN-NEXT: granulated_wavefront_sgpr_count = 0 +; GCN-NEXT: priority = 0 +; GCN-NEXT: float_mode = 240 +; GCN-NEXT: priv = 0 +; GCN-NEXT: enable_dx10_clamp = 1 +; GCN-NEXT: debug_mode = 0 +; GCN-NEXT: enable_ieee_mode = 1 +; GCN-NEXT: enable_wgp_mode = 0 +; GCN-NEXT: enable_mem_ordered = 0 +; GCN-NEXT: enable_fwd_progress = 0 +; GCN-NEXT: enable_sgpr_private_segment_wave_byte_offset = 0 +; GCN-NEXT: user_sgpr_count = 12 +; GCN-NEXT: enable_trap_handler = 0 +; GCN-NEXT: enable_sgpr_workgroup_id_x = 1 +; GCN-NEXT: enable_sgpr_workgroup_id_y = 1 +; GCN-NEXT: enable_sgpr_workgroup_id_z = 1 +; GCN-NEXT: enable_sgpr_workgroup_info = 0 +; GCN-NEXT: enable_vgpr_workitem_id = 2 +; GCN-NEXT: enable_exception_msb = 0 +; GCN-NEXT: granulated_lds_size = 0 +; GCN-NEXT: enable_exception = 0 +; GCN-NEXT: enable_sgpr_private_segment_buffer = 1 +; GCN-NEXT: enable_sgpr_dispatch_ptr = 1 +; GCN-NEXT: enable_sgpr_queue_ptr = 1 +; GCN-NEXT: enable_sgpr_kernarg_segment_ptr = 1 +; GCN-NEXT: enable_sgpr_dispatch_id = 1 +; GCN-NEXT: enable_sgpr_flat_scratch_init = 0 +; GCN-NEXT: enable_sgpr_private_segment_size = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_x = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_y = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_z = 0 +; GCN-NEXT: enable_wavefront_size32 = 0 +; GCN-NEXT: enable_ordered_append_gds = 0 +; GCN-NEXT: private_element_size = 1 +; GCN-NEXT: is_ptr64 = 1 +; GCN-NEXT: is_dynamic_callstack = 0 +; GCN-NEXT: is_debug_enabled = 0 +; GCN-NEXT: is_xnack_enabled = 0 +; GCN-NEXT: workitem_private_segment_byte_size = 0 +; GCN-NEXT: workgroup_group_segment_byte_size = 0 +; GCN-NEXT: gds_segment_byte_size = 0 +; GCN-NEXT: kernarg_segment_byte_size = 16 +; GCN-NEXT: workgroup_fbarrier_count = 0 +; GCN-NEXT: wavefront_sgpr_count = 4 +; GCN-NEXT: workitem_vgpr_count = 2 +; GCN-NEXT: reserved_vgpr_first = 0 +; GCN-NEXT: reserved_vgpr_count = 0 +; GCN-NEXT: reserved_sgpr_first = 0 +; GCN-NEXT: reserved_sgpr_count = 0 +; GCN-NEXT: debug_wavefront_private_segment_offset_sgpr = 0 +; GCN-NEXT: debug_private_segment_buffer_sgpr = 0 +; GCN-NEXT: kernarg_segment_alignment = 4 +; GCN-NEXT: group_segment_alignment = 4 +; GCN-NEXT: private_segment_alignment = 4 +; GCN-NEXT: wavefront_size = 6 +; GCN-NEXT: call_convention = -1 +; GCN-NEXT: runtime_loader_kernel_symbol = 0 +; GCN-NEXT: .end_amd_kernel_code_t +; GCN-NEXT: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0 +; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1 +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -30,14 +181,87 @@ define amdgpu_kernel void @kernel_ieee_mode_on() #1 { ret void } -; GCN-LABEL: {{^}}kernel_ieee_mode_off: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-NOT: [[VAL0]] -; GCN-NOT: [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]] -; GCN-NOT: v_mul_f32 define amdgpu_kernel void @kernel_ieee_mode_off() #2 { +; GCN-LABEL: kernel_ieee_mode_off: +; GCN: .amd_kernel_code_t +; GCN-NEXT: amd_code_version_major = 1 +; GCN-NEXT: amd_code_version_minor = 2 +; GCN-NEXT: amd_machine_kind = 1 +; GCN-NEXT: amd_machine_version_major = 6 +; GCN-NEXT: amd_machine_version_minor = 0 +; GCN-NEXT: amd_machine_version_stepping = 0 +; GCN-NEXT: kernel_code_entry_byte_offset = 256 +; GCN-NEXT: kernel_code_prefetch_byte_size = 0 +; GCN-NEXT: granulated_workitem_vgpr_count = 0 +; GCN-NEXT: granulated_wavefront_sgpr_count = 0 +; GCN-NEXT: priority = 0 +; GCN-NEXT: float_mode = 240 +; GCN-NEXT: priv = 0 +; GCN-NEXT: enable_dx10_clamp = 1 +; GCN-NEXT: debug_mode = 0 +; GCN-NEXT: enable_ieee_mode = 0 +; GCN-NEXT: enable_wgp_mode = 0 +; GCN-NEXT: enable_mem_ordered = 0 +; GCN-NEXT: enable_fwd_progress = 0 +; GCN-NEXT: enable_sgpr_private_segment_wave_byte_offset = 0 +; GCN-NEXT: user_sgpr_count = 12 +; GCN-NEXT: enable_trap_handler = 0 +; GCN-NEXT: enable_sgpr_workgroup_id_x = 1 +; GCN-NEXT: enable_sgpr_workgroup_id_y = 1 +; GCN-NEXT: enable_sgpr_workgroup_id_z = 1 +; GCN-NEXT: enable_sgpr_workgroup_info = 0 +; GCN-NEXT: enable_vgpr_workitem_id = 2 +; GCN-NEXT: enable_exception_msb = 0 +; GCN-NEXT: granulated_lds_size = 0 +; GCN-NEXT: enable_exception = 0 +; GCN-NEXT: enable_sgpr_private_segment_buffer = 1 +; GCN-NEXT: enable_sgpr_dispatch_ptr = 1 +; GCN-NEXT: enable_sgpr_queue_ptr = 1 +; GCN-NEXT: enable_sgpr_kernarg_segment_ptr = 1 +; GCN-NEXT: enable_sgpr_dispatch_id = 1 +; GCN-NEXT: enable_sgpr_flat_scratch_init = 0 +; GCN-NEXT: enable_sgpr_private_segment_size = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_x = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_y = 0 +; GCN-NEXT: enable_sgpr_grid_workgroup_count_z = 0 +; GCN-NEXT: enable_wavefront_size32 = 0 +; GCN-NEXT: enable_ordered_append_gds = 0 +; GCN-NEXT: private_element_size = 1 +; GCN-NEXT: is_ptr64 = 1 +; GCN-NEXT: is_dynamic_callstack = 0 +; GCN-NEXT: is_debug_enabled = 0 +; GCN-NEXT: is_xnack_enabled = 0 +; GCN-NEXT: workitem_private_segment_byte_size = 0 +; GCN-NEXT: workgroup_group_segment_byte_size = 0 +; GCN-NEXT: gds_segment_byte_size = 0 +; GCN-NEXT: kernarg_segment_byte_size = 16 +; GCN-NEXT: workgroup_fbarrier_count = 0 +; GCN-NEXT: wavefront_sgpr_count = 4 +; GCN-NEXT: workitem_vgpr_count = 2 +; GCN-NEXT: reserved_vgpr_first = 0 +; GCN-NEXT: reserved_vgpr_count = 0 +; GCN-NEXT: reserved_sgpr_first = 0 +; GCN-NEXT: reserved_sgpr_count = 0 +; GCN-NEXT: debug_wavefront_private_segment_offset_sgpr = 0 +; GCN-NEXT: debug_private_segment_buffer_sgpr = 0 +; GCN-NEXT: kernarg_segment_alignment = 4 +; GCN-NEXT: group_segment_alignment = 4 +; GCN-NEXT: private_segment_alignment = 4 +; GCN-NEXT: wavefront_size = 6 +; GCN-NEXT: call_convention = -1 +; GCN-NEXT: runtime_loader_kernel_symbol = 0 +; GCN-NEXT: .end_amd_kernel_code_t +; GCN-NEXT: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -45,14 +269,22 @@ define amdgpu_kernel void @kernel_ieee_mode_off() #2 { ret void } -; GCN-LABEL: {{^}}func_ieee_mode_default: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]] -; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]] -; GCN-NOT: v_mul_f32 define void @func_ieee_mode_default() #0 { +; GCN-LABEL: func_ieee_mode_default: +; GCN: ; %bb.0: +; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GCN-NEXT: s_mov_b32 s7, 0xf000 +; GCN-NEXT: s_mov_b32 s6, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[4:7], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[4:7], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0 +; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1 +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) +; GCN-NEXT: s_setpc_b64 s[30:31] %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -60,14 +292,22 @@ define void @func_ieee_mode_default() #0 { ret void } -; GCN-LABEL: {{^}}func_ieee_mode_on: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]] -; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]] -; GCN-NOT: v_mul_f32 define void @func_ieee_mode_on() #1 { +; GCN-LABEL: func_ieee_mode_on: +; GCN: ; %bb.0: +; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GCN-NEXT: s_mov_b32 s7, 0xf000 +; GCN-NEXT: s_mov_b32 s6, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[4:7], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[4:7], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0 +; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1 +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) +; GCN-NEXT: s_setpc_b64 s[30:31] %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -75,14 +315,20 @@ define void @func_ieee_mode_on() #1 { ret void } -; GCN-LABEL: {{^}}func_ieee_mode_off: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-NOT: [[VAL0]] -; GCN-NOT: [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]] -; GCN-NOT: v_mul_f32 define void @func_ieee_mode_off() #2 { +; GCN-LABEL: func_ieee_mode_off: +; GCN: ; %bb.0: +; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; GCN-NEXT: s_mov_b32 s7, 0xf000 +; GCN-NEXT: s_mov_b32 s6, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[4:7], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[4:7], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) +; GCN-NEXT: s_setpc_b64 s[30:31] %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -90,14 +336,19 @@ define void @func_ieee_mode_off() #2 { ret void } -; GCN-LABEL: {{^}}cs_ieee_mode_default: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-NOT: [[VAL0]] -; GCN-NOT: [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]] -; GCN-NOT: v_mul_f32 define amdgpu_cs void @cs_ieee_mode_default() #0 { +; GCN-LABEL: cs_ieee_mode_default: +; GCN: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -105,14 +356,21 @@ define amdgpu_cs void @cs_ieee_mode_default() #0 { ret void } -; GCN-LABEL: {{^}}cs_ieee_mode_on: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]] -; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]] -; GCN-NOT: v_mul_f32 define amdgpu_cs void @cs_ieee_mode_on() #1 { +; GCN-LABEL: cs_ieee_mode_on: +; GCN: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0 +; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1 +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -120,14 +378,19 @@ define amdgpu_cs void @cs_ieee_mode_on() #1 { ret void } -; GCN-LABEL: {{^}}cs_ieee_mode_off: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-NOT: [[VAL0]] -; GCN-NOT: [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]] -; GCN-NOT: v_mul_f32 define amdgpu_cs void @cs_ieee_mode_off() #2 { +; GCN-LABEL: cs_ieee_mode_off: +; GCN: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -135,14 +398,19 @@ define amdgpu_cs void @cs_ieee_mode_off() #2 { ret void } -; GCN-LABEL: {{^}}ps_ieee_mode_default: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-NOT: [[VAL0]] -; GCN-NOT: [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]] -; GCN-NOT: v_mul_f32 define amdgpu_ps void @ps_ieee_mode_default() #0 { +; GCN-LABEL: ps_ieee_mode_default: +; GCN: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -150,14 +418,21 @@ define amdgpu_ps void @ps_ieee_mode_default() #0 { ret void } -; GCN-LABEL: {{^}}ps_ieee_mode_on: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-DAG: v_mul_f32_e32 [[QUIET0:v[0-9]+]], 1.0, [[VAL0]] -; GCN-DAG: v_mul_f32_e32 [[QUIET1:v[0-9]+]], 1.0, [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[QUIET0]], [[QUIET1]] -; GCN-NOT: v_mul_f32 define amdgpu_ps void @ps_ieee_mode_on() #1 { +; GCN-LABEL: ps_ieee_mode_on: +; GCN: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0 +; GCN-NEXT: v_mul_f32_e32 v1, 1.0, v1 +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) @@ -165,14 +440,19 @@ define amdgpu_ps void @ps_ieee_mode_on() #1 { ret void } -; GCN-LABEL: {{^}}ps_ieee_mode_off: -; GCN: {{buffer|global|flat}}_load_dword [[VAL0:v[0-9]+]] -; GCN: {{buffer|global|flat}}_load_dword [[VAL1:v[0-9]+]] -; GCN-NOT: [[VAL0]] -; GCN-NOT: [[VAL1]] -; GCN: v_min_f32_e32 [[MIN:v[0-9]+]], [[VAL0]], [[VAL1]] -; GCN-NOT: v_mul_f32 define amdgpu_ps void @ps_ieee_mode_off() #2 { +; GCN-LABEL: ps_ieee_mode_off: +; GCN: ; %bb.0: +; GCN-NEXT: s_mov_b32 s3, 0xf000 +; GCN-NEXT: s_mov_b32 s2, -1 +; GCN-NEXT: buffer_load_dword v0, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: buffer_load_dword v1, off, s[0:3], 0 glc +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: v_min_f32_e32 v0, v0, v1 +; GCN-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GCN-NEXT: s_waitcnt vmcnt(0) +; GCN-NEXT: s_endpgm %val0 = load volatile float, ptr addrspace(1) poison %val1 = load volatile float, ptr addrspace(1) poison %min = call float @llvm.minnum.f32(float %val0, float %val1) diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll index df9c97fa2372..117af9590ff6 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.1024bit.ll @@ -6551,271 +6551,205 @@ define <128 x i8> @bitcast_v32i32_to_v128i8(<32 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v160.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v33.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v36.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v149.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v146.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v145.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v71.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v69.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v133.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v132.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v130.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v9.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v10.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v50.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v117.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v115.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v113.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v99.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v15.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v16.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v17.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v19.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v21.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v23.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v71.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v25.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v26.h, v27.l ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v32i32_to_v128i8: @@ -15709,61 +15643,61 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:368 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:360 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:352 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:344 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:328 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:320 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:312 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:304 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:296 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:288 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:288 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:280 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:280 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:272 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:264 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:264 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:256 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:256 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:248 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:240 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:240 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:232 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:224 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:216 -; GFX11-TRUE16-NEXT: scratch_load_b32 v114, off, s32 offset:388 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:40 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:48 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_b32 v113, off, s32 offset:388 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:8 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:48 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:56 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:104 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:112 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128 @@ -15778,121 +15712,123 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:204 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:196 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:196 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:188 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:148 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:60 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:52 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:84 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:52 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:44 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.l, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v21.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v29.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54) -; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v81.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v82.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v83.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v85.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v97.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(9) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v98.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(3) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v99.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v102.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v161.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v162.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v162.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v164.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v164.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v165.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v71.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v52.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v29.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v51.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(56) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v54.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(26) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.l, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v71.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v113 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v83.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v84.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v85.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v96.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.l, 8, v97.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v101.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v102.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v163.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v164.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.l, 8, v165.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v55.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -15903,215 +15839,179 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB14_4 ; GFX11-TRUE16-NEXT: .LBB14_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB14_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v146.l -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v0.l, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v151.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v0.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v150.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v149, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v1.h, v150.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v145.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v133.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v131.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v148.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v130.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v149, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v2.l, v148.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v134.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v145.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v130.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v149, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v3.l, v147.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v147.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v118.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v149, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v4.l, v144.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v4.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v133.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v117.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v129.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v128.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v116.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v149, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v5.l, v135.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v5.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v132.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v129.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v112.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v149, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v6.l, v133.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v6.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v128.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v103.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v149, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v7.l, v131.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v7.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v100.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v99.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v149, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v8.l, v129.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v96.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v149, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v9.l, v128.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v9.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v149, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v10.l, v117.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v10.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v102.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v103.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v99.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v99.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v87.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v86.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v81.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v149, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v11.l, v116.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v11.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v149, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v12.l, v114.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v12.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v97.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v69.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v68.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v149, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v13.l, v112.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v13.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v87.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v67.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v65.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v149, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v14.l, v102.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v14.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v85.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v83.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v68.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v65.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v53.l ; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v149, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v15.l, v100.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v15.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v82.l ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l ; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v149, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v16.l, v98.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v16.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.h ; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v149, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v17.l, v97.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v17.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v70.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v68.l ; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v38.h ; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v149, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v18.l, v87.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v18.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v66.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v37.l ; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v149, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v19.l, v85.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v19.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v149, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v20.l, v83.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v20.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v55.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v149, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v21.l, v81.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v21.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v53.l ; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v33.l ; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v149, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v22.l, v71.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v22.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v150.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v150.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v151.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v151.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v146.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v147.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v147.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v148.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v134.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v144.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v130.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v131.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v131.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v117.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v118.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v113.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v113.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v114.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v114.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v100.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v101.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v102.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v96.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v97.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v97.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v83.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v84.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v85.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v70.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v80.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v66.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v67.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v67.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v53.h +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v55.h +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v50.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v30.l, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v51.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v52.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 @@ -16133,433 +16033,329 @@ define <32 x i32> @bitcast_v128i8_to_v32i32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v149, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v23.l, v70.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v23.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v51.l -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v149, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v24.l, v67.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v24.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v149, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v25.l, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v25.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v149, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v26.l, v64.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v26.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v149, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v27.l, v54.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v27.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v149, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v28.l, v53.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v28.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v149, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v29.l, v52.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v29.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v149, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v30.l, v51.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v30.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v149, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v31.l, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v149, v31 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB14_2 ; GFX11-TRUE16-NEXT: .LBB14_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v146.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v145.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v133.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v129.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v128.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v116.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v103.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v99.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v99.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v87.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v87.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v86.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v83.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v82.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v82.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v70.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v69.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v68.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v65.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v65.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v64.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v53.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v50.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v48.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, v32.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v134.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v31, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v148.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v130.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v130.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v144.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v145.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v118.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v117.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v133.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v133.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v131.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v132.l, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v113.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v129.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v129.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v128.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v128.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v103.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v101.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v117.h, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v118.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v116.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v116.h, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v99.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v96.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v96.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v113.l, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v86.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v84.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v84.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v102.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v102.h, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v16 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v100.h, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.l, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v82.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v80.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v80.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v98.h, v15.l -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v99.l, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v97.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v97.h, v16.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v68.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v67.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v87.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v87.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l ; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v31, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v85.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v85.h, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v65.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v31, v21 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v55.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v50.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v83.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v83.h, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v31, v22 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v21.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v81.h, v20.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v82.l, v20.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v31, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.l, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v71.h, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v31, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v70.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v70.h, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v31, v25 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v67.h, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v68.l, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v26 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v25.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.l, v24.l -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v66.h, v24.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v27 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v25.h -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v64.l, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v64.h, v25.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v31, v28 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v55.l, v26.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v31, v29 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v53.h, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v54.l, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v31, v30 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v34.h, 0x300, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v53.l, v28.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v31, v34 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v29.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v51.h, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v52.l, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h ; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l ; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v33 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v50.h, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v51.l, v30.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v31, v33 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.h, 0x300, v32.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v150.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v150.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v151.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v151.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v147.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v148.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v134.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v144.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v144.h, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v131.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v131.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v132.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v117.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v117.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v118.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v118.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v119.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v115.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v100.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v101.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v102.l, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v102.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v96.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v97.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v97.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v98.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v83.h, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v84.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v84.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v85.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v85.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.h, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v71.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v80.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v80.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v67.l, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v67.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v68.l, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v53.h, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v54.h, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v55.l, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v55.h, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v51.l, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v52.l, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v52.h, v31.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v16.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v17.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v18.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v19.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v20.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v21.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v22.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v23.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v24.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v25.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v26.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v26.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v27.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v29.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v30.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v31.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v31.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -42692,271 +42488,205 @@ define <128 x i8> @bitcast_v32f32_to_v128i8(<32 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v160.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v33.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v36.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v149.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v146.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v145.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v71.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v69.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v133.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v132.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v130.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v9.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v10.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v50.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v117.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v115.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v113.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v99.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v15.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v16.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v17.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v19.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v21.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v23.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v71.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v25.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v26.h, v27.l ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v32f32_to_v128i8: @@ -53003,61 +52733,61 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:368 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:360 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:352 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:344 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:328 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:320 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:312 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:304 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:296 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:288 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:288 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:280 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:280 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:272 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:264 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:264 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:256 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:256 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:248 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:240 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:240 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:232 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:224 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:216 -; GFX11-TRUE16-NEXT: scratch_load_b32 v114, off, s32 offset:388 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:40 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:48 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_b32 v113, off, s32 offset:388 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:8 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:48 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:56 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:104 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:112 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128 @@ -53072,121 +52802,123 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:204 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:196 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:196 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:188 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:148 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:60 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:52 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:84 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:52 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:44 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.l, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v21.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v29.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54) -; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v81.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v82.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v83.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v85.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v97.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(9) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v98.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(3) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v99.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v102.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v161.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v162.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v162.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v164.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v164.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v165.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v71.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v52.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v29.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v51.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(56) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v54.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(26) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.l, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v71.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v113 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v83.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v84.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v85.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v96.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.l, 8, v97.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v101.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v102.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v163.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v164.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.l, 8, v165.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v55.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -53197,215 +52929,179 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB38_4 ; GFX11-TRUE16-NEXT: .LBB38_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB38_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v146.l -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v0.l, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v151.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v0.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v150.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v149, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v1.h, v150.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v145.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v133.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v131.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v148.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v130.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v149, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v2.l, v148.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v134.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v145.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v130.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v149, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v3.l, v147.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v147.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v118.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v149, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v4.l, v144.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v4.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v133.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v117.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v129.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v128.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v116.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v149, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v5.l, v135.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v5.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v132.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v129.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v112.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v149, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v6.l, v133.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v6.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v128.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v103.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v149, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v7.l, v131.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v7.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v100.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v99.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v149, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v8.l, v129.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v96.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v149, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v9.l, v128.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v9.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v149, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v10.l, v117.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v10.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v102.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v103.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v99.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v99.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v87.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v86.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v81.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v149, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v11.l, v116.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v11.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v149, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v12.l, v114.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v12.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v97.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v69.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v68.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v149, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v13.l, v112.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v13.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v87.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v67.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v65.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v149, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v14.l, v102.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v14.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v85.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v83.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v68.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v65.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v53.l ; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v149, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v15.l, v100.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v15.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v82.l ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l ; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v149, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v16.l, v98.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v16.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.h ; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v149, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v17.l, v97.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v17.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v70.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v68.l ; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v38.h ; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v149, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v18.l, v87.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v18.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v66.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v37.l ; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v149, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v19.l, v85.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v19.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v149, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v20.l, v83.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v20.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v55.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v149, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v21.l, v81.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v21.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v53.l ; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v33.l ; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v149, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v22.l, v71.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v22.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v150.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v150.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v151.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v151.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v146.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v147.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v147.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v148.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v134.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v144.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v130.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v131.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v131.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v117.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v118.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v113.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v113.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v114.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v114.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v100.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v101.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v102.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v96.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v97.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v97.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v83.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v84.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v85.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v70.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v80.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v66.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v67.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v67.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v53.h +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v55.h +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v50.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v30.l, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v51.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v52.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 @@ -53427,433 +53123,329 @@ define <32 x float> @bitcast_v128i8_to_v32f32(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v149, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v23.l, v70.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v23.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v51.l -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v149, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v24.l, v67.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v24.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v149, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v25.l, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v25.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v149, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v26.l, v64.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v26.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v149, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v27.l, v54.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v27.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v149, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v28.l, v53.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v28.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v149, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v29.l, v52.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v29.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v149, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v30.l, v51.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v30.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v149, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v31.l, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v149, v31 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB38_2 ; GFX11-TRUE16-NEXT: .LBB38_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v146.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v145.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v133.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v129.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v128.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v116.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v103.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v99.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v99.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v87.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v87.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v86.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v83.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v82.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v82.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v70.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v69.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v68.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v65.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v65.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v64.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v53.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v50.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v48.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, v32.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v134.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v31, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v148.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v130.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v130.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v144.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v145.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v118.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v117.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v133.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v133.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v131.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v132.l, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v113.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v129.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v129.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v128.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v128.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v103.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v101.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v117.h, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v118.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v116.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v116.h, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v99.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v96.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v96.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v113.l, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v86.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v84.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v84.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v102.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v102.h, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v16 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v100.h, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.l, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v82.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v80.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v80.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v98.h, v15.l -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v99.l, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v97.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v97.h, v16.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v68.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v67.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v87.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v87.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l ; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v31, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v85.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v85.h, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v65.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v31, v21 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v55.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v50.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v83.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v83.h, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v31, v22 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v21.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v81.h, v20.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v82.l, v20.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v31, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.l, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v71.h, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v31, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v70.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v70.h, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v31, v25 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v67.h, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v68.l, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v26 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v25.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.l, v24.l -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v66.h, v24.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v27 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v25.h -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v64.l, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v64.h, v25.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v31, v28 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v55.l, v26.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v31, v29 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v53.h, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v54.l, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v31, v30 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v34.h, 0x300, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v53.l, v28.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v31, v34 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v29.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v51.h, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v52.l, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h ; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l ; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v33 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v50.h, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v51.l, v30.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v31, v33 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.h, 0x300, v32.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v150.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v150.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v151.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v151.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v147.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v148.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v134.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v144.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v144.h, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v131.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v131.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v132.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v117.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v117.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v118.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v118.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v119.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v115.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v100.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v101.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v102.l, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v102.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v96.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v97.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v97.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v98.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v83.h, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v84.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v84.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v85.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v85.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.h, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v71.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v80.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v80.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v67.l, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v67.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v68.l, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v53.h, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v54.h, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v55.l, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v55.h, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v51.l, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v52.l, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v52.h, v31.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v16.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v17.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v18.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v19.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v20.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v21.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v22.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v23.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v24.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v25.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v26.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v26.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v27.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v29.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v30.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v31.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v31.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -78968,271 +78560,205 @@ define <128 x i8> @bitcast_v16i64_to_v128i8(<16 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v160.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v33.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v36.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v149.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v146.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v145.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v71.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v69.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v133.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v132.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v130.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v9.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v10.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v50.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v117.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v115.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v113.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v99.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v15.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v16.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v17.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v19.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v21.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v23.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v71.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v25.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v26.h, v27.l ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v16i64_to_v128i8: @@ -88136,61 +87662,61 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:368 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:360 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:352 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:344 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:328 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:320 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:312 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:304 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:296 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:288 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:288 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:280 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:280 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:272 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:264 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:264 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:256 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:256 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:248 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:240 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:240 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:232 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:224 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:216 -; GFX11-TRUE16-NEXT: scratch_load_b32 v114, off, s32 offset:388 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:40 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:48 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_b32 v113, off, s32 offset:388 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:8 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:48 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:56 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:104 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:112 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128 @@ -88205,121 +87731,123 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:204 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:196 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:196 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:188 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:148 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:60 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:52 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:84 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:52 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:44 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.l, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v21.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v29.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54) -; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v81.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v82.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v83.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v85.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v97.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(9) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v98.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(3) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v99.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v102.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v161.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v162.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v162.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v164.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v164.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v165.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v71.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v52.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v29.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v51.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(56) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v54.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(26) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.l, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v71.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v113 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v83.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v84.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v85.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v96.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.l, 8, v97.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v101.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v102.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v163.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v164.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.l, 8, v165.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v55.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -88330,215 +87858,179 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB58_4 ; GFX11-TRUE16-NEXT: .LBB58_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB58_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v146.l -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v0.l, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v151.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v0.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v150.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v149, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v1.h, v150.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v145.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v133.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v131.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v148.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v130.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v149, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v2.l, v148.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v134.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v145.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v130.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v149, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v3.l, v147.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v147.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v118.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v149, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v4.l, v144.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v4.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v133.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v117.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v129.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v128.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v116.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v149, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v5.l, v135.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v5.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v132.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v129.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v112.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v149, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v6.l, v133.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v6.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v128.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v103.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v149, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v7.l, v131.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v7.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v100.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v99.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v149, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v8.l, v129.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v96.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v149, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v9.l, v128.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v9.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v149, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v10.l, v117.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v10.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v102.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v103.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v99.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v99.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v87.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v86.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v81.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v149, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v11.l, v116.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v11.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v149, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v12.l, v114.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v12.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v97.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v69.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v68.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v149, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v13.l, v112.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v13.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v87.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v67.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v65.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v149, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v14.l, v102.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v14.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v85.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v83.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v68.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v65.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v53.l ; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v149, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v15.l, v100.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v15.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v82.l ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l ; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v149, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v16.l, v98.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v16.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.h ; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v149, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v17.l, v97.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v17.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v70.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v68.l ; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v38.h ; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v149, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v18.l, v87.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v18.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v66.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v37.l ; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v149, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v19.l, v85.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v19.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v149, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v20.l, v83.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v20.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v55.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v149, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v21.l, v81.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v21.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v53.l ; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v33.l ; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v149, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v22.l, v71.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v22.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v150.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v150.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v151.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v151.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v146.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v147.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v147.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v148.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v134.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v144.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v130.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v131.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v131.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v117.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v118.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v113.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v113.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v114.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v114.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v100.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v101.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v102.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v96.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v97.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v97.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v83.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v84.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v85.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v70.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v80.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v66.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v67.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v67.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v53.h +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v55.h +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v50.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v30.l, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v51.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v52.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 @@ -88560,433 +88052,329 @@ define <16 x i64> @bitcast_v128i8_to_v16i64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v149, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v23.l, v70.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v23.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v51.l -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v149, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v24.l, v67.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v24.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v149, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v25.l, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v25.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v149, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v26.l, v64.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v26.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v149, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v27.l, v54.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v27.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v149, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v28.l, v53.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v28.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v149, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v29.l, v52.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v29.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v149, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v30.l, v51.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v30.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v149, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v31.l, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v149, v31 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB58_2 ; GFX11-TRUE16-NEXT: .LBB58_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v146.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v145.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v133.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v129.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v128.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v116.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v103.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v99.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v99.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v87.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v87.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v86.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v83.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v82.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v82.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v70.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v69.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v68.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v65.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v65.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v64.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v53.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v50.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v48.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, v32.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v134.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v31, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v148.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v130.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v130.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v144.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v145.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v118.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v117.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v133.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v133.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v131.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v132.l, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v113.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v129.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v129.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v128.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v128.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v103.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v101.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v117.h, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v118.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v116.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v116.h, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v99.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v96.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v96.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v113.l, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v86.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v84.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v84.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v102.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v102.h, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v16 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v100.h, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.l, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v82.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v80.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v80.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v98.h, v15.l -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v99.l, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v97.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v97.h, v16.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v68.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v67.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v87.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v87.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l ; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v31, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v85.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v85.h, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v65.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v31, v21 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v55.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v50.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v83.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v83.h, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v31, v22 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v21.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v81.h, v20.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v82.l, v20.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v31, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.l, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v71.h, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v31, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v70.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v70.h, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v31, v25 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v67.h, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v68.l, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v26 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v25.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.l, v24.l -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v66.h, v24.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v27 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v25.h -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v64.l, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v64.h, v25.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v31, v28 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v55.l, v26.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v31, v29 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v53.h, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v54.l, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v31, v30 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v34.h, 0x300, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v53.l, v28.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v31, v34 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v29.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v51.h, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v52.l, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h ; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l ; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v33 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v50.h, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v51.l, v30.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v31, v33 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.h, 0x300, v32.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v150.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v150.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v151.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v151.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v147.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v148.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v134.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v144.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v144.h, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v131.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v131.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v132.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v117.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v117.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v118.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v118.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v119.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v115.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v100.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v101.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v102.l, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v102.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v96.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v97.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v97.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v98.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v83.h, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v84.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v84.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v85.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v85.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.h, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v71.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v80.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v80.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v67.l, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v67.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v68.l, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v53.h, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v54.h, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v55.l, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v55.h, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v51.l, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v52.l, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v52.h, v31.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v16.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v17.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v18.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v19.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v20.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v21.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v22.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v23.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v24.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v25.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v26.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v26.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v27.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v29.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v30.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v31.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v31.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -113114,271 +112502,205 @@ define <128 x i8> @bitcast_v16f64_to_v128i8(<16 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v39.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v160.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v33.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v36.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v150.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v149.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v149.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v148.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v147.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v146.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v145.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v71.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v69.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v68.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v133.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v132.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v131.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v130.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v9.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v10.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v50.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v117.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v115.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v113.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v99.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v15.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v16.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v87.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v17.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v19.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v21.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v81.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v23.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v71.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v25.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v26.h, v27.l ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v16f64_to_v128i8: @@ -123405,61 +122727,61 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v32, off, s32 offset:380 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v31, off, s32 offset:376 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v32, off, s32 offset:372 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:368 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:368 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v33, off, s32 offset:364 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v51, off, s32 offset:360 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v33, off, s32 offset:356 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v51, off, s32 offset:352 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v50, off, s32 offset:352 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v34, off, s32 offset:348 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v52, off, s32 offset:344 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v34, off, s32 offset:340 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v52, off, s32 offset:336 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v35, off, s32 offset:332 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:328 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:328 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v35, off, s32 offset:324 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:320 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:320 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v36, off, s32 offset:316 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v54, off, s32 offset:312 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v53, off, s32 offset:312 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v36, off, s32 offset:308 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v54, off, s32 offset:304 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:304 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v37, off, s32 offset:300 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v55, off, s32 offset:296 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:296 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v37, off, s32 offset:292 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:288 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:288 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v38, off, s32 offset:284 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:280 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:280 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:276 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v66, off, s32 offset:272 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v39, off, s32 offset:268 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v66, off, s32 offset:264 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:264 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v39, off, s32 offset:260 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v67, off, s32 offset:256 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:256 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v48, off, s32 offset:252 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v68, off, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:248 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v48, off, s32 offset:244 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:240 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:240 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v49, off, s32 offset:236 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v70, off, s32 offset:232 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v49, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v71, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:224 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v50, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v71, off, s32 offset:216 -; GFX11-TRUE16-NEXT: scratch_load_b32 v114, off, s32 offset:388 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:40 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:48 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_b32 v113, off, s32 offset:388 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v83, off, s32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:8 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v85, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v85, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:48 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v97, off, s32 offset:56 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v97, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v100, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v101, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v102, off, s32 offset:104 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v102, off, s32 offset:112 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v160, off, s32 offset:120 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v160, off, s32 offset:128 @@ -123474,121 +122796,123 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v164, off, s32 offset:192 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v165, off, s32 offset:200 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v165, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v55, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:204 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:196 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v67, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v80, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v80, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v53, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v64, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v64, off, s32 offset:196 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v65, off, s32 offset:188 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v65, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v68, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v69, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v69, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v70, off, s32 offset:148 ; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v81, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v84, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v84, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v96, off, s32 offset:92 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v96, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v98, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:60 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v101, off, s32 offset:52 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v113, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v81, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v82, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v82, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v83, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v86, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v86, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v87, off, s32 offset:84 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v87, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v98, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v99, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v99, off, s32 offset:52 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v100, off, s32 offset:44 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v103, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v103, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v112, off, s32 offset:20 ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_load_d16_b16 v115, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v112, off, s32 offset:12 ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v115, off, s32 offset:4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.l, v30.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v118.h, v28.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.l, v26.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.l, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v116.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v119.h, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v132.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v144.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v145.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v146.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v145.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v150.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v151.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v146.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v147.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v148.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v134.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.l, 8, v21.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v135.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v133.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v29.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(54) -; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v114 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(17) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v81.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(16) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.l, 8, v82.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v129.h, 8, v83.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.l, 8, v83.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v128.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v85.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v87.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.l, 8, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v116.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v97.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(9) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v98.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(3) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v112.h, 8, v99.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v100.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v101.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v102.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.h, 8, v161.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v99.l, 8, v161.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v162.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v162.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.l, 8, v163.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v87.h, 8, v163.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v164.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v164.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.l, 8, v165.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v81.h, 8, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v82.l, 8, v71.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.l, 8, v68.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v66.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.l, 8, v55.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v64.h, 8, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v52.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v52.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v51.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v31.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v144.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.l, 8, v29.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(62) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v51.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v51.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(56) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v50.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v54.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(26) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.h, 8, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.l, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v66.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v71.l, 8, v71.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(18) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v70.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v113 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(13) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v130.h, 8, v83.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v131.h, 8, v84.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v132.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.l, 8, v85.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v117.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.l, 8, v96.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v118.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v119.l, 8, v97.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(8) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.l, 8, v98.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v113.h, 8, v100.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.l, 8, v101.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v114.h, 8, v101.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v115.l, 8, v102.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v100.h, 8, v102.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.l, 8, v160.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v101.h, 8, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.l, 8, v161.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v102.h, 8, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.l, 8, v162.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v96.h, 8, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.l, 8, v163.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v97.h, 8, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v98.l, 8, v164.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v83.h, 8, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.l, 8, v165.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v84.h, 8, v165.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v85.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v80.h, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v67.h, 8, v55.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v68.l, 8, v55.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v31.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) @@ -123599,215 +122923,179 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB74_4 ; GFX11-TRUE16-NEXT: .LBB74_2: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB74_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v149.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v146.l -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v0.l, v151.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v151.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v0.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v150.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v149, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v1.h, v150.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v146.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v145.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v133.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v133.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v132.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v131.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v148.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v130.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v149, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v2.l, v148.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v2.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v134.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v134.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v145.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v130.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v149, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v3.l, v147.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v147.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v3.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v119.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v118.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v149, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v4.l, v144.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v4.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v133.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v117.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v129.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v129.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v128.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v119.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v116.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v116.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v115.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v115.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v149, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v5.l, v135.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v5.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v132.l -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v129.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v112.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v149, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v6.l, v133.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v6.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v128.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v103.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v101.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v149, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v7.l, v131.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v7.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v100.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v99.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v98.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v149, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v8.l, v129.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v8.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v116.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v114.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v96.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v149, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v9.l, v128.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v9.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v113.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v84.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v149, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v10.l, v117.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v10.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v102.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v84.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v112.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v103.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v99.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v99.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v87.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v87.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v86.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v82.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v81.h ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v81.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v149, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v11.l, v116.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v11.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v80.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v149, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v12.l, v114.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v12.l, v149.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v97.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v69.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v68.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v149, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v13.l, v112.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v13.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v87.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v67.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v65.h -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v65.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v149, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v14.l, v102.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v14.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v85.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v83.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v69.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v68.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v65.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v53.l ; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v149, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v15.l, v100.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v15.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v82.l ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v49.l ; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v149, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v16.l, v98.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v16.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.h ; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v48.l ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v39.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v149, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v17.l, v97.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v17.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v70.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v68.l ; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v38.h ; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v149, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v18.l, v87.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v18.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v66.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v37.l ; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v149, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v19.l, v85.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v19.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v149, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v20.l, v83.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v20.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v55.l -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v54.l ; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v149, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v21.l, v81.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v21.l, v149.h ; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v53.l ; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v33.l ; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v149, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v22.l, v71.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v22.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v150.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v150.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v151.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v151.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v146.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v147.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v147.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v148.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v134.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v135.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v144.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v130.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v130.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v131.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v131.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v117.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v117.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v118.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v118.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v113.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v113.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v114.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v114.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v115.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v100.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v101.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v101.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v102.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v17.l, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v96.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.l, v97.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v97.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v19.l, v98.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v83.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v84.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v84.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v21.l, v85.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v85.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v22.l, v70.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v71.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.l, v71.h +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v80.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v24.l, v80.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v66.l +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v25.l, v66.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v67.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v26.l, v67.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v68.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v27.l, v53.h +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v28.l, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v55.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v29.l, v55.h +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v50.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v30.l, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v51.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v31.l, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v52.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 @@ -123829,433 +123117,329 @@ define <16 x double> @bitcast_v128i8_to_v16f64(<128 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v149, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v23.l, v70.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v23.l, v149.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v51.l -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v149, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v24.l, v67.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v24.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v149, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v25.l, v66.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v25.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v149, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v26.l, v64.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v26.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v149, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v27.l, v54.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v27.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v149, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v28.l, v53.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v28.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v149, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v29.l, v52.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v29.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v149, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v30.l, v51.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v30.l, v149.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v149, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v149.l, v31.l, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v149, v31 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB74_2 ; GFX11-TRUE16-NEXT: .LBB74_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v149.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v149.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v146.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v146.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v145.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v133.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v133.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v129.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v129.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v128.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v128.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v116.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v116.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v112.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v103.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v99.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v99.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v98.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v87.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v87.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v86.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v83.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v82.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v82.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v81.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, v70.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, v69.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v68.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, v65.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, v65.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v64.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v64.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, v53.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, v50.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, v48.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, v48.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, v38.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, v36.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, v32.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v151.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v151.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v150.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v150.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v145.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v144.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v134.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v31, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v134.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v148.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v147.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v132.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v131.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v130.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v130.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v144.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v145.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v119.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v119.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v118.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v117.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v133.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v133.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v131.h, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v132.l, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v115.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v115.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v113.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v112.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v129.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v129.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v128.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v128.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v103.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v103.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v101.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v100.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v117.h, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v118.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v116.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v116.h, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v99.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v98.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v96.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v96.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v112.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v113.l, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v86.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v86.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v84.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v84.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v102.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v102.h, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v16 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v100.h, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.l, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v82.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v81.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, v80.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, v80.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v98.h, v15.l -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v99.l, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v97.l, v16.l -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v97.h, v16.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v69.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v69.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v19 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v17.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v17.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, v68.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, v67.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v87.l, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v87.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l ; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v31, v20 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v85.l, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v85.h, v18.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v65.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v65.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v31, v21 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v19.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, v55.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, v50.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v83.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v83.h, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v31, v22 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v21.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v81.h, v20.l -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v82.l, v20.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v31, v23 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v21.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, v48.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.l, v21.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v71.h, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v31, v24 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v70.l, v22.l -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v70.h, v22.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v39.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v39.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v31, v25 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v23.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v23.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, v38.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v67.h, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v68.l, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v31, v26 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v25.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.l, v24.l -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v66.h, v24.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v37.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v31, v27 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v25.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v25.h -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v64.l, v25.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v64.h, v25.h +; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h ; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v31, v28 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v54.h, v26.l -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v55.l, v26.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v31, v29 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v27.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v53.h, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v54.l, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v31, v30 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v34.h, 0x300, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v52.h, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v53.l, v28.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v33.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v31, v34 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v29.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v29.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v51.h, v29.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v52.l, v29.h +; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h ; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l ; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v31, v33 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v33.h, 0x300, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.l, v50.h, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v51.l, v30.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v31, v33 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v32.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v32.h, 0x300, v32.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v31, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v150.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v150.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v151.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v151.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v146.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v147.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v147.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v148.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v148.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v134.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v135.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v135.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v144.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v144.h, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v130.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v130.h, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v131.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v131.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v132.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v117.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v117.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v118.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v118.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v119.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v113.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v113.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v114.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v114.h, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v115.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v100.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v101.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v101.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v102.l, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v102.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v96.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v96.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v97.l, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v97.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v98.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v83.h, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v84.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v84.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v85.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v85.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v70.h, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v71.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v71.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v80.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v80.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v66.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v25.l, v66.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v67.l, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v26.l, v67.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v68.l, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v53.h, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v54.l, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v28.l, v54.h, v28.l +; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v55.l, v28.h +; GFX11-TRUE16-NEXT: v_or_b16 v29.l, v55.h, v29.l +; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v50.h, v29.h +; GFX11-TRUE16-NEXT: v_or_b16 v30.l, v51.l, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v51.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v52.l, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v52.h, v31.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v15.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.l, 0x300, v16.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v16.h, 0x300, v16.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.l, 0x300, v17.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v17.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.l, 0x300, v18.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v18.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.l, 0x300, v19.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v19.h, 0x300, v19.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.l, 0x300, v20.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v20.h, 0x300, v20.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v21.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.h, 0x300, v21.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.l, 0x300, v22.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v22.h, 0x300, v22.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.l, 0x300, v23.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v23.h, 0x300, v23.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.l, 0x300, v24.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v24.h, 0x300, v24.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v25.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.h, 0x300, v25.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.l, 0x300, v26.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v26.h, 0x300, v26.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.l, 0x300, v27.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v27.h, 0x300, v27.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.l, 0x300, v28.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v28.h, 0x300, v28.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.l, 0x300, v29.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v29.h, 0x300, v29.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.l, 0x300, v30.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v30.h, 0x300, v30.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.l, 0x300, v31.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v31.h, 0x300, v31.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -161654,179 +160838,182 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:236 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:232 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:224 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:216 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:204 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:200 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:196 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:192 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:184 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:176 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:168 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:160 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:152 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:148 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:144 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:136 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:128 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:120 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:112 -; GFX11-TRUE16-NEXT: s_clause 0x18 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:92 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:60 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:52 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:40 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:20 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v40, s32 offset:248 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v41, s32 offset:244 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v42, s32 offset:240 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v43, s32 offset:236 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v44, s32 offset:232 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v45, s32 offset:228 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v46, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v47, s32 offset:220 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v56, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v57, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v58, s32 offset:208 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v59, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v60, s32 offset:200 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v61, s32 offset:196 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v62, s32 offset:192 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v63, s32 offset:188 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v72, s32 offset:184 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v73, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v74, s32 offset:176 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v75, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v76, s32 offset:168 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v77, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v78, s32 offset:160 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v79, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v88, s32 offset:152 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v89, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v90, s32 offset:144 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v91, s32 offset:140 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v92, s32 offset:136 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v93, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v94, s32 offset:128 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v95, s32 offset:124 +; GFX11-TRUE16-NEXT: s_clause 0x1b +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v104, s32 offset:120 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v105, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v106, s32 offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v107, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v108, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v109, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v110, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v111, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v120, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v121, s32 offset:84 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v122, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v123, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v124, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v125, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v126, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v127, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v136, s32 offset:56 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v137, s32 offset:52 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v138, s32 offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v139, s32 offset:44 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v140, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v141, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v142, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v143, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v152, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v153, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v154, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_store_b32 off, v155, s32 offset:12 ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32 offset:8 -; GFX11-TRUE16-NEXT: scratch_load_b32 v99, off, s32 offset:4 -; GFX11-TRUE16-NEXT: scratch_load_b32 v98, off, s32 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr178_hi16 +; GFX11-TRUE16-NEXT: scratch_load_b32 v81, off, s32 offset:4 +; GFX11-TRUE16-NEXT: scratch_load_b32 v80, off, s32 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr181_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr152_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr177_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr180_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr143_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr141_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr183_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr140_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr139_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr40_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr138_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr179_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr137_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr56_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr126_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr42_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr182_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr127_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr125_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr41_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr123_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr40_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr121_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr79_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr111_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr60_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr72_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr109_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr46_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr107_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr106_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr95_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr104_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr92_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr76_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr93_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr74_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr91_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr127_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr89_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr136_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr79_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr104_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr106_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr77_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr75_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr142_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr73_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr125_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr63_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr153_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr62_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr139_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr61_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr143_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr58_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr141_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr59_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr155_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr57_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr154_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr47_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr124_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr44_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr142_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr122_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr110_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr138_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr137_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr126_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr108_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr105_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr94_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr124_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr122_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr120_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr92_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr90_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr88_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr110_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr108_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr95_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr77_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr74_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr72_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr94_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr93_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr90_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr62_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr59_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr57_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr46_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr44_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr88_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr78_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr167_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr76_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr43_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr41_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr183_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr73_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr63_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr177_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr60_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr176_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr182_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr167_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr181_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr180_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr58_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr56_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr179_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr45_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr178_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr43_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr42_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) ; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v31 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 @@ -161835,136 +161022,136 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB90_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[100:101], 24, v[15:16] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[84:85], 24, v[27:28] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[101:102], 24, v[13:14] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[96:97], 24, v[15:16] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[99:100], 24, v[13:14] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[114:115], 24, v[11:12] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[117:118], 24, v[9:10] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[130:131], 24, v[7:8] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[144:145], 24, v[3:4] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[85:86], 24, v[25:26] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v44, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v73, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 8, v13 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v78, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v89, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v91, 24, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v93, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v95, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v109, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v111, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v121, 24, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v123, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v126, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v136, 24, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v137, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v138, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v139, 24, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v140, 8, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v77, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v89, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v91, 8, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v92, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v109, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v111, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v121, 8, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v123, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v125, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v127, 8, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v140, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v141, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v143, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v152, 8, v1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v180, 24, v99 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 8, v99 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 24, v81 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v81 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v98 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 24, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v41, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v44, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 8, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v77, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v92, 8, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v21 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v110, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v120, 24, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v122, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v124, 8, v17 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[131:132], 24, v[5:6] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 8, v80 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v73, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v76, 8, v27 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v78, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v93, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v95, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v110, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v120, 8, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v122, 24, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v124, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v126, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v137, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v138, 8, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v142, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[128:129], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[134:135], 24, v[5:6] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[145:146], 24, v[1:2] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[98:99] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[29:30] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[86:87], 24, v[23:24] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[102:103], 24, v[21:22] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[115:116], 24, v[19:20] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[118:119], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v178.h, v1.l +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[82:83], 24, v[80:81] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[86:87], 24, v[29:30] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[100:101], 24, v[27:28] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[97:98], 24, v[25:26] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[112:113], 24, v[23:24] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[115:116], 24, v[21:22] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[118:119], 24, v[19:20] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[132:133], 24, v[17:18] +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v181.h, v1.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v177.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v180.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.h, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v40.h, v3.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v183.h, v3.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.h, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v179.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v182.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v56.h, v5.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v41.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.h, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v42.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v40.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.h, v6.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v79.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v60.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v106.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.h, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v76.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.h, v10.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v127.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.h, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v104.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v142.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.h, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v125.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.h, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v143.h, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.h, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v141.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.h, v16.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v135.h, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v72.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.h, v7.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v46.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.h, v8.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v104.h, v9.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.h, v9.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v74.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v10.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v136.h, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v102.h, v11.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v106.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.h, v12.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v153.h, v13.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.h, v13.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v139.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.h, v14.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v155.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v154.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, v17.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.h, v19.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.h, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.h, v20.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v150.h, v21.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v161.h, v21.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v21.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v149.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v160.h, v22.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v22.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v160.h, v23.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v163.h, v23.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.h, v23.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v151.h, v24.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v162.h, v24.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.h, v24.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v162.h, v25.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v165.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v25.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v161.h, v26.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v164.h, v26.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v26.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v164.h, v27.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v167.h, v27.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.h, v27.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v163.h, v28.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v166.h, v28.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v51.h, v28.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v166.h, v29.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v177.h, v29.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v165.h, v30.l +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v176.h, v30.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v30.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v176.h, v98.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.h, v98.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v167.h, v99.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v99.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v179.h, v80.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.h, v80.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v178.h, v81.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v81.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5 @@ -161980,7 +161167,7 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81 ; GFX11-TRUE16-NEXT: .LBB90_2: ; %Flow ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB90_4 @@ -162019,10 +161206,10 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v37, v48, v17, 0x7fff ; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v120, 24, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v122, 8, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v137, 24, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v138, 8, v32 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v135, v37, v49, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v149, v37, v49, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v34, v34 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v19 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v19, 16, v19 @@ -162036,97 +161223,101 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v20, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v34, 0x400000, v18 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v135.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v31.l, v149.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v36, 16, 1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v147, v33, v35, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v150, v33, v35, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v19 ; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v36, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, 0x400000, v36 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v124, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[132:133], 24, v[31:32] ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v34, v17, v34, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v34.l, v147.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v142, 8, v31 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v19, 0x7fff -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v34 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 8, v34 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v148, v17, v33, vcc_lo +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v151, v17, v33 :: v_dual_and_b32 v18, 0xffff0000, v22 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v36, v36 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v33, v20, v35 :: v_dual_and_b32 v18, 0xffff0000, v22 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_cndmask_b32 v33, v20, v35 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v21 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v21 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v22 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v18 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_add_f32 v22, 0x40c00000, v22 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_add_f32 v21, 0x40c00000, v21 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v34.l, v150.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v33.l, v151.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v22, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, 0x400000, v22 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v122, 24, v34 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v124, 8, v34 ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v22, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v149, v19, v35 :: v_dual_lshlrev_b32 v22, 16, v24 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v126, 8, v33 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v160, v19, v35, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v24 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v21 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v17, v36, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v21, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v36.l, v149.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v21, 0x7fff ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v23 ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, 0x400000, v20 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v150, v17, v24, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v161, v17, v24, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v20, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v23 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v22 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v21 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 24, v36 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v36.l, v160.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v35, v19, v35, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v22, 16, 1 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v33.l, v148.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v92, 8, v36 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v35.l, v161.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v22, 0x7fff ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v26 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v110, 8, v33 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v151, v19, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v120, 8, v35 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v162, v19, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v22, 0x40c00000, v22 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v20, 0x7fff ; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v18 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v26 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_4) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v38, v17, v24, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v21, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v20 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v38.l, v151.h ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v21, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v160, v17, v23 :: v_dual_lshlrev_b32 v21, 16, v25 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v108, 24, v36 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v110, 8, v36 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v163, v17, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v22 @@ -162139,8 +161330,10 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v18 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v22, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v161, v19, v23 :: v_dual_lshlrev_b32 v22, 16, v28 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v28 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v37.l, v163.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v164, v19, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21 @@ -162153,10 +161346,9 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v21, 0x7fff ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v27 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v49.l, v161.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v35.l, v150.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v162, v17, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v95, 8, v37 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v165, v17, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v22 @@ -162169,10 +161361,10 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v22, 0x7fff ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 24, v49 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 8, v49 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v35 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v163, v19, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v38.l, v162.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v48.l, v165.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v166, v19, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21 @@ -162185,10 +161377,10 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v21, 0x7fff ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v29 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v51.l, v163.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v74, 24, v38 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v77, 8, v38 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v164, v17, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[112:113], 24, v[37:38] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v93, 24, v38 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v94, 8, v38 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v167, v17, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v22 @@ -162201,14 +161393,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v22, 0x7fff ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v99 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v44, 24, v51 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v46, 8, v51 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v165, v19, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v81 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v50.l, v167.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v176, v19, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v99 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v81 ; GFX11-TRUE16-NEXT: v_dual_add_f32 v22, 0x40c00000, v22 :: v_dual_cndmask_b32 v53, v17, v24 ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v21, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21 @@ -162217,14 +161409,14 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v21, 0x7fff ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v98 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v37.l, v160.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v53.l, v165.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v166, v17, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v80 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v49.l, v164.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v177, v17, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v22 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v98 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v80 ; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_cndmask_b32 v52, v19, v24 ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v22, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 @@ -162233,10 +161425,10 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20 ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v22, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v21 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v183, 24, v53 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v41, 8, v53 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v37 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v167, v19, v23, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v52.l, v177.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v78, 24, v49 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v88, 8, v49 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v178, v19, v23, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v2 @@ -162249,10 +161441,9 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v21, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v55.l, v167.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v176, v17, v22, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v179, v17, v22, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 ; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v1 @@ -162263,13 +161454,12 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff ; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v20, 0x40c00000, v20 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v180, 24, v55 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v181, 8, v55 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v54.l, v179.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v20, 16, 1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v177, v19, v21, vcc_lo +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v180, v19, v21, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v4 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 @@ -162282,11 +161472,10 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18 ; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v48.l, v162.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v65.l, v177.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v90, 8, v48 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v18, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v178, v17, v19, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v181, v17, v19, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v3 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 @@ -162301,9 +161490,9 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v4, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v139, 24, v65 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v140, 8, v65 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v179, v2, v19, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v51.l, v166.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v64.l, v181.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v182, v2, v19, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18 ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v17, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v3 @@ -162313,13 +161502,13 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v17, 0x7fff ; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v67.l, v179.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 24, v51 ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v5 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v50.l, v164.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v136, 24, v67 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v40, v1, v18, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v53.l, v176.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v73, 8, v51 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v183, v1, v18, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v4, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v6 @@ -162330,13 +161519,13 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v4, 0x7fff ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[84:85], 24, v[50:51] +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v66.l, v183.h ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v6, 0x7fff ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[85:86], 24, v[48:49] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[86:87], 24, v[37:38] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v137, 8, v67 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v42, v2, v17, vcc_lo +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v56, 24, v53 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v53 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v76, 8, v50 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v40, v2, v17, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v8 @@ -162350,28 +161539,27 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v6, 16, 1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 8, v50 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v56, v2, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v41, v2, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v4, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v3, v5, v6, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v6 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v72, 8, v48 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v68, v1, v17, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v2, v4, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v7 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v10 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v56.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.l, v41.h ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v60, v3, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v46, v3, v5, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v7 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v10 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v83, v1, v8 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v71, v1, v8 ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 @@ -162380,29 +161568,29 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v6, 16, 1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.l, v60.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v52.l, v166.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v82, v1, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.l, v46.h +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[86:87], 24, v[52:53] +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v70, v1, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v9 ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v6, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v6 ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v79, v4, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v72, v4, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v12 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v11 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.l, v79.h -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v76, v1, v4 :: v_dual_lshlrev_b32 v1, 16, v9 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v55.l, v178.h +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v74, v1, v4 :: v_dual_lshlrev_b32 v1, 16, v9 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v12 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v97, v2, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v2, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 @@ -162410,40 +161598,40 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v13 ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.l, v76.h -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v54.l, v176.h -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[130:131], 24, v[82:83] -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v96, v2, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v67.l, v182.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.l, v74.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.l, v72.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v84, v2, v6, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v7, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[54:55] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[52:53] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[128:129], 24, v[70:71] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[82:83], 24, v[54:55] ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v6, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v106, v2, v3, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v104, v2, v3, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v8 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_add3_u32 v3, v7, v4, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v14 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.l, v106.h -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v104, v1, v2 :: v_dual_lshlrev_b32 v1, 16, v11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v84.l, v104.h +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v106, v1, v2 :: v_dual_lshlrev_b32 v1, 16, v11 ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v6, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[117:118], 24, v[96:97] +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v65.l, v180.h ; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v113, v3, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v103, v3, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v6 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.l, v104.h -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[118:119], 24, v[31:32] -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v112, v2, v3, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.l, v106.h +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[117:118], 24, v[84:85] +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v102, v2, v3, vcc_lo ; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v4 :: v_dual_add_f32 v3, 0x40c00000, v5 ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v1 @@ -162452,8 +161640,8 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v64.l, v178.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v127, v4, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[118:119], 24, v[33:34] +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v136, v4, v5, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v2, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 @@ -162461,19 +161649,19 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v1, 16, 1 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v9 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v129, v4, v5, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v131, v4, v5, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 ; GFX11-TRUE16-NEXT: v_add3_u32 v3, v8, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v16 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v15 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v125, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v139, v6, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v5 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.l, v40.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v128, v3, v4, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v131.l, v139.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v130, v3, v4, vcc_lo ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v16 ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v6, v2, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 @@ -162481,11 +161669,11 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v129.l, v125.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.l, v127.h +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v102.l, v136.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.l, v40.h ; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v142, v4, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v153, v4, v6, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 @@ -162494,389 +161682,322 @@ define <128 x i8> @bitcast_v64bf16_to_v128i8(<64 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v8 ; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v128.l, v142.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v141, v2, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v130.l, v153.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v154, v2, v9, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.l, v42.h -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[114:115], 24, v[112:113] -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v143, v7, v11, vcc_lo +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[99:100], 24, v[130:131] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[114:115], 24, v[102:103] +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v155, v7, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v10, v6, 0x7fff ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[144:145], 24, v[66:67] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[131:132], 24, v[68:69] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[134:135], 24, v[68:69] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[145:146], 24, v[64:65] -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v134, v4, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v148, v4, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v134.l, v141.h -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[115:116], 24, v[33:34] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 24, v129 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v63, 8, v129 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v133, v2, v3, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e64 v133.l, v143.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 24, v134 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 8, v134 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v73, 8, v128 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 24, v113 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[100:101], 24, v[133:134] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[101:102], 24, v[128:129] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[102:103], 24, v[35:36] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v58, 8, v133 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v78, 8, v113 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v89, 8, v112 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v91, 24, v97 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v93, 8, v97 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v95, 8, v96 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 24, v83 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v109, 8, v83 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v111, 8, v82 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v121, 24, v69 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v123, 8, v69 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v126, 8, v68 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v138, 8, v66 +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v148.l, v154.h +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[100:101], 24, v[50:51] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[115:116], 24, v[35:36] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v59, 24, v131 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v147, v2, v3, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e64 v147.l, v155.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v44, 24, v148 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v47, 8, v148 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v61, 8, v131 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v62, 8, v130 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[96:97], 24, v[147:148] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[97:98], 24, v[48:49] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v57, 8, v147 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v75, 24, v103 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v77, 8, v103 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v79, 8, v102 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v89, 24, v85 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v91, 8, v85 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v92, 8, v84 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v105, 24, v71 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v107, 8, v71 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v109, 8, v70 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v111, 24, v69 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v121, 8, v69 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v123, 8, v68 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v125, 24, v67 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v127, 8, v67 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v140, 8, v66 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v141, 24, v65 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v143, 8, v65 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v152, 8, v64 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v182, 8, v54 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v52 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v42, 24, v55 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v43, 8, v55 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v45, 8, v54 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v60, 8, v52 ; GFX11-TRUE16-NEXT: .LBB90_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v178.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v181.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v152.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v64.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v65.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.l, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v139.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v180.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v143.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v141.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v183.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v140.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v144.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v1.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v177.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v140.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v67.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v40.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v138.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v136.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v182.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v127.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v125.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v41.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v123.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v68.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v179.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v137.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v134.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v40.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v121.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v5, v3 -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v82.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v56.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v126.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v83.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v107.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v5.l, v5.h ; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v42.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v123.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v96.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v5, v6 -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v97.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v79.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v111.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v91.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v112.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v5, v7 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v60.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v109.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v75.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v5, v8 -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v128.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v106.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v95.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v129.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v5, v9 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v61.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v111.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v72.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v109.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v46.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v107.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v105.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v104.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v92.l +; GFX11-TRUE16-NEXT: s_clause 0x1 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v85.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v89.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v136.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v79.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v102.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v106.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v77.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v103.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v75.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v84.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v117.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v74.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v91.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v153.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v62.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v130.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v139.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v61.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v131.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v59.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v155.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v57.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v147.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v96.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v154.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v47.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v148.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v44.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v149.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v142.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v132.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v138.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v137.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v151.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v126.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v118.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v150.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v124.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h ; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v76.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v93.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v133.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v5, v10 -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v134.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v127.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v89.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v45.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v5, v11 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v104.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v78.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v120.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v5, v12 -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v142.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v73.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v34.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v5, v13 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v105.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v125.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v63.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v5, v14 -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v122.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v161.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v120.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v35.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v115.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v160.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v110.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v108.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v15.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v16.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v163.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v95.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v37.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v162.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v94.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v93.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v165.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v90.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h ; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v143.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v58.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v90.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v5, v15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v141.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v47.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v74.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v5, v16 -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v135.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v124.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v5, v17 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v59.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v18.l, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v122.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v50.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v5, v18 -; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v51.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v19.l, v19.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v20.l, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v148.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v110.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v44.l -; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v5, v19 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v80.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v20.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v17.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v48.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v164.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v88.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v49.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v78.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v167.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v76.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v50.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v100.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v19.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v21.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v166.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v73.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v51.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v63.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v177.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v60.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v52.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v176.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v58.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h ; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v147.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v108.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v53.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v183.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v5, v20 -; GFX11-TRUE16-NEXT: v_and_b16 v33.l, 0xff, v54.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v150.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v94.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_and_b16 v34.l, 0xff, v55.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v5, v21 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v180.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v22.l, v22.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v149.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v92.l -; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[6:9], off offset:16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v5, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v23.l, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v160.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v88.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v5, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v24.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v25.l, v25.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v151.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v77.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v5, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v25.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v26.l, v26.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v162.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v72.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v5, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v26.l, v26.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v27.l, v27.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v161.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v62.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v5, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v27.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v28.l, v28.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v164.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v57.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v5, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v28.l, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v29.l, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v163.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v46.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v5, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.l, v29.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v30.l, v30.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v166.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v43.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v5, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v30.l, v30.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v31.l, v31.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v165.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v31.h, 8, v41.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v5, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v31.l, v31.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v32.l, v32.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v176.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v32.h, 8, v182.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v5, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v32.l, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v33.l, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v33.l, 0xff, v167.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v181.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v5, v32 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v33.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v33.h, v34.l, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v5.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, v5, v33 +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v23.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v56.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v179.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v45.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v178.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v43.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v55.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v42.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v25.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v26.h, v27.l ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[10:13], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[14:17], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[18:21], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[22:25], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[26:29], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[30:33], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 ; GFX11-TRUE16-NEXT: s_clause 0x1f -; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:12 -; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:16 -; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:20 -; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:24 -; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:28 -; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:32 -; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:36 -; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:40 -; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:44 -; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:48 -; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:52 -; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:56 -; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:60 -; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:64 -; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:68 -; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:72 -; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:76 -; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:80 -; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:84 -; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:88 -; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:92 -; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:96 -; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:100 -; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:104 -; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:108 -; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:112 -; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:116 -; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:120 -; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:124 -; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:128 -; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:132 -; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:136 -; GFX11-TRUE16-NEXT: s_clause 0x18 -; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:140 -; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:144 -; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:148 -; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:152 -; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:156 -; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:160 -; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:164 -; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:168 -; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:172 -; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:176 -; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:180 -; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:184 -; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:188 -; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:192 -; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:196 -; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:200 -; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:204 -; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:208 -; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:212 -; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:216 -; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:220 -; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:224 -; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:228 -; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:232 -; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:236 +; GFX11-TRUE16-NEXT: scratch_load_b32 v155, off, s32 offset:12 +; GFX11-TRUE16-NEXT: scratch_load_b32 v154, off, s32 offset:16 +; GFX11-TRUE16-NEXT: scratch_load_b32 v153, off, s32 offset:20 +; GFX11-TRUE16-NEXT: scratch_load_b32 v152, off, s32 offset:24 +; GFX11-TRUE16-NEXT: scratch_load_b32 v143, off, s32 offset:28 +; GFX11-TRUE16-NEXT: scratch_load_b32 v142, off, s32 offset:32 +; GFX11-TRUE16-NEXT: scratch_load_b32 v141, off, s32 offset:36 +; GFX11-TRUE16-NEXT: scratch_load_b32 v140, off, s32 offset:40 +; GFX11-TRUE16-NEXT: scratch_load_b32 v139, off, s32 offset:44 +; GFX11-TRUE16-NEXT: scratch_load_b32 v138, off, s32 offset:48 +; GFX11-TRUE16-NEXT: scratch_load_b32 v137, off, s32 offset:52 +; GFX11-TRUE16-NEXT: scratch_load_b32 v136, off, s32 offset:56 +; GFX11-TRUE16-NEXT: scratch_load_b32 v127, off, s32 offset:60 +; GFX11-TRUE16-NEXT: scratch_load_b32 v126, off, s32 offset:64 +; GFX11-TRUE16-NEXT: scratch_load_b32 v125, off, s32 offset:68 +; GFX11-TRUE16-NEXT: scratch_load_b32 v124, off, s32 offset:72 +; GFX11-TRUE16-NEXT: scratch_load_b32 v123, off, s32 offset:76 +; GFX11-TRUE16-NEXT: scratch_load_b32 v122, off, s32 offset:80 +; GFX11-TRUE16-NEXT: scratch_load_b32 v121, off, s32 offset:84 +; GFX11-TRUE16-NEXT: scratch_load_b32 v120, off, s32 offset:88 +; GFX11-TRUE16-NEXT: scratch_load_b32 v111, off, s32 offset:92 +; GFX11-TRUE16-NEXT: scratch_load_b32 v110, off, s32 offset:96 +; GFX11-TRUE16-NEXT: scratch_load_b32 v109, off, s32 offset:100 +; GFX11-TRUE16-NEXT: scratch_load_b32 v108, off, s32 offset:104 +; GFX11-TRUE16-NEXT: scratch_load_b32 v107, off, s32 offset:108 +; GFX11-TRUE16-NEXT: scratch_load_b32 v106, off, s32 offset:112 +; GFX11-TRUE16-NEXT: scratch_load_b32 v105, off, s32 offset:116 +; GFX11-TRUE16-NEXT: scratch_load_b32 v104, off, s32 offset:120 +; GFX11-TRUE16-NEXT: scratch_load_b32 v95, off, s32 offset:124 +; GFX11-TRUE16-NEXT: scratch_load_b32 v94, off, s32 offset:128 +; GFX11-TRUE16-NEXT: scratch_load_b32 v93, off, s32 offset:132 +; GFX11-TRUE16-NEXT: scratch_load_b32 v92, off, s32 offset:136 +; GFX11-TRUE16-NEXT: s_clause 0x1b +; GFX11-TRUE16-NEXT: scratch_load_b32 v91, off, s32 offset:140 +; GFX11-TRUE16-NEXT: scratch_load_b32 v90, off, s32 offset:144 +; GFX11-TRUE16-NEXT: scratch_load_b32 v89, off, s32 offset:148 +; GFX11-TRUE16-NEXT: scratch_load_b32 v88, off, s32 offset:152 +; GFX11-TRUE16-NEXT: scratch_load_b32 v79, off, s32 offset:156 +; GFX11-TRUE16-NEXT: scratch_load_b32 v78, off, s32 offset:160 +; GFX11-TRUE16-NEXT: scratch_load_b32 v77, off, s32 offset:164 +; GFX11-TRUE16-NEXT: scratch_load_b32 v76, off, s32 offset:168 +; GFX11-TRUE16-NEXT: scratch_load_b32 v75, off, s32 offset:172 +; GFX11-TRUE16-NEXT: scratch_load_b32 v74, off, s32 offset:176 +; GFX11-TRUE16-NEXT: scratch_load_b32 v73, off, s32 offset:180 +; GFX11-TRUE16-NEXT: scratch_load_b32 v72, off, s32 offset:184 +; GFX11-TRUE16-NEXT: scratch_load_b32 v63, off, s32 offset:188 +; GFX11-TRUE16-NEXT: scratch_load_b32 v62, off, s32 offset:192 +; GFX11-TRUE16-NEXT: scratch_load_b32 v61, off, s32 offset:196 +; GFX11-TRUE16-NEXT: scratch_load_b32 v60, off, s32 offset:200 +; GFX11-TRUE16-NEXT: scratch_load_b32 v59, off, s32 offset:204 +; GFX11-TRUE16-NEXT: scratch_load_b32 v58, off, s32 offset:208 +; GFX11-TRUE16-NEXT: scratch_load_b32 v57, off, s32 offset:212 +; GFX11-TRUE16-NEXT: scratch_load_b32 v56, off, s32 offset:216 +; GFX11-TRUE16-NEXT: scratch_load_b32 v47, off, s32 offset:220 +; GFX11-TRUE16-NEXT: scratch_load_b32 v46, off, s32 offset:224 +; GFX11-TRUE16-NEXT: scratch_load_b32 v45, off, s32 offset:228 +; GFX11-TRUE16-NEXT: scratch_load_b32 v44, off, s32 offset:232 +; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s32 offset:236 +; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s32 offset:240 +; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s32 offset:244 +; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s32 offset:248 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -186713,69 +185834,69 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:8 ; GFX11-TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:4 ; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr176_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr167_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) ; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 @@ -186784,95 +185905,91 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB94_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[15:16] ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[13:14] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[11:12] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[5:6] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[27:28] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 24, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v21 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 24, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 8, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v27 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 8, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[9:10] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[3:4] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[1:2] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[25:26] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[23:24] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[21:22] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[19:20] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[17:18] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 8, v17 ; GFX11-TRUE16-NEXT: .LBB94_2: ; %Flow ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB94_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true +; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1) ; GFX11-TRUE16-NEXT: v_pk_add_f16 v32, 0x200, v32 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_pk_add_f16 v31, 0x200, v31 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v6, 0x200, v6 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v5, 0x200, v5 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v28, 0x200, v28 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v27, 0x200, v27 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v8, 0x200, v8 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v12, 0x200, v12 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v14, 0x200, v14 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v13, 0x200, v13 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v11, 0x200, v11 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v7, 0x200, v7 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v18, 0x200, v18 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v17, 0x200, v17 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v20, 0x200, v20 op_sel_hi:[0,1] @@ -186883,345 +186000,283 @@ define <128 x i8> @bitcast_v64f16_to_v128i8(<64 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_pk_add_f16 v23, 0x200, v23 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v26, 0x200, v26 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v25, 0x200, v25 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v30, 0x200, v30 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v29, 0x200, v29 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v2, 0x200, v2 op_sel_hi:[0,1] ; GFX11-TRUE16-NEXT: v_pk_add_f16 v1, 0x200, v1 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v16, 0x200, v16 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_pk_add_f16 v15, 0x200, v15 op_sel_hi:[0,1] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v4, 0x200, v4 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v3, 0x200, v3 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v10, 0x200, v10 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_pk_add_f16 v9, 0x200, v9 op_sel_hi:[0,1] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[15:16] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[13:14] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[11:12] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[5:6] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[27:28] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[9:10] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[3:4] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[25:26] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[23:24] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[21:22] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[19:20] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[17:18] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 24, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v21 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 24, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 8, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v27 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 8, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 8, v17 ; GFX11-TRUE16-NEXT: .LBB94_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v166.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v176.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v80.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v165.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v164.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v167.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v166.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v165.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v33.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v69.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v36.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v160.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v150.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v69.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v149.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v66.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v134.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v133.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v148.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v83.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v118.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v9.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v101.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v10.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v67.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v164.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v163.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v161.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v149.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v147.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v15.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v16.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v132.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v131.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v130.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v17.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v117.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v115.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v113.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v19.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v21.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v23.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v87.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v82.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v25.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v26.h, v27.l ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v64f16_to_v128i8: @@ -209415,69 +208470,69 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_b32 v33, off, s32 offset:8 ; GFX11-TRUE16-NEXT: scratch_load_b32 v32, off, s32 offset:4 ; GFX11-TRUE16-NEXT: scratch_load_b32 v31, off, s32 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr176_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr167_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr166_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr165_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr164_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr163_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr162_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr161_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr160_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr151_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr149_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr147_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr145_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr135_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr133_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr150_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr148_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr146_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr144_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr134_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr132_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(2) ; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v33 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 @@ -209486,95 +208541,91 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[15:16] ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[13:14] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[11:12] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[5:6] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[27:28] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 24, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v21 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 24, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 8, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v27 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 8, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[9:10] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[3:4] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[1:2] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[25:26] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[23:24] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[21:22] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[19:20] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[17:18] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 8, v17 ; GFX11-TRUE16-NEXT: .LBB98_2: ; %Flow ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true +; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(1) ; GFX11-TRUE16-NEXT: v_pk_add_u16 v32, v32, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_pk_add_u16 v31, v31, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v6, v6, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v5, v5, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v28, v28, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v27, v27, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v8, v8, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v12, v12, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v14, v14, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v13, v13, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v11, v11, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v7, v7, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v18, v18, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v17, v17, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v20, v20, 3 op_sel_hi:[1,0] @@ -209585,345 +208636,283 @@ define <128 x i8> @bitcast_v64i16_to_v128i8(<64 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_pk_add_u16 v23, v23, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v26, v26, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v25, v25, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v30, v30, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v29, v29, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v2, v2, 3 op_sel_hi:[1,0] ; GFX11-TRUE16-NEXT: v_pk_add_u16 v1, v1, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v16, v16, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_pk_add_u16 v15, v15, 3 op_sel_hi:[1,0] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v4, v4, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v3, v3, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v10, v10, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_pk_add_u16 v9, v9, 3 op_sel_hi:[1,0] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[15:16] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[13:14] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[11:12] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[66:67], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[5:6] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[33:34], 24, v[31:32] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[68:69], 24, v[5:6] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[37:38], 24, v[27:28] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[9:10] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[3:4] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[80:81], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[29:30] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[13:14] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[53:54], 24, v[11:12] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[9:10] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[69:70], 24, v[3:4] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[27:28] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[15:16] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[70:71], 24, v[1:2] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[25:26] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[21:22] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[54:55], 24, v[19:20] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[65:66], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 24, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v13 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v10 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 8, v9 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 8, v5 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 24, v4 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[38:39], 24, v[25:26] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[23:24] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[52:53], 24, v[21:22] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[64:65], 24, v[19:20] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[67:68], 24, v[17:18] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 24, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v16 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v13 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v12 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 8, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v133, 24, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 8, v8 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 8, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v160, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v162, 8, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 8, v3 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 24, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 8, v1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 24, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v32 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 8, v31 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 24, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v30 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v29 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 24, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v26 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v25 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 24, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 8, v23 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v21 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v134, 24, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v144, 8, v20 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v146, 8, v19 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v148, 24, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v150, 8, v18 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v37, 8, v17 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v165, 8, v3 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v166, 24, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v167, 8, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v176, 8, v1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 8, v31 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 24, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 8, v30 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v29 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 24, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v28 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v27 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v26 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v132, 8, v23 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v135, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v145, 8, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v147, 8, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v149, 24, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v151, 8, v20 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v161, 8, v19 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v163, 24, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v164, 8, v18 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v35, 8, v17 ; GFX11-TRUE16-NEXT: .LBB98_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v166.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v176.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v80.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v165.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v39.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v164.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v167.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v34.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v166.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v165.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v39, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v2.l, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v70.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v33.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v163.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v69.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v39, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v3.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v36.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v162.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v161.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v160.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v150.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v39, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v4.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v160.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v69.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v148.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v39, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v5.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v151.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v149.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v144.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v39, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v6.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v147.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v66.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v36.h, 8, v134.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v39, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v7.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v145.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v135.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v37.h, 8, v133.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v36.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v129.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v39, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v8.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v133.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v39, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v9.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v131.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v129.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v39, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v10.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v39, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v11.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v39, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v12.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v113.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v39, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v13.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v103.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v39, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v14.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v39, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v15.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v97.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v87.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v39, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v16.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v39, v16 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v17.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v150.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v148.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, v39, v17 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v18.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v18.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v146.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, v39, v18 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v19.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v19.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v144.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v134.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, v39, v19 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v20.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v20.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v132.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, v39, v20 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v21.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v21.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, v39, v21 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v22.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, v39, v22 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v23.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v23.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, v39, v23 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v24.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v24.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v112.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, v39, v24 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v25.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v25.h, v25.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v102.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.l, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v27.h, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, v39, v25 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v26.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v26.h, v26.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_and_b16 v28.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, v39, v26 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v27.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.h, v27.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v86.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.l, 0xff, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v29.h, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, v39, v27 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v28.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v28.h, v28.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v34.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.l, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v30.h, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v28, v39, v28 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v29.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v29.h, v29.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v84.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v83.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b16 v31.l, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v31.h, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, v39, v29 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v30.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v30.h, v30.h, v34.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v82.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.l, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v32.h, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, v39, v30 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v31.l, v33.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.h, v31.h, v33.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v81.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v80.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, v39, v31 -; GFX11-TRUE16-NEXT: v_or_b16 v39.l, v32.l, v33.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v32.h, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v39.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v32, v39, v32 +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v118.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v51.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v112.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v35.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v9.h, v34.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v101.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v13.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v98.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v96.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v10.l, v35.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v86.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v16.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v84.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v67.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v164.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v163.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v161.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v64.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v151.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v13.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v149.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v147.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v52.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v145.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v135.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v15.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v16.h, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v132.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v131.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v130.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v17.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v18.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v17.l, v19.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v117.l +; GFX11-TRUE16-NEXT: v_and_b16 v19.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v115.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.h, 0xff, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v113.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_or_b16 v17.h, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.l, v18.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v18.h, v19.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.l, v20.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v19.h, v21.h, v22.l +; GFX11-TRUE16-NEXT: v_and_b16 v20.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v103.l +; GFX11-TRUE16-NEXT: v_and_b16 v21.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v22.l, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v99.l +; GFX11-TRUE16-NEXT: v_and_b16 v23.l, 0xff, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v34.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v97.l +; GFX11-TRUE16-NEXT: v_or_b16 v20.l, v20.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v20.h, v21.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v22.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v21.h, v23.l, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v22.l, v24.l, v24.h +; GFX11-TRUE16-NEXT: v_and_b16 v22.h, 0xff, v30.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v87.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_and_b16 v23.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.l, 8, v85.l +; GFX11-TRUE16-NEXT: v_and_b16 v24.h, 0xff, v31.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v25.h, 0xff, v32.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.l, 8, v83.l +; GFX11-TRUE16-NEXT: v_and_b16 v26.h, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v82.l +; GFX11-TRUE16-NEXT: v_or_b16 v22.h, v22.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.l, v23.h, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v23.h, v24.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v25.h, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v24.h, v26.h, v27.l ; GFX11-TRUE16-NEXT: s_clause 0x5 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:32 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:48 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:64 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:80 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[25:28], off offset:96 -; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[29:32], off offset:112 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:48 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[9:12], off offset:64 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[13:16], off offset:80 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[17:20], off offset:96 +; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[21:24], off offset:112 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v64i16_to_v128i8: diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll index 64b5ecc8f6b8..582f31b0361a 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.128bit.ll @@ -4125,19 +4125,19 @@ define <4 x i32> @bitcast_v16i8_to_v4i32(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -4152,94 +4152,71 @@ define <4 x i32> @bitcast_v16i8_to_v4i32(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v6.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v8.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -8614,19 +8591,19 @@ define <4 x float> @bitcast_v16i8_to_v4f32(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -8641,94 +8618,71 @@ define <4 x float> @bitcast_v16i8_to_v4f32(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v6.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 ; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v8.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -12703,19 +12657,19 @@ define <2 x i64> @bitcast_v16i8_to_v2i64(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -12730,94 +12684,71 @@ define <2 x i64> @bitcast_v16i8_to_v2i64(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v6.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v8.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -16408,19 +16339,19 @@ define <2 x double> @bitcast_v16i8_to_v2f64(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -16435,94 +16366,71 @@ define <2 x double> @bitcast_v16i8_to_v2f64(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v6.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2 ; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v8.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -19833,19 +19741,19 @@ define <8 x i16> @bitcast_v16i8_to_v8i16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -19860,94 +19768,71 @@ define <8 x i16> @bitcast_v16i8_to_v8i16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB98_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v6.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2 ; GFX11-TRUE16-NEXT: .LBB98_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v8.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -22745,19 +22630,19 @@ define <8 x half> @bitcast_v16i8_to_v8f16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -22772,94 +22657,71 @@ define <8 x half> @bitcast_v16i8_to_v8f16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB106_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v6.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB106_2 ; GFX11-TRUE16-NEXT: .LBB106_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v8.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -24960,19 +24822,19 @@ define <8 x bfloat> @bitcast_v16i8_to_v8bf16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v7.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v15.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16 @@ -24987,94 +24849,71 @@ define <8 x bfloat> @bitcast_v16i8_to_v8bf16(<16 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB110_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v1.h, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v6.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v11, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v2.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v11, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.l, v4.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v11.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v11, v3 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB110_2 ; GFX11-TRUE16-NEXT: .LBB110_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v10.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v9.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v12.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v8.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v7.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v9, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v9, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v9.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v8.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v9.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v5.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v9, v3 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll index cb4b3bd4382a..0a73571fe3c1 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.256bit.ll @@ -6298,31 +6298,33 @@ define <8 x i32> @bitcast_v32i8_to_v8i32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v13.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB26_3 @@ -6335,48 +6337,43 @@ define <8 x i32> @bitcast_v32i8_to_v8i32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v0.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v15.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v2.l, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v3.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v4.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v10.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 @@ -6387,122 +6384,88 @@ define <8 x i32> @bitcast_v32i8_to_v8i32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v6.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v7.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v17.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v14.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v13.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v19.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v18.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v15.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v14.h, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v12.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v13.l, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v11.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v9.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v6.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v15.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v15.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v16.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v11.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v12.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v8.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -13349,31 +13312,33 @@ define <8 x float> @bitcast_v32i8_to_v8f32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v13.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB50_3 @@ -13386,48 +13351,43 @@ define <8 x float> @bitcast_v32i8_to_v8f32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v0.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v15.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v2.l, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v3.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v4.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v10.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 @@ -13438,122 +13398,88 @@ define <8 x float> @bitcast_v32i8_to_v8f32(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v6.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v7.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 ; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v17.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v14.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v13.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v19.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v18.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v15.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v14.h, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v12.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v13.l, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v11.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v9.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v6.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v15.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v15.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v16.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v11.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v12.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v8.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -19888,31 +19814,33 @@ define <4 x i64> @bitcast_v32i8_to_v4i64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v13.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB70_3 @@ -19925,48 +19853,43 @@ define <4 x i64> @bitcast_v32i8_to_v4i64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v0.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v15.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v2.l, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v3.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v4.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v10.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 @@ -19977,122 +19900,88 @@ define <4 x i64> @bitcast_v32i8_to_v4i64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v6.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v7.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v17.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v14.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v13.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v19.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v18.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v15.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v14.h, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v12.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v13.l, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v11.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v9.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v6.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v15.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v15.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v16.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v11.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v12.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v8.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -25929,31 +25818,33 @@ define <4 x double> @bitcast_v32i8_to_v4f64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v19.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v17.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v13.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v0.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v23.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v23.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v32 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB86_3 @@ -25966,48 +25857,43 @@ define <4 x double> @bitcast_v32i8_to_v4f64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v0.l, v19.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v19.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v18.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v14.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v1.h, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v15.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v2.l, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v21.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v3.l, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v10.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v4.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v10.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 @@ -26018,122 +25904,88 @@ define <4 x double> @bitcast_v32i8_to_v4f64(<32 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v5.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v6.l, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v21.l, v7.l, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2 ; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v21.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v17.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v16.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v18.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v14.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v13.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v14.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v19.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v18.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v17.h, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v15.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v13.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v12.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v21, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v14.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v14.h, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v15.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v21, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v12.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v13.l, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v21, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v11.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v11.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v21, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v10.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v10.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v21, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v9.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v21, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v8.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.h, v6.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v21, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v21.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v19.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v17.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v17.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v15.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v15.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v16.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v10.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v11.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v11.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v12.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v8.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v8.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v9.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v9.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v21, v7 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll index 3aaf25423a18..b622e6ee58b9 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.320bit.ll @@ -3044,91 +3044,66 @@ define <40 x i8> @bitcast_v10i32_to_v40i8(<10 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v10i32_to_v40i8: @@ -5025,39 +5000,41 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v23.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v19.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v35.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v35.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v33.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v35.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v36 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB14_3 @@ -5071,63 +5048,53 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB14_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v26.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v0.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v25.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v23.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v27, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v1.h, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v15.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v27, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v2.l, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v15.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v27, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v3.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v14.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v27, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v4.l, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v27, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v5.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v27, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v6.l, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v12.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 @@ -5140,147 +5107,110 @@ define <10 x i32> @bitcast_v40i8_to_v10i32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v27, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v7.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v27.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v27, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v8.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v27.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v27, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v9.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v27.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v27, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB14_2 ; GFX11-TRUE16-NEXT: .LBB14_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v26.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v25.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v22.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v25.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v15.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v31.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v24.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v25.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v19.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v25, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v16.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v20.h, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v21.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v25.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v18.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v25, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v15.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v15.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v25, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v14.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v13.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v13.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v25, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v12.h, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v25, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v31.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v25, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.h, v8.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v21.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v22.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v24.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v16.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v17.h, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v18.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v19.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v13.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v13.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v14.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v14.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v10.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v11.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v12.l, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v25, v9 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -9991,91 +9921,66 @@ define <40 x i8> @bitcast_v10f32_to_v40i8(<10 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v10f32_to_v40i8: @@ -11997,39 +11902,41 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v23.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v19.l -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v17.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v35.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v30.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v29.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v28.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v28.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v33.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v33.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v34.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v35.l +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(7) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v33.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v33.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(5) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v34.h +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(4) +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v35.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v36 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB34_3 @@ -12043,63 +11950,53 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB34_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v26.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v0.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v25.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v23.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v27, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v1.h, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v15.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v21.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v27, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v2.l, v20.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v17.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v15.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v27, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v3.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v14.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v27, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v4.l, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v27, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v5.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v27.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v27, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v6.l, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v27.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v23.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v12.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 @@ -12112,147 +12009,110 @@ define <10 x float> @bitcast_v40i8_to_v10f32(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr14_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v27, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v7.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v27.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v27, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v8.l, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v27.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v27, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v27.l, v9.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v27.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v27, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr12_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB34_2 ; GFX11-TRUE16-NEXT: .LBB34_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v26.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v25.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v22.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v25.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v21.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v20.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v15.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v15.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v31.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v24.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v25.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v19.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v19.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v17.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v25, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v16.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v20.h, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v21.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v25, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v25.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v18.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v16.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v18.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v25, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v15.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v15.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v25, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v14.l, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v14.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v25, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v13.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v13.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v25, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v12.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v12.h, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v25, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v31.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v11.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v25, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.h, v8.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v25, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v25.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v21.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v22.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v23.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v23.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v24.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v16.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v17.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v17.h, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v18.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v19.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v12.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v13.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v13.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v14.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v14.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v10.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v11.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v12.l, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v25, v9 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -16367,91 +16227,66 @@ define <40 x i8> @bitcast_v20i16_to_v40i8(<20 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v20i16_to_v40i8: @@ -22484,91 +22319,66 @@ define <40 x i8> @bitcast_v20f16_to_v40i8(<20 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v20f16_to_v40i8: @@ -28791,39 +28601,38 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v27.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v48.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v38.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v36.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v37.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v49 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB72_3 @@ -28837,65 +28646,55 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB72_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v0.l, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v1.h, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v2.l, v28.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v3.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v26.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v22.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v5.l, v21.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v6.l, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v29.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 @@ -28906,146 +28705,110 @@ define <5 x double> @bitcast_v40i8_to_v5f64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v16.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.l, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v9.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB72_2 ; GFX11-TRUE16-NEXT: .LBB72_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v35.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v30.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v23.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v23.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v22.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v21.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v31.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v34.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v34.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v27.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v25.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v24.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v28.h, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v29.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v26.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v21.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v23.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v23.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v22.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v19.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v19.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v18.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v18.h, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v31.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v17.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v16.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v16.h, v8.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v29.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v30.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v34.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v24.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v25.h, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v26.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v27.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v18.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v19.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v19.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v20.h, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v21.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v16.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v16.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v17.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v18.l, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -30878,91 +30641,66 @@ define <40 x i8> @bitcast_v5f64_to_v40i8(<5 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v5f64_to_v40i8: @@ -32912,39 +32650,38 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v27.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v25.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v18.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.l, 8, v5.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v33.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v28.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.l, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v21.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v39.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v34.l, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v48.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v38.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v36.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v37.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v36.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v36.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v38.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v49 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB76_3 @@ -32958,65 +32695,55 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB76_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v35.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v35.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v30.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v0.l, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v34.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v33.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v1.h, v33.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v34.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v28.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v23.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v2.l, v28.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v10.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v3.l, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v26.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v22.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v32.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v32.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v5.l, v21.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v31.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v31.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v6.l, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v29.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v33.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v33.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 @@ -33027,146 +32754,110 @@ define <5 x i64> @bitcast_v40i8_to_v5i64(<40 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v16.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v7.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v8.l, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v9.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB76_2 ; GFX11-TRUE16-NEXT: .LBB76_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v35.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v30.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v29.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v28.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v23.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v23.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v22.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v21.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v20.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v22.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v31.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v34.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v34.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v27.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v27.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v25.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v10, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v24.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v28.h, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v29.l, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v26.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v21.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v20.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v10, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v20.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v22.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v23.l, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v23.h, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v10, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v21.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v22.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v10, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v19.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v19.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v10, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v18.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v18.h, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v10, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v31.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v17.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v10, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v16.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v16.h, v8.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v10, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v29.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v30.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v33.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v33.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v34.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v24.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v25.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v25.h, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v26.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v27.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v18.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v19.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v19.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v20.h, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v21.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v16.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v16.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v17.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v17.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v18.l, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v10, v9 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -35022,91 +34713,66 @@ define <40 x i8> @bitcast_v5i64_to_v40i8(<5 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v16.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v1.l, v11.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v15, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v2.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v28.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v12.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v15, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v3.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v26.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v15, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v4.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v15, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v5.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v15.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v24.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v23.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v22.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v22.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v12.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v28.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v15, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v7.l, v11.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v13.l ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v15, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v8.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v15.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v19.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v11.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v15, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v11.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v9.h, v11.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v15.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v18.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v15, v11 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v15, v9 +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v18.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v13.l ; GFX11-TRUE16-NEXT: s_clause 0x2 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 -; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[11:12], off offset:32 +; GFX11-TRUE16-NEXT: scratch_store_b64 v0, v[9:10], off offset:32 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: bitcast_v5i64_to_v40i8: diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll index 632b03ca51b8..e6c7b1a0f484 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.32bit.ll @@ -2279,17 +2279,13 @@ define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB22_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2 ; GFX11-TRUE16-NEXT: .LBB22_4: ; %cmp.true @@ -2301,13 +2297,9 @@ define i32 @bitcast_v4i8_to_i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4530,17 +4522,13 @@ define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB42_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB42_2 ; GFX11-TRUE16-NEXT: .LBB42_4: ; %cmp.true @@ -4552,13 +4540,9 @@ define float @bitcast_v4i8_to_f32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -6487,17 +6471,13 @@ define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB58_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB58_2 ; GFX11-TRUE16-NEXT: .LBB58_4: ; %cmp.true @@ -6509,13 +6489,9 @@ define <2 x i16> @bitcast_v4i8_to_v2i16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -8138,17 +8114,13 @@ define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true @@ -8160,13 +8132,9 @@ define <2 x half> @bitcast_v4i8_to_v2f16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -9502,17 +9470,13 @@ define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB78_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB78_2 ; GFX11-TRUE16-NEXT: .LBB78_4: ; %cmp.true @@ -9524,13 +9488,9 @@ define <2 x bfloat> @bitcast_v4i8_to_v2bf16(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -10212,17 +10172,13 @@ define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB82_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v1.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v2.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1_hi16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v2, v0 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB82_2 ; GFX11-TRUE16-NEXT: .LBB82_4: ; %cmp.true @@ -10234,13 +10190,9 @@ define <1 x i32> @bitcast_v4i8_to_v1i32(<4 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l ; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v1, v0 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll index d3fbba3cf4dd..bff054f872ee 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll @@ -8921,133 +8921,98 @@ define <64 x i8> @bitcast_v16i32_to_v64i8(<16 x i32> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v24, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v18.l ; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v19.l ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -12574,53 +12539,52 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v20.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v80.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v68.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v69.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v80.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v81 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB26_3 @@ -12633,98 +12597,82 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v53.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v52.l -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v0.l, v54.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v55.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v54.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v64, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v1.h, v53.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v50.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v50.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v29.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v51.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v64, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v2.l, v51.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v64.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v64, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v3.l, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v64, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v4.l, v39.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v64, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v5.l, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v64, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v6.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v64, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v7.l, v25.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v64, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v8.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v9.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v64, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v10.l, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v51.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v52.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v48.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v18.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 @@ -12745,226 +12693,170 @@ define <16 x i32> @bitcast_v64i8_to_v16i32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v64, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v11.l, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v64, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v12.l, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v64, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v13.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v64, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v14.l, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v64.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v64, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v15.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v64, v15 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v52.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v53.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v50.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v50.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v49.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v29.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v23.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v35.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v33.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v31.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v55.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v54.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v53.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v52, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v39.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v51.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v51.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v52, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v52.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v50.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v50.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v52, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v24.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v39.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v52, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v27.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v27.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v52, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v52, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v23.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v23.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v52, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v52, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v21.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v21.h, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v52, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v20.h, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v52, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v19.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v19.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v18.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v31.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v17.h, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v52, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v14.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v52, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v54.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v51.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v51.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v52.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v52.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v53.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v30.h, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v39.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v48.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v24.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v25.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v26.h, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v27.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v21.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v21.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v22.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v23.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v18.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v19.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v20.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v20.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v16.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.l, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v18.l, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v15.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v52, v15 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -23576,133 +23468,98 @@ define <64 x i8> @bitcast_v16f32_to_v64i8(<16 x float> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v24, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v18.l ; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v19.l ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -27358,53 +27215,52 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v20.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v80.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v68.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v69.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v80.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v81 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB50_3 @@ -27417,98 +27273,82 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v53.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v52.l -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v0.l, v54.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v55.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v54.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v64, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v1.h, v53.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v50.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v50.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v29.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v51.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v64, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v2.l, v51.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v64.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v64, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v3.l, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v64, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v4.l, v39.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v64, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v5.l, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v64, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v6.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v64, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v7.l, v25.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v64, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v8.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v9.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v64, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v10.l, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v51.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v52.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v48.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v18.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 @@ -27529,226 +27369,170 @@ define <16 x float> @bitcast_v64i8_to_v16f32(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v64, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v11.l, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v64, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v12.l, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v64, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v13.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v64, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v14.l, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v64.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v64, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v15.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v64, v15 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 ; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v52.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v53.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v50.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v50.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v49.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v29.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v23.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v35.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v33.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v31.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v55.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v54.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v53.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v52, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v39.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v51.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v51.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v52, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v52.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v50.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v50.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v52, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v24.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v39.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v52, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v27.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v27.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v52, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v52, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v23.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v23.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v52, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v52, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v21.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v21.h, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v52, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v20.h, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v52, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v19.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v19.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v18.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v31.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v17.h, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v52, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v14.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v52, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v54.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v51.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v51.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v52.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v52.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v53.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v30.h, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v39.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v48.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v24.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v25.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v26.h, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v27.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v21.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v21.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v22.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v23.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v18.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v19.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v20.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v20.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v16.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.l, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v18.l, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v15.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v52, v15 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -37760,133 +37544,98 @@ define <64 x i8> @bitcast_v8i64_to_v64i8(<8 x i64> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v24, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v18.l ; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v19.l ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -41418,53 +41167,52 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v20.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v80.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v68.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v69.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v80.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v81 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB70_3 @@ -41477,98 +41225,82 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v53.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v52.l -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v0.l, v54.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v55.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v54.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v64, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v1.h, v53.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v50.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v50.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v29.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v51.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v64, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v2.l, v51.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v64.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v64, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v3.l, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v64, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v4.l, v39.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v64, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v5.l, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v64, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v6.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v64, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v7.l, v25.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v64, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v8.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v9.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v64, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v10.l, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v51.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v52.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v48.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v18.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 @@ -41589,226 +41321,170 @@ define <8 x i64> @bitcast_v64i8_to_v8i64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v64, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v11.l, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v64, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v12.l, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v64, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v13.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v64, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v14.l, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v64.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v64, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v15.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v64, v15 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v52.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v53.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v50.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v50.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v49.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v29.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v23.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v35.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v33.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v31.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v55.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v54.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v53.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v52, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v39.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v51.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v51.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v52, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v52.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v50.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v50.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v52, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v24.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v39.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v52, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v27.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v27.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v52, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v52, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v23.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v23.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v52, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v52, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v21.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v21.h, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v52, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v20.h, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v52, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v19.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v19.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v18.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v31.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v17.h, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v52, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v14.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v52, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v54.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v51.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v51.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v52.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v52.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v53.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v30.h, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v39.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v48.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v24.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v25.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v26.h, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v27.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v21.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v21.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v22.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v23.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v18.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v19.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v20.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v20.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v16.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.l, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v18.l, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v15.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v52, v15 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -50954,133 +50630,98 @@ define <64 x i8> @bitcast_v8f64_to_v64i8(<8 x double> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v24, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v18.l ; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v19.l ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -54638,53 +54279,52 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: scratch_load_d16_hi_b16 v38, off, s32 offset:4 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(11) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v29.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v20.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v22.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v20.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v18.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v14.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v48.h, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v10.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.l, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v49.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.l, v8.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v50.h, v6.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v4.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v55.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v7.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.l, 8, v13.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v50.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v17.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v29.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.h, 8, v27.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v54.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v51.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.l, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v52.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v53.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v30.h, 8, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v39.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v48.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v24.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v80.h ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v64.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v65.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v66.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v25.h, 8, v67.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.h, 8, v68.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v68.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v69.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v70.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v70.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v71.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v80.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v64.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v65.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v66.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v26.h, 8, v67.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v27.l, 8, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v68.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v22.h, 8, v69.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v23.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v70.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v71.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v80.l ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v81 ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB86_3 @@ -54697,98 +54337,82 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v55.h -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v53.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v52.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v52.l -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v0.l, v54.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v55.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.l, v54.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v49.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v64, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v1.h, v53.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v55.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v50.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v50.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v49.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v29.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v28.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v51.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v64, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v2.l, v51.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v64.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v48.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v39.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v64, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v3.l, v50.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v50.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v30.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v24.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v26.l ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v64, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v4.l, v39.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v27.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v30.l ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v38.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v38.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v64, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v5.l, v29.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v25.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v23.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v37.h ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v37.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v64, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v6.l, v27.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v36.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v35.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v64, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v7.l, v25.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v21.h ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v35.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v34.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v34.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v64, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v8.l, v23.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v33.h ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v33.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v64, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v9.l, v22.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v64.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v31.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v64, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v10.l, v21.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v31.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v54.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v51.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v52.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v30.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v39.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v48.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v48.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v24.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v25.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v25.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v26.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v27.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v22.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v16.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v18.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_lo16 @@ -54809,226 +54433,170 @@ define <8 x double> @bitcast_v64i8_to_v8f64(<64 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v64, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v11.l, v20.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v64.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v64, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v12.l, v19.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v64, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v13.l, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v64, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v14.l, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v64.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v64, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v64.l, v15.l, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v64.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr16_lo16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v64, v15 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2 ; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v55.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v53.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v52.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v52.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v55.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v53.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v50.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v50.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, v49.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, v29.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, v27.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, v23.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, v28.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, v30.l, 3 +; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, v37.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, v37.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, v35.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, v35.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, v33.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, v33.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, v31.h, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, v31.l, 3 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v55.l, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v54.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v53.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v49.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v49.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v48.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v52, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v39.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v51.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v51.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v52, v4 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v52.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v50.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v50.h, v2.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v29.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v28.h, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v52, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v3.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v3.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, v26.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, v24.h, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v39.h, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v52, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v29.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v30.h, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v24.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v26.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v52, v7 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v5.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v5.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, v28.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, v30.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v27.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v27.h, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v52, v8 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v25.h, v6.h -; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v38.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v38.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v52, v9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v7.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v7.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, v37.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, v37.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v23.l, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v23.h, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v52, v10 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v22.l, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.h, v8.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v36.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v36.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v52, v11 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v9.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v9.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, v35.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, v35.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v21.l, v9.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v21.h, v9.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v52, v12 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v20.l, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v20.h, v10.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v34.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v34.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v52, v13 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v11.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v11.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, v33.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, v33.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v19.l, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v19.h, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v52, v14 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v18.l, v12.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v18.h, v12.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v32.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v32.l, 3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v52, v15 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v13.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v18.h, 0x300, v13.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, v31.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, v31.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.l, v13.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v17.h, v13.h +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v52, v18 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v17.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v16.l, v14.l -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v14.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v52, v17 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v52.l, 0x300, v15.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v54.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v54.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v51.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v51.h, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v52.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v52.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v53.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v30.h, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v39.l, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v39.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v48.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v48.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v24.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v25.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v25.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v26.h, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v27.l, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v21.l, v8.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v21.h, v9.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v22.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v22.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v23.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v18.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v19.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v19.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v20.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v20.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v16.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.h, v14.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v17.l, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v17.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v18.l, v15.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v3.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v3.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x300, v4.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x300, v5.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v5.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.l, 0x300, v6.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v6.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v7.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x300, v7.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.l, 0x300, v8.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v8.h, 0x300, v8.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.l, 0x300, v9.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v9.h, 0x300, v9.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.l, 0x300, v10.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v10.h, 0x300, v10.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.l, 0x300, v11.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v11.h, 0x300, v11.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.l, 0x300, v12.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v12.h, 0x300, v12.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.l, 0x300, v13.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v13.h, 0x300, v13.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.l, 0x300, v14.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v14.h, 0x300, v14.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.l, 0x300, v15.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v15.h, 0x300, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v52.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v52, v15 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -64107,133 +63675,98 @@ define <64 x i8> @bitcast_v32i16_to_v64i8(<32 x i16> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v24, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v18.l ; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v19.l ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -76401,133 +75934,98 @@ define <64 x i8> @bitcast_v32f16_to_v64i8(<32 x half> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v25.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v24.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0 ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v17.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v17.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v64.l ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v55.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v54.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v23.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v54.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v17.h ; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v53.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v52.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v18.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v52.l ; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v51.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v3.l, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v24, v3 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v22.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v22.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.l, v18.h ; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v39.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v17.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v39.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v38.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v37.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v20.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v36.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v36.l ; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v53.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v50.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v8.l, v19.h +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v35.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v19.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v33.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v12.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v32.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v31.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v31.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v21.h +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v6.l, v20.h +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.h, 8, v38.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v21.l, 8, v37.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v20.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.h, v19.h ; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v12.h, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v13.l, v19.l ; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v14.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v29.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v15.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v28.l ; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v17.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v18.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v28.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.h, 8, v27.l ; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v16.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v17.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v26.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v6.h, v21.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v8.h, v20.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v9.l, v21.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v11.l, v20.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v14.l, v19.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v17.h +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.l, v18.l ; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v17.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v26.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v24, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v16.l, v17.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v16.l, v18.h +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v16.h, v19.l ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 @@ -85053,57 +84551,57 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr131_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr130_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr129_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr128_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr119_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr118_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr116_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr25_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr99_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr84_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr117_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr113_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr115_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_lo16 ; GFX11-TRUE16-NEXT: s_and_saveexec_b32 s0, vcc_lo @@ -85111,29 +84609,29 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB108_2 ; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[7:8] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[5:6] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[29:30], 24, v[15:16] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[11:12] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[3:4] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[5:6] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[29:30], 24, v[15:16] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[9:10] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[3:4] ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 24, v16 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v15 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v14 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v15 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v14 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v13 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v12 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v12 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v11 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v10 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v11 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 24, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v10 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v9 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 24, v8 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v8 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v7 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v6 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v7 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v6 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v5 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v4 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v3 @@ -85141,11 +84639,11 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v2 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v1 ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[30:31], 24, v[13:14] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[9:10] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[1:2] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[7:8] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[1:2] ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v1.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.h, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v2.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.h, v2.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v3.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.h, v3.h @@ -85155,26 +84653,26 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.h, v5.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v64.h, v6.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.h, v6.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.h, v7.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.h, v7.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.h, v7.h ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v66.h, v8.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.h, v9.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.h, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.h, v10.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v68.h, v10.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v10.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.h, v11.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.h, v11.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v80.h, v12.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v71.h, v12.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v112.h, v13.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.h, v13.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.h, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.h, v14.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v98.h, v14.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v113.h, v15.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.h, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v103.h, v16.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.h, v16.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v117.h, v15.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.h, v15.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v115.h, v16.l +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v16.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr1 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5 @@ -85187,71 +84685,72 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB108_4 ; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true -; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v4 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_lshlrev_b32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v18 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v20, 0x40c00000, v18 ; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v17, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v20, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v17 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17 ; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v17, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_add3_u32 v17, v24, v20, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v52, v18, v23 :: v_dual_lshlrev_b32 v1, 16, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v39, v18, v23 :: v_dual_and_b32 v2, 0xffff0000, v2 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v1, 16, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v1, 0x40c00000, v1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v20, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_add3_u32 v21, v21, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, 0x400000, v1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_add3_u32 v17, v24, v20, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v21, v21, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 ; GFX11-TRUE16-NEXT: v_add3_u32 v23, v25, v1, 0x7fff ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v21, v22, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v19 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v20 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v39.h +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v52.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v53, v23, v26, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v3 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v19, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_add_f32_e32 v22, 0x40c00000, v20 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v22, 0x40c00000, v20 :: v_dual_add_f32 v3, 0x40c00000, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v4, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v53.h ; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v22, 16, 1 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v18 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v54, v19, v21, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v22 ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v22, 0x7fff -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v18 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v22 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v6 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v1, v20, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v54.h ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v130, 8, v18 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v131, 8, v17 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v3, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v55, v1, v19 :: v_dual_and_b32 v2, 0xffff0000, v6 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v55, v1, v19, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v54.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v4, v21, vcc_lo ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v5 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5 @@ -85304,305 +84803,266 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v10 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v65.h ; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2 -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v66, v4, v5 :: v_dual_lshlrev_b32 v5, 16, v10 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v66, v4, v5, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v7 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v7 :: v_dual_lshlrev_b32 v5, 16, v10 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v21 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v24, v1, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v4, 16, 1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v66.h ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff ; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v115, 24, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v116, 8, v22 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v117, 8, v21 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[21:22] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[19:20] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[17:18] ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v23, v1, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 24, v22 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v22 ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v68, v3, v8 :: v_dual_and_b32 v3, 0xffff0000, v9 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v67, v3, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v12 +; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v12 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v9 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v68, v1, v4 :: v_dual_add_f32 v5, 0x40c00000, v5 ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v6, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v6 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v67, v1, v4, vcc_lo ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v9 -; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v6, 0x7fff ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 +; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v6, 0x7fff +; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v12 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v11 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v66.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v2, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v68.h +; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v67.h ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v3, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v12 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v67.h -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[48:49], 24, v[23:24] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 24, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v24 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 8, v23 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v25, v2, v6, vcc_lo ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v7, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[21:22] -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[50:51], 24, v[19:20] +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v68.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_4) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v1, v6, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v82, v2, v3, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v80, v2, v3, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, 0x400000, v5 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v8 :: v_dual_lshlrev_b32 v5, 16, v14 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v82.h -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[17:18] -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v80, v1, v2 :: v_dual_lshlrev_b32 v1, 16, v11 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v71, v1, v2, vcc_lo +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v11 ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v6, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 24, v26 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1 ; GFX11-TRUE16-NEXT: v_add3_u32 v3, v7, v4, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v4 ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v6, 0x7fff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v14 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 24, v26 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v80.h ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v26 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v3, v7, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v6 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v13 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v80.h -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 24, v24 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v71.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 8, v25 ; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v27, v2, v3 :: v_dual_add_f32 v2, 0x40c00000, v4 ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v5 ; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v1, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v6 ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v6 ; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v28 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v97, v4, v5, vcc_lo -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v96, v4, v5, vcc_lo ; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v2, 0x7fff ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 ; GFX11-TRUE16-NEXT: v_add3_u32 v6, v8, v3, 0x7fff ; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v9 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v28 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v4, v5, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v13 ; GFX11-TRUE16-NEXT: v_add3_u32 v3, v8, v1, 0x7fff -; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v97.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v87, v6, v7, vcc_lo +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v98, v6, v7 :: v_dual_and_b32 v5, 0xffff0000, v16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v9 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v86, 8, v28 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v24 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[27:28] -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v87.h +; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v15 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v15 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v96.h ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v32, v3, v4, vcc_lo ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v16 ; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v2, 16, 1 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[25:26] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 24, v33 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v33 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v6, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v16 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v27 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v114, 8, v23 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v112, v4, v6 :: v_dual_add_f32 v1, 0x40c00000, v5 -; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v8 :: v_dual_lshlrev_b32 v5, 16, v15 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v112.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 -; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v32 -; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.l, v98.h ; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3 -; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v5 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v6, v2, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v2 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[34:35], 24, v[27:28] +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[35:36], 24, v[25:26] ; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v3, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v113, v4, v6 :: v_dual_add_f32 v6, 0x40c00000, v8 ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 -; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v1, 0x7fff +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v6 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v103, v2, v9, vcc_lo -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v115, v2, v9, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v113.h +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[36:37], 24, v[23:24] ; GFX11-TRUE16-NEXT: v_add3_u32 v2, v10, v6, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v113, v7, v11, vcc_lo +; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v33 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v83, 8, v33 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v32 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v27 +; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v5 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v117, v7, v11, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v38, v4, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v39, v4, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v103.h -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v37, v2, v3, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.l, v113.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.l, v115.h +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v38, v2, v3, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.l, v117.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 24, v38 -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v38 -; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[29:30], 24, v[37:38] +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v69, 24, v39 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v70, 8, v39 +; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[29:30], 24, v[38:39] ; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[30:31], 24, v[32:33] -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v37 +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v38 ; GFX11-TRUE16-NEXT: .LBB108_4: ; %end ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v53.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v131.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v17.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v51.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v50.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v18.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v1.l, v1.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v129.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v1.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v55.h ; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v52.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v130.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v19.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v50.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v31, v1 -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v20.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v2.l, v2.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v128.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v19.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v49.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v54.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v119.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v2.h ; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v55.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v128.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v118.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v21.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v31, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v49.l -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v3.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v54.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v119.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v22.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v115.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v31, v3 +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v5.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v6.l, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v20.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v118.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v65.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v116.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v21.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v48.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v64.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v114.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v22.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v112.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v4.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v5.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v7.h, v8.l +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v8.h, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v67.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v103.l ; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v23.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v4.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v65.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v117.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v48.l -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v24.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v31, v4 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v101.l -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v5.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v64.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v116.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v25.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v35.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v31, v5 -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v26.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v6.l, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v7.l, v7.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v68.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v114.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v98.l -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v27.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v31, v6 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v34.l -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v7.l, v7.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v36.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v66.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v24.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v101.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v80.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v100.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v7.l, v7.h ; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v66.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v102.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v9.l, v9.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v10.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v11.l, v11.h +; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v25.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v35.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v68.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.l, 8, v99.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v26.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v97.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.h, 0xff, v96.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v87.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v27.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v34.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v10.h, v11.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.h, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v13.h, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v71.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v86.l ; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v28.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v85.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v31, v7 -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v32.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v8.l, v8.h -; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v82.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v100.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v30.l -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v33.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v31, v8 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v81.l -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v9.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v67.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v99.l -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v37.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v31, v9 -; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v38.h -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v10.l, v10.h -; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v97.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v96.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v69.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v31, v10 -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v11.l, v11.h -; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v12.l, v12.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v80.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v86.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v31, v11 -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v113.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v84.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v32.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v30.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v98.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v83.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v12.h ; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v112.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v84.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v31, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v13.l, v13.h -; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v87.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v83.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v31, v13 -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v14.l, v14.h -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v113.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v71.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v31, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v15.l, v15.h -; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v31.h -; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v103.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v70.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, v31, v15 -; GFX11-TRUE16-NEXT: v_or_b16 v31.l, v16.l, v16.h -; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v17.l, v17.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v31.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v31, v16 +; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v14.l, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v15.l, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v16.l, v16.h +; GFX11-TRUE16-NEXT: v_and_b16 v14.h, 0xff, v33.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v82.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v117.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v81.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v38.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v17.h, 0xff, v115.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v18.l, 8, v70.l +; GFX11-TRUE16-NEXT: v_and_b16 v18.h, 0xff, v39.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v19.l, 8, v69.l +; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v14.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v15.h, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v16.h, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.l, v17.h, v18.l +; GFX11-TRUE16-NEXT: v_or_b16 v16.h, v18.h, v19.l ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[1:4], off ; GFX11-TRUE16-NEXT: scratch_store_b128 v0, v[5:8], off offset:16 diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll index ecc715cfb52f..11f90b906956 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.64bit.ll @@ -3067,9 +3067,9 @@ define i64 @bitcast_v8i8_to_i64(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 @@ -3085,52 +3085,47 @@ define i64 @bitcast_v8i8_to_i64(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB26_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB26_2 ; GFX11-TRUE16-NEXT: .LBB26_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -6210,9 +6205,9 @@ define double @bitcast_v8i8_to_f64(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 @@ -6228,52 +6223,47 @@ define double @bitcast_v8i8_to_f64(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB50_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB50_2 ; GFX11-TRUE16-NEXT: .LBB50_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -9050,9 +9040,9 @@ define <2 x i32> @bitcast_v8i8_to_v2i32(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 @@ -9068,52 +9058,47 @@ define <2 x i32> @bitcast_v8i8_to_v2i32(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB70_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB70_2 ; GFX11-TRUE16-NEXT: .LBB70_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -11590,9 +11575,9 @@ define <2 x float> @bitcast_v8i8_to_v2f32(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 @@ -11608,52 +11593,47 @@ define <2 x float> @bitcast_v8i8_to_v2f32(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB86_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB86_2 ; GFX11-TRUE16-NEXT: .LBB86_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -13809,9 +13789,9 @@ define <4 x i16> @bitcast_v8i8_to_v4i16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 @@ -13827,52 +13807,47 @@ define <4 x i16> @bitcast_v8i8_to_v4i16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB98_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB98_2 ; GFX11-TRUE16-NEXT: .LBB98_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -15630,9 +15605,9 @@ define <4 x half> @bitcast_v8i8_to_v4f16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 @@ -15648,52 +15623,47 @@ define <4 x half> @bitcast_v8i8_to_v4f16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB106_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB106_2 ; GFX11-TRUE16-NEXT: .LBB106_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -16934,9 +16904,9 @@ define <4 x bfloat> @bitcast_v8i8_to_v4bf16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v5.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v5.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v7.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1 @@ -16952,52 +16922,47 @@ define <4 x bfloat> @bitcast_v8i8_to_v4bf16(<8 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB110_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v0.l, v2.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.l, v3.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v5, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v1.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v5, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v2.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.h ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB110_2 ; GFX11-TRUE16-NEXT: .LBB110_4: ; %cmp.true ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.l, 3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v4.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v4.l, 3 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v3.l, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.h, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v2.l, v1.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v4.h -; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v4, v3 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v4, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v2.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v2.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v3.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.h, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll index 685e2fbdecfa..9a6ea1b93dfc 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.96bit.ll @@ -1104,16 +1104,15 @@ define <3 x i32> @bitcast_v12i8_to_v3i32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -1128,37 +1127,28 @@ define <3 x i32> @bitcast_v12i8_to_v3i32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB6_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v4.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v1.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v2.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v7.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB6_2 ; GFX11-TRUE16-NEXT: .LBB6_4: ; %cmp.true @@ -1166,36 +1156,26 @@ define <3 x i32> @bitcast_v12i8_to_v3i32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v4.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.l, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v2 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -4254,16 +4234,15 @@ define <3 x float> @bitcast_v12i8_to_v3f32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v8.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v8.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -4278,37 +4257,28 @@ define <3 x float> @bitcast_v12i8_to_v3f32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB22_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v7.h ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.h -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v0.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v5.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.l, v4.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v6.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v3.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v3.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v4.l +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v1.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v3.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v4 -; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v2.l, v3.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v7.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v2 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2 ; GFX11-TRUE16-NEXT: .LBB22_4: ; %cmp.true @@ -4316,36 +4286,26 @@ define <3 x float> @bitcast_v12i8_to_v3f32(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v7.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v6.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v6.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v4.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v7, v5 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v5.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v4.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v3.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.l, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v3.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v7, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v7.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v7, v2 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -6909,12 +6869,12 @@ define <6 x bfloat> @bitcast_v12i8_to_v6bf16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -6929,37 +6889,28 @@ define <6 x bfloat> @bitcast_v12i8_to_v6bf16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB36_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v6.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v4.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB36_2 ; GFX11-TRUE16-NEXT: .LBB36_4: ; %cmp.true @@ -6967,36 +6918,26 @@ define <6 x bfloat> @bitcast_v12i8_to_v6bf16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v6.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -8669,12 +8610,12 @@ define <6 x half> @bitcast_v12i8_to_v6f16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -8689,37 +8630,28 @@ define <6 x half> @bitcast_v12i8_to_v6f16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB40_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v6.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v4.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB40_2 ; GFX11-TRUE16-NEXT: .LBB40_4: ; %cmp.true @@ -8727,36 +8659,26 @@ define <6 x half> @bitcast_v12i8_to_v6f16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v6.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -10079,12 +10001,12 @@ define <6 x i16> @bitcast_v12i8_to_v6i16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.h, v4.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.h, v2.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v10.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v9.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v10.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v9.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v11.l ; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v12 @@ -10099,37 +10021,28 @@ define <6 x i16> @bitcast_v12i8_to_v6i16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: .LBB44_3: ; %cmp.false ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v8.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v6.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v6.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v1.l, v5.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v7.h +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v6.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v0.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.l, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v2.l, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v5.l ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_hi16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr10_lo16 -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v0 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v1.h, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v4.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v5 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.l, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h -; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr6_lo16 ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_lo16 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4_hi16 +; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5_lo16 ; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0 ; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB44_2 ; GFX11-TRUE16-NEXT: .LBB44_4: ; %cmp.true @@ -10137,36 +10050,26 @@ define <6 x i16> @bitcast_v12i8_to_v6i16(<12 x i8> %a, i32 %b) { ; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v8.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, v7.h, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, v7.l, 3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, v8.l, 3 ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, v10.l, 3 +; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v0.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v5.l, v1.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v6.h, 0x300, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v5.h, v1.h ; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v6 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v6.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v5.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v6.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v4.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v5.l, v2.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, 0x300, v0.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, 0x300, v0.h +; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.l, 0x300, v1.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v1.h, 0x300, v1.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v4.l, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v2.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v3, v1 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v3.l, 0x300, v2.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.l, 0x300, v2.l ; GFX11-TRUE16-NEXT: v_add_nc_u16 v2.h, 0x300, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v3.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2 ; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-break-large-phis-heuristics.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-break-large-phis-heuristics.ll index c94b33334646..1f36902762f0 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-break-large-phis-heuristics.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-break-large-phis-heuristics.ll @@ -726,16 +726,16 @@ define amdgpu_kernel void @used_by_unbreakable_and_breakable_phi(<5 x double> %i ; CHECK-NEXT: [[LARGEPHI_EXTRACTSLICE815:%.*]] = extractelement <5 x double> [[LARGEPHI_INSERTSLICE4]], i64 4 ; CHECK-NEXT: br label [[END]] ; CHECK: end: -; CHECK-NEXT: [[TMP5:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE01]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE1]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP6:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE22]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE3]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP7:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE43]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE5]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP8:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE64]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE7]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP9:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE85]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE9]], [[FINALLY]] ] -; CHECK-NEXT: [[TMP10:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE011]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] -; CHECK-NEXT: [[TMP11:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE212]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] -; CHECK-NEXT: [[TMP12:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE413]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] -; CHECK-NEXT: [[TMP13:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE614]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] -; CHECK-NEXT: [[TMP14:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE815]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP5:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE01]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP6:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE22]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP7:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE43]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP8:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE64]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP9:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE85]], [[THEN1]] ], [ 0.000000e+00, [[FINALLY]] ] +; CHECK-NEXT: [[TMP10:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE011]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE1]], [[FINALLY]] ] +; CHECK-NEXT: [[TMP11:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE212]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE3]], [[FINALLY]] ] +; CHECK-NEXT: [[TMP12:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE413]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE5]], [[FINALLY]] ] +; CHECK-NEXT: [[TMP13:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE614]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE7]], [[FINALLY]] ] +; CHECK-NEXT: [[TMP14:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE815]], [[THEN1]] ], [ [[LARGEPHI_EXTRACTSLICE9]], [[FINALLY]] ] ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE016:%.*]] = insertelement <5 x double> poison, double [[TMP10]], i64 0 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE117:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE016]], double [[TMP11]], i64 1 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE218:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE117]], double [[TMP12]], i64 2 @@ -746,8 +746,8 @@ define amdgpu_kernel void @used_by_unbreakable_and_breakable_phi(<5 x double> %i ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE28:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE17]], double [[TMP7]], i64 2 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE39:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE28]], double [[TMP8]], i64 3 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE410:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE39]], double [[TMP9]], i64 4 -; CHECK-NEXT: store <5 x double> [[LARGEPHI_INSERTSLICE410]], ptr [[OUT]], align 1 ; CHECK-NEXT: store <5 x double> [[LARGEPHI_INSERTSLICE420]], ptr [[OUT]], align 1 +; CHECK-NEXT: store <5 x double> [[LARGEPHI_INSERTSLICE410]], ptr [[OUT]], align 1 ; CHECK-NEXT: ret void ; entry: @@ -1187,11 +1187,11 @@ define amdgpu_kernel void @test_breakable_chain_5_out_of_7(<5 x double> %in, ptr ; CHECK-NEXT: [[LARGEPHI_EXTRACTSLICE960:%.*]] = extractelement <5 x double> [[IN]], i64 4 ; CHECK-NEXT: br i1 [[COND]], label [[END:%.*]], label [[COND5_END]] ; CHECK: cond5.end: -; CHECK-NEXT: [[TMP25:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE041]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE1]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP26:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE242]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE3]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP27:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE443]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE5]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP28:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE644]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE7]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP29:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE845]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE9]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP25:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE041]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE152]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP26:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE242]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE354]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP27:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE443]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE556]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP28:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE644]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE758]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP29:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE845]], [[COND4_END]] ], [ [[LARGEPHI_EXTRACTSLICE960]], [[COND5_TRUE]] ] ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE046:%.*]] = insertelement <5 x double> poison, double [[TMP25]], i64 0 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE147:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE046]], double [[TMP26]], i64 1 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE248:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE147]], double [[TMP27]], i64 2 @@ -1204,11 +1204,11 @@ define amdgpu_kernel void @test_breakable_chain_5_out_of_7(<5 x double> %in, ptr ; CHECK-NEXT: [[LARGEPHI_EXTRACTSLICE859:%.*]] = extractelement <5 x double> [[LARGEPHI_INSERTSLICE450]], i64 4 ; CHECK-NEXT: br label [[END]] ; CHECK: end: -; CHECK-NEXT: [[TMP30:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE051]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE152]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP31:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE253]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE354]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP32:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE455]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE556]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP33:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE657]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE758]], [[COND5_TRUE]] ] -; CHECK-NEXT: [[TMP34:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE859]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE960]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP30:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE051]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE1]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP31:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE253]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE3]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP32:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE455]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE5]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP33:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE657]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE7]], [[COND5_TRUE]] ] +; CHECK-NEXT: [[TMP34:%.*]] = phi double [ [[LARGEPHI_EXTRACTSLICE859]], [[COND5_END]] ], [ [[LARGEPHI_EXTRACTSLICE9]], [[COND5_TRUE]] ] ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE061:%.*]] = insertelement <5 x double> poison, double [[TMP30]], i64 0 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE162:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE061]], double [[TMP31]], i64 1 ; CHECK-NEXT: [[LARGEPHI_INSERTSLICE263:%.*]] = insertelement <5 x double> [[LARGEPHI_INSERTSLICE162]], double [[TMP32]], i64 2 diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-fdiv.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-fdiv.ll index a4f9ce3e7350..7ff86ac152fe 100644 --- a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-fdiv.ll +++ b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-fdiv.ll @@ -2160,7 +2160,22 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; IEEE-GOODFREXP-NEXT: [[TMP38:%.*]] = insertelement <2 x float> poison, float [[TMP27]], i64 0 ; IEEE-GOODFREXP-NEXT: [[MD_1ULP_UNDEF:%.*]] = insertelement <2 x float> [[TMP38]], float [[TMP37]], i64 1 ; IEEE-GOODFREXP-NEXT: store volatile <2 x float> [[MD_1ULP_UNDEF]], ptr addrspace(1) [[OUT]], align 4 -; IEEE-GOODFREXP-NEXT: [[SQRT_X_3ULP:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]), !fpmath [[META3:![0-9]+]] +; IEEE-GOODFREXP-NEXT: [[TMP56:%.*]] = extractelement <2 x float> [[X]], i64 0 +; IEEE-GOODFREXP-NEXT: [[TMP57:%.*]] = extractelement <2 x float> [[X]], i64 1 +; IEEE-GOODFREXP-NEXT: [[TMP58:%.*]] = fcmp olt float [[TMP56]], 0x3810000000000000 +; IEEE-GOODFREXP-NEXT: [[TMP59:%.*]] = select i1 [[TMP58]], i32 32, i32 0 +; IEEE-GOODFREXP-NEXT: [[TMP60:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP56]], i32 [[TMP59]]) +; IEEE-GOODFREXP-NEXT: [[TMP61:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP60]]) +; IEEE-GOODFREXP-NEXT: [[TMP62:%.*]] = select i1 [[TMP58]], i32 -16, i32 0 +; IEEE-GOODFREXP-NEXT: [[TMP63:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP61]], i32 [[TMP62]]) +; IEEE-GOODFREXP-NEXT: [[TMP64:%.*]] = fcmp olt float [[TMP57]], 0x3810000000000000 +; IEEE-GOODFREXP-NEXT: [[TMP65:%.*]] = select i1 [[TMP64]], i32 32, i32 0 +; IEEE-GOODFREXP-NEXT: [[TMP66:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP57]], i32 [[TMP65]]) +; IEEE-GOODFREXP-NEXT: [[TMP67:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP66]]) +; IEEE-GOODFREXP-NEXT: [[TMP68:%.*]] = select i1 [[TMP64]], i32 -16, i32 0 +; IEEE-GOODFREXP-NEXT: [[TMP69:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP67]], i32 [[TMP68]]) +; IEEE-GOODFREXP-NEXT: [[TMP70:%.*]] = insertelement <2 x float> poison, float [[TMP63]], i64 0 +; IEEE-GOODFREXP-NEXT: [[SQRT_X_3ULP:%.*]] = insertelement <2 x float> [[TMP70]], float [[TMP69]], i64 1 ; IEEE-GOODFREXP-NEXT: [[TMP39:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 0 ; IEEE-GOODFREXP-NEXT: [[TMP40:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 1 ; IEEE-GOODFREXP-NEXT: [[TMP41:%.*]] = extractelement <2 x float> [[X]], i64 0 @@ -2231,7 +2246,22 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; IEEE-BADFREXP-NEXT: [[TMP38:%.*]] = insertelement <2 x float> poison, float [[TMP27]], i64 0 ; IEEE-BADFREXP-NEXT: [[MD_1ULP_UNDEF:%.*]] = insertelement <2 x float> [[TMP38]], float [[TMP37]], i64 1 ; IEEE-BADFREXP-NEXT: store volatile <2 x float> [[MD_1ULP_UNDEF]], ptr addrspace(1) [[OUT]], align 4 -; IEEE-BADFREXP-NEXT: [[SQRT_X_3ULP:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]), !fpmath [[META3:![0-9]+]] +; IEEE-BADFREXP-NEXT: [[TMP56:%.*]] = extractelement <2 x float> [[X]], i64 0 +; IEEE-BADFREXP-NEXT: [[TMP57:%.*]] = extractelement <2 x float> [[X]], i64 1 +; IEEE-BADFREXP-NEXT: [[TMP58:%.*]] = fcmp olt float [[TMP56]], 0x3810000000000000 +; IEEE-BADFREXP-NEXT: [[TMP59:%.*]] = select i1 [[TMP58]], i32 32, i32 0 +; IEEE-BADFREXP-NEXT: [[TMP60:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP56]], i32 [[TMP59]]) +; IEEE-BADFREXP-NEXT: [[TMP61:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP60]]) +; IEEE-BADFREXP-NEXT: [[TMP62:%.*]] = select i1 [[TMP58]], i32 -16, i32 0 +; IEEE-BADFREXP-NEXT: [[TMP63:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP61]], i32 [[TMP62]]) +; IEEE-BADFREXP-NEXT: [[TMP64:%.*]] = fcmp olt float [[TMP57]], 0x3810000000000000 +; IEEE-BADFREXP-NEXT: [[TMP65:%.*]] = select i1 [[TMP64]], i32 32, i32 0 +; IEEE-BADFREXP-NEXT: [[TMP66:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP57]], i32 [[TMP65]]) +; IEEE-BADFREXP-NEXT: [[TMP67:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP66]]) +; IEEE-BADFREXP-NEXT: [[TMP68:%.*]] = select i1 [[TMP64]], i32 -16, i32 0 +; IEEE-BADFREXP-NEXT: [[TMP69:%.*]] = call float @llvm.ldexp.f32.i32(float [[TMP67]], i32 [[TMP68]]) +; IEEE-BADFREXP-NEXT: [[TMP70:%.*]] = insertelement <2 x float> poison, float [[TMP63]], i64 0 +; IEEE-BADFREXP-NEXT: [[SQRT_X_3ULP:%.*]] = insertelement <2 x float> [[TMP70]], float [[TMP69]], i64 1 ; IEEE-BADFREXP-NEXT: [[TMP39:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 0 ; IEEE-BADFREXP-NEXT: [[TMP40:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 1 ; IEEE-BADFREXP-NEXT: [[TMP41:%.*]] = extractelement <2 x float> [[X]], i64 0 @@ -2258,7 +2288,12 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; DAZ-NEXT: [[SQRT_X_NO_MD:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]) ; DAZ-NEXT: [[NO_MD:%.*]] = fdiv contract <2 x float> splat (float 1.000000e+00), [[SQRT_X_NO_MD]] ; DAZ-NEXT: store volatile <2 x float> [[NO_MD]], ptr addrspace(1) [[OUT]], align 4 -; DAZ-NEXT: [[SQRT_MD_1ULP:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]), !fpmath [[META2:![0-9]+]] +; DAZ-NEXT: [[TMP39:%.*]] = extractelement <2 x float> [[X]], i64 0 +; DAZ-NEXT: [[TMP40:%.*]] = extractelement <2 x float> [[X]], i64 1 +; DAZ-NEXT: [[TMP41:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP39]]) +; DAZ-NEXT: [[TMP42:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP40]]) +; DAZ-NEXT: [[TMP43:%.*]] = insertelement <2 x float> poison, float [[TMP41]], i64 0 +; DAZ-NEXT: [[SQRT_MD_1ULP:%.*]] = insertelement <2 x float> [[TMP43]], float [[TMP42]], i64 1 ; DAZ-NEXT: [[TMP1:%.*]] = extractelement <2 x float> [[SQRT_MD_1ULP]], i64 0 ; DAZ-NEXT: [[TMP2:%.*]] = extractelement <2 x float> [[SQRT_MD_1ULP]], i64 1 ; DAZ-NEXT: [[TMP3:%.*]] = extractelement <2 x float> [[X]], i64 0 @@ -2276,7 +2311,9 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; DAZ-NEXT: [[SQRT_MD_1ULP_UNDEF:%.*]] = insertelement <2 x float> [[TMP12]], float [[TMP11]], i64 1 ; DAZ-NEXT: [[TMP13:%.*]] = extractelement <2 x float> [[SQRT_MD_1ULP_UNDEF]], i64 0 ; DAZ-NEXT: [[TMP14:%.*]] = extractelement <2 x float> [[SQRT_MD_1ULP_UNDEF]], i64 1 -; DAZ-NEXT: [[TMP15:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP13]]) +; DAZ-NEXT: [[TMP44:%.*]] = extractelement <2 x float> [[X]], i64 0 +; DAZ-NEXT: [[TMP45:%.*]] = extractelement <2 x float> [[X]], i64 1 +; DAZ-NEXT: [[TMP15:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[TMP44]]) ; DAZ-NEXT: [[TMP16:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP14]]) ; DAZ-NEXT: [[TMP17:%.*]] = extractvalue { float, i32 } [[TMP16]], 0 ; DAZ-NEXT: [[TMP18:%.*]] = extractvalue { float, i32 } [[TMP16]], 1 @@ -2290,7 +2327,12 @@ define amdgpu_kernel void @rsq_f32_vector_fpmath(ptr addrspace(1) %out, <2 x flo ; DAZ-NEXT: [[TMP26:%.*]] = insertelement <2 x float> poison, float [[TMP15]], i64 0 ; DAZ-NEXT: [[MD_1ULP_UNDEF:%.*]] = insertelement <2 x float> [[TMP26]], float [[TMP25]], i64 1 ; DAZ-NEXT: store volatile <2 x float> [[MD_1ULP_UNDEF]], ptr addrspace(1) [[OUT]], align 4 -; DAZ-NEXT: [[SQRT_X_3ULP:%.*]] = call contract <2 x float> @llvm.sqrt.v2f32(<2 x float> [[X]]), !fpmath [[META3:![0-9]+]] +; DAZ-NEXT: [[TMP34:%.*]] = extractelement <2 x float> [[X]], i64 0 +; DAZ-NEXT: [[TMP35:%.*]] = extractelement <2 x float> [[X]], i64 1 +; DAZ-NEXT: [[TMP36:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP34]]) +; DAZ-NEXT: [[TMP37:%.*]] = call float @llvm.amdgcn.sqrt.f32(float [[TMP35]]) +; DAZ-NEXT: [[TMP38:%.*]] = insertelement <2 x float> poison, float [[TMP36]], i64 0 +; DAZ-NEXT: [[SQRT_X_3ULP:%.*]] = insertelement <2 x float> [[TMP38]], float [[TMP37]], i64 1 ; DAZ-NEXT: [[TMP27:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 0 ; DAZ-NEXT: [[TMP28:%.*]] = extractelement <2 x float> [[SQRT_X_3ULP]], i64 1 ; DAZ-NEXT: [[TMP29:%.*]] = extractelement <2 x float> [[X]], i64 0 @@ -3200,9 +3242,13 @@ define <4 x float> @rsq_f32_vector_mixed_constant_numerator(<4 x float> %arg) { ; DAZ-NEXT: [[TMP13:%.*]] = extractelement <4 x float> [[DENOM]], i64 1 ; DAZ-NEXT: [[TMP14:%.*]] = extractelement <4 x float> [[DENOM]], i64 2 ; DAZ-NEXT: [[TMP15:%.*]] = extractelement <4 x float> [[DENOM]], i64 3 -; DAZ-NEXT: [[TMP16:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP12]]) -; DAZ-NEXT: [[TMP17:%.*]] = fneg contract float [[TMP13]] -; DAZ-NEXT: [[TMP18:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP17]]) +; DAZ-NEXT: [[TMP42:%.*]] = extractelement <4 x float> [[ARG]], i64 0 +; DAZ-NEXT: [[TMP17:%.*]] = extractelement <4 x float> [[ARG]], i64 1 +; DAZ-NEXT: [[TMP43:%.*]] = extractelement <4 x float> [[ARG]], i64 2 +; DAZ-NEXT: [[TMP44:%.*]] = extractelement <4 x float> [[ARG]], i64 3 +; DAZ-NEXT: [[TMP16:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[TMP42]]) +; DAZ-NEXT: [[TMP45:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float [[TMP17]]) +; DAZ-NEXT: [[TMP18:%.*]] = fneg contract float [[TMP45]] ; DAZ-NEXT: [[TMP19:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP14]]) ; DAZ-NEXT: [[TMP20:%.*]] = extractvalue { float, i32 } [[TMP19]], 0 ; DAZ-NEXT: [[TMP21:%.*]] = extractvalue { float, i32 } [[TMP19]], 1 @@ -3675,9 +3721,13 @@ define <4 x float> @rsq_f32_vector_mixed_constant_numerator_arcp(<4 x float> %ar ; DAZ-NEXT: [[TMP13:%.*]] = extractelement <4 x float> [[DENOM]], i64 1 ; DAZ-NEXT: [[TMP14:%.*]] = extractelement <4 x float> [[DENOM]], i64 2 ; DAZ-NEXT: [[TMP15:%.*]] = extractelement <4 x float> [[DENOM]], i64 3 -; DAZ-NEXT: [[TMP16:%.*]] = call arcp contract float @llvm.amdgcn.rcp.f32(float [[TMP12]]) -; DAZ-NEXT: [[TMP17:%.*]] = fneg arcp contract float [[TMP13]] -; DAZ-NEXT: [[TMP18:%.*]] = call arcp contract float @llvm.amdgcn.rcp.f32(float [[TMP17]]) +; DAZ-NEXT: [[TMP26:%.*]] = extractelement <4 x float> [[ARG]], i64 0 +; DAZ-NEXT: [[TMP17:%.*]] = extractelement <4 x float> [[ARG]], i64 1 +; DAZ-NEXT: [[TMP27:%.*]] = extractelement <4 x float> [[ARG]], i64 2 +; DAZ-NEXT: [[TMP28:%.*]] = extractelement <4 x float> [[ARG]], i64 3 +; DAZ-NEXT: [[TMP16:%.*]] = call arcp contract float @llvm.amdgcn.rsq.f32(float [[TMP26]]) +; DAZ-NEXT: [[TMP29:%.*]] = call arcp contract float @llvm.amdgcn.rsq.f32(float [[TMP17]]) +; DAZ-NEXT: [[TMP18:%.*]] = fneg arcp contract float [[TMP29]] ; DAZ-NEXT: [[TMP19:%.*]] = call arcp contract float @llvm.amdgcn.rcp.f32(float [[TMP14]]) ; DAZ-NEXT: [[TMP20:%.*]] = fmul arcp contract float 4.000000e+00, [[TMP19]] ; DAZ-NEXT: [[TMP21:%.*]] = call arcp contract float @llvm.amdgcn.rcp.f32(float [[TMP15]]) @@ -3850,19 +3900,9 @@ define <4 x float> @rsq_f32_vector_const_denom(ptr addrspace(1) %out, <2 x float ; IEEE-GOODFREXP-NEXT: [[TMP9:%.*]] = extractelement <4 x float> [[SQRT]], i64 1 ; IEEE-GOODFREXP-NEXT: [[TMP10:%.*]] = extractelement <4 x float> [[SQRT]], i64 2 ; IEEE-GOODFREXP-NEXT: [[TMP11:%.*]] = extractelement <4 x float> [[SQRT]], i64 3 -; IEEE-GOODFREXP-NEXT: [[TMP12:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP8]]) -; IEEE-GOODFREXP-NEXT: [[TMP13:%.*]] = extractvalue { float, i32 } [[TMP12]], 0 -; IEEE-GOODFREXP-NEXT: [[TMP14:%.*]] = extractvalue { float, i32 } [[TMP12]], 1 -; IEEE-GOODFREXP-NEXT: [[TMP15:%.*]] = sub i32 0, [[TMP14]] -; IEEE-GOODFREXP-NEXT: [[TMP16:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP13]]) -; IEEE-GOODFREXP-NEXT: [[TMP17:%.*]] = call contract float @llvm.ldexp.f32.i32(float [[TMP16]], i32 [[TMP15]]) -; IEEE-GOODFREXP-NEXT: [[TMP18:%.*]] = fneg contract float [[TMP9]] -; IEEE-GOODFREXP-NEXT: [[TMP48:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP18]]) -; IEEE-GOODFREXP-NEXT: [[TMP49:%.*]] = extractvalue { float, i32 } [[TMP48]], 0 -; IEEE-GOODFREXP-NEXT: [[TMP50:%.*]] = extractvalue { float, i32 } [[TMP48]], 1 -; IEEE-GOODFREXP-NEXT: [[TMP22:%.*]] = sub i32 0, [[TMP50]] -; IEEE-GOODFREXP-NEXT: [[TMP51:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP49]]) -; IEEE-GOODFREXP-NEXT: [[TMP24:%.*]] = call contract float @llvm.ldexp.f32.i32(float [[TMP51]], i32 [[TMP22]]) +; IEEE-GOODFREXP-NEXT: [[TMP17:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 4.000000e+00) +; IEEE-GOODFREXP-NEXT: [[TMP13:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 2.000000e+00) +; IEEE-GOODFREXP-NEXT: [[TMP24:%.*]] = fneg contract float [[TMP13]] ; IEEE-GOODFREXP-NEXT: [[TMP29:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP10]]) ; IEEE-GOODFREXP-NEXT: [[TMP30:%.*]] = extractvalue { float, i32 } [[TMP29]], 0 ; IEEE-GOODFREXP-NEXT: [[TMP31:%.*]] = extractvalue { float, i32 } [[TMP29]], 1 @@ -3903,19 +3943,9 @@ define <4 x float> @rsq_f32_vector_const_denom(ptr addrspace(1) %out, <2 x float ; IEEE-BADFREXP-NEXT: [[TMP9:%.*]] = extractelement <4 x float> [[SQRT]], i64 1 ; IEEE-BADFREXP-NEXT: [[TMP10:%.*]] = extractelement <4 x float> [[SQRT]], i64 2 ; IEEE-BADFREXP-NEXT: [[TMP11:%.*]] = extractelement <4 x float> [[SQRT]], i64 3 -; IEEE-BADFREXP-NEXT: [[TMP12:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP8]]) -; IEEE-BADFREXP-NEXT: [[TMP13:%.*]] = extractvalue { float, i32 } [[TMP12]], 0 -; IEEE-BADFREXP-NEXT: [[TMP14:%.*]] = call i32 @llvm.amdgcn.frexp.exp.i32.f32(float [[TMP8]]) -; IEEE-BADFREXP-NEXT: [[TMP15:%.*]] = sub i32 0, [[TMP14]] -; IEEE-BADFREXP-NEXT: [[TMP16:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP13]]) -; IEEE-BADFREXP-NEXT: [[TMP17:%.*]] = call contract float @llvm.ldexp.f32.i32(float [[TMP16]], i32 [[TMP15]]) -; IEEE-BADFREXP-NEXT: [[TMP18:%.*]] = fneg contract float [[TMP9]] -; IEEE-BADFREXP-NEXT: [[TMP48:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP18]]) -; IEEE-BADFREXP-NEXT: [[TMP49:%.*]] = extractvalue { float, i32 } [[TMP48]], 0 -; IEEE-BADFREXP-NEXT: [[TMP21:%.*]] = call i32 @llvm.amdgcn.frexp.exp.i32.f32(float [[TMP18]]) -; IEEE-BADFREXP-NEXT: [[TMP22:%.*]] = sub i32 0, [[TMP21]] -; IEEE-BADFREXP-NEXT: [[TMP50:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP49]]) -; IEEE-BADFREXP-NEXT: [[TMP24:%.*]] = call contract float @llvm.ldexp.f32.i32(float [[TMP50]], i32 [[TMP22]]) +; IEEE-BADFREXP-NEXT: [[TMP17:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 4.000000e+00) +; IEEE-BADFREXP-NEXT: [[TMP13:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 2.000000e+00) +; IEEE-BADFREXP-NEXT: [[TMP24:%.*]] = fneg contract float [[TMP13]] ; IEEE-BADFREXP-NEXT: [[TMP29:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP10]]) ; IEEE-BADFREXP-NEXT: [[TMP30:%.*]] = extractvalue { float, i32 } [[TMP29]], 0 ; IEEE-BADFREXP-NEXT: [[TMP27:%.*]] = call i32 @llvm.amdgcn.frexp.exp.i32.f32(float [[TMP10]]) @@ -3956,9 +3986,9 @@ define <4 x float> @rsq_f32_vector_const_denom(ptr addrspace(1) %out, <2 x float ; DAZ-NEXT: [[TMP9:%.*]] = extractelement <4 x float> [[SQRT]], i64 1 ; DAZ-NEXT: [[TMP10:%.*]] = extractelement <4 x float> [[SQRT]], i64 2 ; DAZ-NEXT: [[TMP11:%.*]] = extractelement <4 x float> [[SQRT]], i64 3 -; DAZ-NEXT: [[TMP12:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP8]]) -; DAZ-NEXT: [[TMP13:%.*]] = fneg contract float [[TMP9]] -; DAZ-NEXT: [[TMP14:%.*]] = call contract float @llvm.amdgcn.rcp.f32(float [[TMP13]]) +; DAZ-NEXT: [[TMP12:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 4.000000e+00) +; DAZ-NEXT: [[TMP13:%.*]] = call contract float @llvm.amdgcn.rsq.f32(float 2.000000e+00) +; DAZ-NEXT: [[TMP14:%.*]] = fneg contract float [[TMP13]] ; DAZ-NEXT: [[TMP15:%.*]] = call { float, i32 } @llvm.frexp.f32.i32(float [[TMP10]]) ; DAZ-NEXT: [[TMP16:%.*]] = extractvalue { float, i32 } [[TMP15]], 0 ; DAZ-NEXT: [[TMP17:%.*]] = extractvalue { float, i32 } [[TMP15]], 1 diff --git a/llvm/test/CodeGen/AMDGPU/bf16.ll b/llvm/test/CodeGen/AMDGPU/bf16.ll index 4b14dc63eeb8..7ee0015f1599 100644 --- a/llvm/test/CodeGen/AMDGPU/bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/bf16.ll @@ -21204,18 +21204,14 @@ define bfloat @v_fabs_bf16(bfloat %a) { ; GCN: ; %bb.0: ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0 -; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GCN-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0 -; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 +; GCN-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0 ; GCN-NEXT: s_setpc_b64 s[30:31] ; ; GFX7-LABEL: v_fabs_bf16: ; GFX7: ; %bb.0: ; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0 -; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX7-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0 -; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 +; GFX7-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0 ; GFX7-NEXT: s_setpc_b64 s[30:31] ; ; GFX8-LABEL: v_fabs_bf16: @@ -21440,10 +21436,7 @@ define bfloat @v_fneg_fabs_bf16(bfloat %a) { ; GCN-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GCN-NEXT: v_mul_f32_e32 v0, 1.0, v0 ; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GCN-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0 -; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GCN-NEXT: v_xor_b32_e32 v0, 0x80000000, v0 -; GCN-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 +; GCN-NEXT: v_or_b32_e32 v0, 0x80000000, v0 ; GCN-NEXT: s_setpc_b64 s[30:31] ; ; GFX7-LABEL: v_fneg_fabs_bf16: @@ -21451,10 +21444,7 @@ define bfloat @v_fneg_fabs_bf16(bfloat %a) { ; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX7-NEXT: v_mul_f32_e32 v0, 1.0, v0 ; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX7-NEXT: v_and_b32_e32 v0, 0x7fffffff, v0 -; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GFX7-NEXT: v_xor_b32_e32 v0, 0x80000000, v0 -; GFX7-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 +; GFX7-NEXT: v_or_b32_e32 v0, 0x80000000, v0 ; GFX7-NEXT: s_setpc_b64 s[30:31] ; ; GFX8-LABEL: v_fneg_fabs_bf16: @@ -21510,23 +21500,17 @@ define amdgpu_ps i32 @s_fneg_fabs_bf16(bfloat inreg %a) { ; GCN-LABEL: s_fneg_fabs_bf16: ; GCN: ; %bb.0: ; GCN-NEXT: v_mul_f32_e64 v0, 1.0, s0 +; GCN-NEXT: v_lshrrev_b32_e32 v0, 16, v0 +; GCN-NEXT: v_or_b32_e32 v0, 0x8000, v0 ; GCN-NEXT: v_readfirstlane_b32 s0, v0 -; GCN-NEXT: s_and_b32 s0, s0, 0xffff0000 -; GCN-NEXT: s_bitset0_b32 s0, 31 -; GCN-NEXT: s_and_b32 s0, s0, 0xffff0000 -; GCN-NEXT: s_xor_b32 s0, s0, 0x80000000 -; GCN-NEXT: s_lshr_b32 s0, s0, 16 ; GCN-NEXT: ; return to shader part epilog ; ; GFX7-LABEL: s_fneg_fabs_bf16: ; GFX7: ; %bb.0: ; GFX7-NEXT: v_mul_f32_e64 v0, 1.0, s0 +; GFX7-NEXT: v_lshrrev_b32_e32 v0, 16, v0 +; GFX7-NEXT: v_or_b32_e32 v0, 0x8000, v0 ; GFX7-NEXT: v_readfirstlane_b32 s0, v0 -; GFX7-NEXT: s_and_b32 s0, s0, 0xffff0000 -; GFX7-NEXT: s_bitset0_b32 s0, 31 -; GFX7-NEXT: s_and_b32 s0, s0, 0xffff0000 -; GFX7-NEXT: s_xor_b32 s0, s0, 0x80000000 -; GFX7-NEXT: s_lshr_b32 s0, s0, 16 ; GFX7-NEXT: ; return to shader part epilog ; ; GFX8-LABEL: s_fneg_fabs_bf16: diff --git a/llvm/test/CodeGen/AMDGPU/calling-conventions.ll b/llvm/test/CodeGen/AMDGPU/calling-conventions.ll index cbf6b66c6acd..7dbbeaabeb71 100644 --- a/llvm/test/CodeGen/AMDGPU/calling-conventions.ll +++ b/llvm/test/CodeGen/AMDGPU/calling-conventions.ll @@ -3632,13 +3632,9 @@ define amdgpu_cs void @amdgpu_cs_v32i1(<32 x i1> %arg0) { ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v2.l, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 ; GFX11-TRUE16-NEXT: global_store_b32 v[0:1], v0, off ; GFX11-TRUE16-NEXT: s_endpgm ; @@ -3813,16 +3809,12 @@ define amdgpu_cs void @amdgpu_cs_v32i1(<32 x i1> %arg0) { ; GFX1250-TRUE16-NEXT: v_bitop3_b16 v2.l, v16.l, v16.h, 15 bitop3:0xec ; GFX1250-TRUE16-NEXT: v_or_b16 v2.h, v2.h, v17.l ; GFX1250-TRUE16-NEXT: v_bitop3_b16 v0.l, v0.l, v0.h, 15 bitop3:0xec -; GFX1250-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4) +; GFX1250-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX1250-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v1.l -; GFX1250-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 ; GFX1250-TRUE16-NEXT: v_bitop3_b16 v1.h, v2.l, v2.h, 0xff bitop3:0xec -; GFX1250-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX1250-TRUE16-NEXT: v_bitop3_b16 v0.l, v0.l, v0.h, 0xff bitop3:0xec -; GFX1250-TRUE16-NEXT: v_mov_b16_e32 v0.h, v1.l -; GFX1250-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX1250-TRUE16-NEXT: v_or_b32_e32 v0, v0, v1 -; GFX1250-TRUE16-NEXT: global_store_b32 v[0:1], v0, off +; GFX1250-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) +; GFX1250-TRUE16-NEXT: v_bitop3_b16 v1.l, v0.l, v0.h, 0xff bitop3:0xec +; GFX1250-TRUE16-NEXT: global_store_b32 v[0:1], v1, off ; GFX1250-TRUE16-NEXT: s_endpgm ; ; GFX1250-FAKE16-LABEL: amdgpu_cs_v32i1: diff --git a/llvm/test/CodeGen/AMDGPU/coalescer-avoid-coalesce-class-with-no-registers.ll b/llvm/test/CodeGen/AMDGPU/coalescer-avoid-coalesce-class-with-no-registers.ll new file mode 100644 index 000000000000..f4665137393b --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/coalescer-avoid-coalesce-class-with-no-registers.ll @@ -0,0 +1,27 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck %s + +; Make sure the coalescer doesn't introduce any uses of +; vreg_1024. None are available to allocate with the register budget +; of this function. + +define void @no_introduce_vreg_1024() #0 { +; CHECK-LABEL: no_introduce_vreg_1024: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; def v[0:7] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: v_mov_b32_e32 v9, v0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use v[0:15] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_setpc_b64 s[30:31] + %tuple = call <8 x i32> asm sideeffect "; def $0","=v"() + %sub0 = extractelement <8 x i32> %tuple, i32 0 + %insert = insertelement <16 x i32> poison, i32 %sub0, i32 9 + call void asm sideeffect "; use $0","v"(<16 x i32> %insert) + ret void +} + +attributes #0 = { nounwind "amdgpu-waves-per-eu"="10,10" } diff --git a/llvm/test/CodeGen/AMDGPU/coalescer-avoid-coalesce-class-with-no-registers.mir b/llvm/test/CodeGen/AMDGPU/coalescer-avoid-coalesce-class-with-no-registers.mir new file mode 100644 index 000000000000..1f414eb2d786 --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/coalescer-avoid-coalesce-class-with-no-registers.mir @@ -0,0 +1,34 @@ +# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6 +# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -run-pass=register-coalescer -o - %s | FileCheck %s + +# The register budget for this function does not permit using 1024-bit +# registers. The coalescer should not introduce a 1024-bit virtual +# register which will fail to allocate. + +--- | + define void @no_introduce_vreg_1024() #0 { + ret void + } + + attributes #0 = { "amdgpu-waves-per-eu"="10,10" } +... +--- +name: no_introduce_vreg_1024 +tracksRegLiveness: true +machineFunctionInfo: + occupancy: 10 +body: | + bb.0: + liveins: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 + + ; CHECK-LABEL: name: no_introduce_vreg_1024 + ; CHECK: liveins: $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 + ; CHECK-NEXT: {{ $}} + ; CHECK-NEXT: [[COPY:%[0-9]+]]:vreg_256 = COPY $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 + ; CHECK-NEXT: undef [[COPY1:%[0-9]+]].sub9:vreg_512 = COPY [[COPY]].sub0 + ; CHECK-NEXT: SI_RETURN implicit [[COPY1]] + %0:vreg_256 = COPY $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 + undef %1.sub9:vreg_512 = COPY %0.sub0 + SI_RETURN implicit %1 + +... diff --git a/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll b/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll index 26f204f29f5a..14897b68bf57 100644 --- a/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll +++ b/llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll @@ -1771,33 +1771,29 @@ define amdgpu_kernel void @load_v4i8_to_v4f32_2_uses(ptr addrspace(1) noalias %o ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_load_b64 s[0:1], s[4:5], 0x34 ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 2, v0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.h ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: global_load_b32 v4, v0, s[0:1] +; GFX11-TRUE16-NEXT: global_load_b32 v5, v0, s[0:1] ; GFX11-TRUE16-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v4.l, 9 -; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v4.h, 9 -; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff00, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff00, v4.h -; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte3_e32 v3, v4 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.l, v5.l, 9 +; GFX11-TRUE16-NEXT: v_add_nc_u16 v0.h, v5.h, 9 +; GFX11-TRUE16-NEXT: v_and_b16 v1.l, 0xff00, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff00, v5.h +; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte3_e32 v3, v5 ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v0.h -; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte2_e32 v2, v4 +; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte2_e32 v2, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v1.l, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte1_e32 v1, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_add_nc_u16 v5.l, 0x900, v0.l -; GFX11-TRUE16-NEXT: v_add_nc_u16 v7.h, 0x900, v0.h -; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte0_e32 v0, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v5, v7 +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v1.l, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte1_e32 v1, v5 +; GFX11-TRUE16-NEXT: v_cvt_f32_ubyte0_e32 v0, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.l, 0x900, v4.l +; GFX11-TRUE16-NEXT: v_add_nc_u16 v4.h, 0x900, v4.h ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: global_store_b128 v6, v[0:3], s[0:1] diff --git a/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll b/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll index 5d184b17f32e..c46fcde739b1 100644 --- a/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/fabs.bf16.ll @@ -218,19 +218,11 @@ define amdgpu_kernel void @s_fabs_v4bf16(ptr addrspace(1) %out, <4 x bfloat> %in ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_and_b32 s4, s3, 0xffff0000 -; CI-NEXT: s_lshl_b32 s3, s3, 16 -; CI-NEXT: s_and_b32 s5, s2, 0xffff0000 -; CI-NEXT: v_mul_f32_e64 v0, 1.0, |s4| -; CI-NEXT: v_mul_f32_e64 v1, 1.0, |s3| -; CI-NEXT: v_mul_f32_e64 v2, 1.0, |s5| -; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_alignbit_b32 v1, v0, v1, 16 -; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v2 -; CI-NEXT: v_mul_f32_e64 v2, 1.0, |s2| -; CI-NEXT: v_alignbit_b32 v0, v0, v2, 16 +; CI-NEXT: s_and_b32 s3, s3, 0x7fff7fff +; CI-NEXT: s_and_b32 s2, s2, 0x7fff7fff ; CI-NEXT: v_mov_b32_e32 v3, s1 +; CI-NEXT: v_mov_b32_e32 v0, s2 +; CI-NEXT: v_mov_b32_e32 v1, s3 ; CI-NEXT: v_mov_b32_e32 v2, s0 ; CI-NEXT: flat_store_dwordx2 v[2:3], v[0:1] ; CI-NEXT: s_endpgm @@ -537,16 +529,15 @@ define amdgpu_kernel void @v_fabs_fold_self_v2bf16(ptr addrspace(1) %out, ptr ad ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 ; CI-NEXT: s_waitcnt vmcnt(0) -; CI-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 -; CI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; CI-NEXT: v_mul_f32_e64 v4, 1.0, |v3| -; CI-NEXT: v_mul_f32_e64 v5, 1.0, |v2| -; CI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 -; CI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 -; CI-NEXT: v_mul_f32_e32 v3, v4, v3 -; CI-NEXT: v_mul_f32_e32 v2, v5, v2 -; CI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 -; CI-NEXT: v_alignbit_b32 v2, v3, v2, 16 +; CI-NEXT: v_and_b32_e32 v3, 0x7fff, v2 +; CI-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; CI-NEXT: v_and_b32_e32 v5, 0xffff0000, v2 +; CI-NEXT: v_and_b32_e32 v2, 0x7fff0000, v2 +; CI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; CI-NEXT: v_mul_f32_e32 v2, v2, v5 +; CI-NEXT: v_mul_f32_e32 v3, v3, v4 +; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 +; CI-NEXT: v_alignbit_b32 v2, v2, v3, 16 ; CI-NEXT: flat_store_dword v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -898,16 +889,13 @@ define amdgpu_kernel void @v_extract_fabs_fold_v2bf16(ptr addrspace(1) %in) #0 { ; CI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc ; CI-NEXT: flat_load_dword v0, v[0:1] ; CI-NEXT: s_waitcnt vmcnt(0) -; CI-NEXT: v_lshlrev_b32_e32 v1, 16, v0 -; CI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; CI-NEXT: v_mul_f32_e64 v1, 1.0, |v1| -; CI-NEXT: v_mul_f32_e64 v0, 1.0, |v0| -; CI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 -; CI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; CI-NEXT: v_mul_f32_e32 v1, 4.0, v1 +; CI-NEXT: v_and_b32_e32 v1, 0x7fff, v0 +; CI-NEXT: v_and_b32_e32 v0, 0x7fff0000, v0 +; CI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; CI-NEXT: v_add_f32_e32 v0, 2.0, v0 -; CI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; CI-NEXT: v_mul_f32_e32 v1, 4.0, v1 ; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 +; CI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 ; CI-NEXT: flat_store_short v[0:1], v1 ; CI-NEXT: s_waitcnt vmcnt(0) ; CI-NEXT: flat_store_short v[0:1], v0 diff --git a/llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll b/llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll index 3983655285e5..38239c550931 100644 --- a/llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll +++ b/llvm/test/CodeGen/AMDGPU/fdiv_flags.f32.ll @@ -1634,29 +1634,18 @@ define float @v_recip_sqrt_f32_ulp25_contract(float %x) { ; IR-IEEE-SDAG-LABEL: v_recip_sqrt_f32_ulp25_contract: ; IR-IEEE-SDAG: ; %bb.0: ; IR-IEEE-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; IR-IEEE-SDAG-NEXT: s_mov_b32 s4, 0xf800000 -; IR-IEEE-SDAG-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0 +; IR-IEEE-SDAG-NEXT: s_mov_b32 s4, 0x800000 ; IR-IEEE-SDAG-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc -; IR-IEEE-SDAG-NEXT: v_sqrt_f32_e32 v1, v0 -; IR-IEEE-SDAG-NEXT: v_add_i32_e64 v2, s[4:5], -1, v1 -; IR-IEEE-SDAG-NEXT: v_fma_f32 v3, -v2, v1, v0 -; IR-IEEE-SDAG-NEXT: v_cmp_ge_f32_e64 s[4:5], 0, v3 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e64 v2, v1, v2, s[4:5] -; IR-IEEE-SDAG-NEXT: v_add_i32_e64 v3, s[4:5], 1, v1 -; IR-IEEE-SDAG-NEXT: v_fma_f32 v1, -v3, v1, v0 -; IR-IEEE-SDAG-NEXT: v_cmp_lt_f32_e64 s[4:5], 0, v1 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e64 v1, v2, v3, s[4:5] -; IR-IEEE-SDAG-NEXT: v_mul_f32_e32 v2, 0x37800000, v1 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc -; IR-IEEE-SDAG-NEXT: v_mov_b32_e32 v2, 0x260 -; IR-IEEE-SDAG-NEXT: v_cmp_class_f32_e32 vcc, v0, v2 -; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc +; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, 32, vcc +; IR-IEEE-SDAG-NEXT: v_ldexp_f32_e32 v0, v0, v1 +; IR-IEEE-SDAG-NEXT: v_sqrt_f32_e32 v0, v0 +; IR-IEEE-SDAG-NEXT: v_cndmask_b32_e64 v1, 0, -16, vcc +; IR-IEEE-SDAG-NEXT: v_ldexp_f32_e32 v0, v0, v1 ; IR-IEEE-SDAG-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 ; IR-IEEE-SDAG-NEXT: v_rcp_f32_e32 v2, v1 -; IR-IEEE-SDAG-NEXT: v_fma_f32 v3, -v1, v2, 1.0 -; IR-IEEE-SDAG-NEXT: v_fma_f32 v2, v3, v2, v2 ; IR-IEEE-SDAG-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; IR-IEEE-SDAG-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; IR-IEEE-SDAG-NEXT: v_fma_f32 v2, v4, v2, v2 ; IR-IEEE-SDAG-NEXT: v_mul_f32_e32 v4, v3, v2 ; IR-IEEE-SDAG-NEXT: v_fma_f32 v5, -v1, v4, v3 ; IR-IEEE-SDAG-NEXT: v_fma_f32 v4, v5, v2, v4 @@ -1668,24 +1657,14 @@ define float @v_recip_sqrt_f32_ulp25_contract(float %x) { ; IR-IEEE-GISEL-LABEL: v_recip_sqrt_f32_ulp25_contract: ; IR-IEEE-GISEL: ; %bb.0: ; IR-IEEE-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; IR-IEEE-GISEL-NEXT: v_mov_b32_e32 v1, 0xf800000 -; IR-IEEE-GISEL-NEXT: v_mul_f32_e32 v2, 0x4f800000, v0 +; IR-IEEE-GISEL-NEXT: v_mov_b32_e32 v1, 0x800000 ; IR-IEEE-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc -; IR-IEEE-GISEL-NEXT: v_sqrt_f32_e32 v1, v0 -; IR-IEEE-GISEL-NEXT: v_add_i32_e64 v2, s[4:5], -1, v1 -; IR-IEEE-GISEL-NEXT: v_fma_f32 v3, -v2, v1, v0 -; IR-IEEE-GISEL-NEXT: v_add_i32_e64 v4, s[4:5], 1, v1 -; IR-IEEE-GISEL-NEXT: v_fma_f32 v5, -v4, v1, v0 -; IR-IEEE-GISEL-NEXT: v_cmp_ge_f32_e64 s[4:5], 0, v3 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e64 v1, v1, v2, s[4:5] -; IR-IEEE-GISEL-NEXT: v_cmp_lt_f32_e64 s[4:5], 0, v5 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e64 v1, v1, v4, s[4:5] -; IR-IEEE-GISEL-NEXT: v_mul_f32_e32 v2, 0x37800000, v1 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc -; IR-IEEE-GISEL-NEXT: v_mov_b32_e32 v2, 0x260 -; IR-IEEE-GISEL-NEXT: v_cmp_class_f32_e32 vcc, v0, v2 -; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc +; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, 1, vcc +; IR-IEEE-GISEL-NEXT: v_lshlrev_b32_e32 v1, 5, v1 +; IR-IEEE-GISEL-NEXT: v_ldexp_f32_e32 v0, v0, v1 +; IR-IEEE-GISEL-NEXT: v_sqrt_f32_e32 v0, v0 +; IR-IEEE-GISEL-NEXT: v_cndmask_b32_e64 v1, 0, -16, vcc +; IR-IEEE-GISEL-NEXT: v_ldexp_f32_e32 v0, v0, v1 ; IR-IEEE-GISEL-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 ; IR-IEEE-GISEL-NEXT: v_rcp_f32_e32 v2, v1 ; IR-IEEE-GISEL-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 @@ -1705,75 +1684,24 @@ define float @v_recip_sqrt_f32_ulp25_contract(float %x) { ; CODEGEN-DAZ-NEXT: v_rsq_f32_e32 v0, v0 ; CODEGEN-DAZ-NEXT: s_setpc_b64 s[30:31] ; -; IR-DAZ-SDAG-LABEL: v_recip_sqrt_f32_ulp25_contract: -; IR-DAZ-SDAG: ; %bb.0: -; IR-DAZ-SDAG-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; IR-DAZ-SDAG-NEXT: s_mov_b32 s4, 0xf800000 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v1, 0x4f800000, v0 -; IR-DAZ-SDAG-NEXT: v_cmp_gt_f32_e32 vcc, s4, v0 -; IR-DAZ-SDAG-NEXT: v_cndmask_b32_e32 v0, v0, v1, vcc -; IR-DAZ-SDAG-NEXT: v_rsq_f32_e32 v1, v0 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v2, v0, v1 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v1, 0.5, v1 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v3, -v1, v2, 0.5 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v2, v2, v3, v2 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v4, -v2, v2, v0 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v1, v1, v3, v1 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v1, v4, v1, v2 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v2, 0x37800000, v1 -; IR-DAZ-SDAG-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc -; IR-DAZ-SDAG-NEXT: v_mov_b32_e32 v2, 0x260 -; IR-DAZ-SDAG-NEXT: v_cmp_class_f32_e32 vcc, v0, v2 -; IR-DAZ-SDAG-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc -; IR-DAZ-SDAG-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 -; IR-DAZ-SDAG-NEXT: v_rcp_f32_e32 v2, v1 -; IR-DAZ-SDAG-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 -; IR-DAZ-SDAG-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v4, -v1, v2, 1.0 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v2, v4, v2, v2 -; IR-DAZ-SDAG-NEXT: v_mul_f32_e32 v4, v3, v2 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v5, -v1, v4, v3 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v4, v5, v2, v4 -; IR-DAZ-SDAG-NEXT: v_fma_f32 v1, -v1, v4, v3 -; IR-DAZ-SDAG-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; IR-DAZ-SDAG-NEXT: v_div_fmas_f32 v1, v1, v2, v4 -; IR-DAZ-SDAG-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 -; IR-DAZ-SDAG-NEXT: s_setpc_b64 s[30:31] -; -; IR-DAZ-GISEL-LABEL: v_recip_sqrt_f32_ulp25_contract: -; IR-DAZ-GISEL: ; %bb.0: -; IR-DAZ-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; IR-DAZ-GISEL-NEXT: v_mov_b32_e32 v1, 0xf800000 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v2, 0x4f800000, v0 -; IR-DAZ-GISEL-NEXT: v_cmp_lt_f32_e32 vcc, v0, v1 -; IR-DAZ-GISEL-NEXT: v_cndmask_b32_e32 v0, v0, v2, vcc -; IR-DAZ-GISEL-NEXT: v_rsq_f32_e32 v1, v0 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v2, v0, v1 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v1, 0.5, v1 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v3, -v1, v2, 0.5 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v2, v2, v3, v2 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v1, v1, v3, v1 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v3, -v2, v2, v0 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v1, v3, v1, v2 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v2, 0x37800000, v1 -; IR-DAZ-GISEL-NEXT: v_cndmask_b32_e32 v1, v1, v2, vcc -; IR-DAZ-GISEL-NEXT: v_mov_b32_e32 v2, 0x260 -; IR-DAZ-GISEL-NEXT: v_cmp_class_f32_e32 vcc, v0, v2 -; IR-DAZ-GISEL-NEXT: v_cndmask_b32_e32 v0, v1, v0, vcc -; IR-DAZ-GISEL-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 -; IR-DAZ-GISEL-NEXT: v_rcp_f32_e32 v2, v1 -; IR-DAZ-GISEL-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 -; IR-DAZ-GISEL-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v4, -v1, v2, 1.0 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v2, v4, v2, v2 -; IR-DAZ-GISEL-NEXT: v_mul_f32_e32 v4, v3, v2 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v5, -v1, v4, v3 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v4, v5, v2, v4 -; IR-DAZ-GISEL-NEXT: v_fma_f32 v1, -v1, v4, v3 -; IR-DAZ-GISEL-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 -; IR-DAZ-GISEL-NEXT: v_div_fmas_f32 v1, v1, v2, v4 -; IR-DAZ-GISEL-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 -; IR-DAZ-GISEL-NEXT: s_setpc_b64 s[30:31] +; IR-DAZ-LABEL: v_recip_sqrt_f32_ulp25_contract: +; IR-DAZ: ; %bb.0: +; IR-DAZ-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) +; IR-DAZ-NEXT: v_sqrt_f32_e32 v0, v0 +; IR-DAZ-NEXT: v_div_scale_f32 v1, s[4:5], v0, v0, 1.0 +; IR-DAZ-NEXT: v_rcp_f32_e32 v2, v1 +; IR-DAZ-NEXT: v_div_scale_f32 v3, vcc, 1.0, v0, 1.0 +; IR-DAZ-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 3 +; IR-DAZ-NEXT: v_fma_f32 v4, -v1, v2, 1.0 +; IR-DAZ-NEXT: v_fma_f32 v2, v4, v2, v2 +; IR-DAZ-NEXT: v_mul_f32_e32 v4, v3, v2 +; IR-DAZ-NEXT: v_fma_f32 v5, -v1, v4, v3 +; IR-DAZ-NEXT: v_fma_f32 v4, v5, v2, v4 +; IR-DAZ-NEXT: v_fma_f32 v1, -v1, v4, v3 +; IR-DAZ-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_MODE, 4, 2), 0 +; IR-DAZ-NEXT: v_div_fmas_f32 v1, v1, v2, v4 +; IR-DAZ-NEXT: v_div_fixup_f32 v0, v1, v0, 1.0 +; IR-DAZ-NEXT: s_setpc_b64 s[30:31] %sqrt = call contract float @llvm.sqrt.f32(float %x), !fpmath !0 %fdiv = fdiv contract float 1.0, %sqrt, !fpmath !0 ret float %fdiv diff --git a/llvm/test/CodeGen/AMDGPU/fma-combine.ll b/llvm/test/CodeGen/AMDGPU/fma-combine.ll index d7cf411da0cc..8fc6904f5009 100644 --- a/llvm/test/CodeGen/AMDGPU/fma-combine.ll +++ b/llvm/test/CodeGen/AMDGPU/fma-combine.ll @@ -1,9 +1,9 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2 -; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math-f32=preserve-sign < %s | FileCheck -enable-var-scope -check-prefixes=SI,SI-NOFMA %s -; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=verde -denormal-fp-math-f32=preserve-sign < %s | FileCheck -enable-var-scope -check-prefixes=SI,SI-NOFMA %s -; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math-f32=ieee -enable-no-infs-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=SI,SI-FMA %s +; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math-f32=preserve-sign < %s | FileCheck -enable-var-scope -check-prefixes=SI,SI-NOFMA,TAHITI %s +; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=verde -denormal-fp-math-f32=preserve-sign < %s | FileCheck -enable-var-scope -check-prefixes=SI,SI-NOFMA,VERDE %s +; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=tahiti -denormal-fp-math-f32=ieee < %s | FileCheck -enable-var-scope -check-prefixes=SI,SI-FMA %s ; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1100 -denormal-fp-math-f32=preserve-sign < %s | FileCheck -enable-var-scope -check-prefixes=GFX11,GFX11-NOFMA %s -; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1100 -denormal-fp-math-f32=ieee -enable-no-infs-fp-math < %s | FileCheck -enable-var-scope -check-prefixes=GFX11,GFX11-FMA %s +; RUN: llc -amdgpu-scalarize-global-loads=false -mtriple=amdgcn -mcpu=gfx1100 -denormal-fp-math-f32=ieee < %s | FileCheck -enable-var-scope -check-prefixes=GFX11,GFX11-FMA %s ; Note: The SI-FMA conversions of type x * (y + 1) --> x * y + x would be ; beneficial even without fp32 denormals, but they do require no-infs-fp-math @@ -648,53 +648,53 @@ define amdgpu_kernel void @combine_to_fma_fsub_2_f64_2uses_mul(ptr addrspace(1) ; fold (fsub (fma x, y, (fmul u, v)), z) -> (fma x, y (fma u, v, (fneg z))) define amdgpu_kernel void @aggressive_combine_to_fma_fsub_0_f64(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #1 { -; SI-FMA-LABEL: aggressive_combine_to_fma_fsub_0_f64: -; SI-FMA: ; %bb.0: -; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 -; SI-FMA-NEXT: s_mov_b32 s7, 0xf000 -; SI-FMA-NEXT: s_mov_b32 s6, 0 -; SI-FMA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 -; SI-FMA-NEXT: v_mov_b32_e32 v1, 0 -; SI-FMA-NEXT: s_waitcnt lgkmcnt(0) -; SI-FMA-NEXT: s_mov_b64 s[4:5], s[2:3] -; SI-FMA-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:8 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:16 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[4:7], 0 addr64 offset:24 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: buffer_load_dwordx2 v[10:11], v[0:1], s[4:7], 0 addr64 offset:32 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: s_mov_b64 s[2:3], s[6:7] -; SI-FMA-NEXT: v_fma_f64 v[6:7], v[8:9], v[10:11], -v[6:7] -; SI-FMA-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[6:7] -; SI-FMA-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 -; SI-FMA-NEXT: s_endpgm +; SI-LABEL: aggressive_combine_to_fma_fsub_0_f64: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, 0 +; SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 +; SI-NEXT: v_mov_b32_e32 v1, 0 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b64 s[4:5], s[2:3] +; SI-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:8 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:16 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[4:7], 0 addr64 offset:24 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[10:11], v[0:1], s[4:7], 0 addr64 offset:32 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: s_mov_b64 s[2:3], s[6:7] +; SI-NEXT: v_fma_f64 v[6:7], v[8:9], v[10:11], -v[6:7] +; SI-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[6:7] +; SI-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: aggressive_combine_to_fma_fsub_0_f64: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-FMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FMA-NEXT: v_lshlrev_b32_e32 v10, 3, v0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[0:1], v10, s[2:3] glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[2:3], v10, s[2:3] offset:8 glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[4:5], v10, s[2:3] offset:16 glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[6:7], v10, s[2:3] offset:24 glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[8:9], v10, s[2:3] offset:32 glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f64 v[4:5], v[6:7], v[8:9], -v[4:5] -; GFX11-FMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FMA-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5] -; GFX11-FMA-NEXT: global_store_b64 v10, v[0:1], s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: aggressive_combine_to_fma_fsub_0_f64: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v0 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_lshlrev_b32_e32 v10, 3, v0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: global_load_b64 v[0:1], v10, s[2:3] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[2:3], v10, s[2:3] offset:8 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[4:5], v10, s[2:3] offset:16 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[6:7], v10, s[2:3] offset:24 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[8:9], v10, s[2:3] offset:32 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f64 v[4:5], v[6:7], v[8:9], -v[4:5] +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[4:5] +; GFX11-NEXT: global_store_b64 v10, v[0:1], s[0:1] +; GFX11-NEXT: s_endpgm %tid = tail call i32 @llvm.amdgcn.workitem.id.x() #0 %gep.0 = getelementptr double, ptr addrspace(1) %in, i32 %tid %gep.1 = getelementptr double, ptr addrspace(1) %gep.0, i32 1 @@ -718,55 +718,55 @@ define amdgpu_kernel void @aggressive_combine_to_fma_fsub_0_f64(ptr addrspace(1) } define amdgpu_kernel void @no_aggressive_combine_to_fma_fsub_0_f64(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #1 { -; SI-NOFMA-LABEL: no_aggressive_combine_to_fma_fsub_0_f64: -; SI-NOFMA: ; %bb.0: -; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 -; SI-NOFMA-NEXT: s_mov_b32 s7, 0xf000 -; SI-NOFMA-NEXT: s_mov_b32 s6, 0 -; SI-NOFMA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 -; SI-NOFMA-NEXT: v_mov_b32_e32 v1, 0 -; SI-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; SI-NOFMA-NEXT: s_mov_b64 s[4:5], s[2:3] -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:8 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:16 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[4:7], 0 addr64 offset:24 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[10:11], v[0:1], s[4:7], 0 addr64 offset:32 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: s_mov_b64 s[2:3], s[6:7] -; SI-NOFMA-NEXT: v_mul_f64 v[8:9], v[8:9], v[10:11] -; SI-NOFMA-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[8:9] -; SI-NOFMA-NEXT: v_add_f64 v[2:3], v[2:3], -v[6:7] -; SI-NOFMA-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 -; SI-NOFMA-NEXT: s_endpgm +; SI-LABEL: no_aggressive_combine_to_fma_fsub_0_f64: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, 0 +; SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 +; SI-NEXT: v_mov_b32_e32 v1, 0 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b64 s[4:5], s[2:3] +; SI-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:8 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:16 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[4:7], 0 addr64 offset:24 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[10:11], v[0:1], s[4:7], 0 addr64 offset:32 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: s_mov_b64 s[2:3], s[6:7] +; SI-NEXT: v_mul_f64 v[8:9], v[8:9], v[10:11] +; SI-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[8:9] +; SI-NEXT: v_add_f64 v[2:3], v[2:3], -v[6:7] +; SI-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 +; SI-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: no_aggressive_combine_to_fma_fsub_0_f64: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_lshlrev_b32_e32 v10, 3, v0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[0:1], v10, s[2:3] glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[2:3], v10, s[2:3] offset:8 glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[4:5], v10, s[2:3] offset:16 glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[6:7], v10, s[2:3] offset:24 glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[8:9], v10, s[2:3] offset:32 glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: v_mul_f64 v[6:7], v[6:7], v[8:9] -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[6:7] -; GFX11-NOFMA-NEXT: v_add_f64 v[0:1], v[0:1], -v[4:5] -; GFX11-NOFMA-NEXT: global_store_b64 v10, v[0:1], s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; GFX11-LABEL: no_aggressive_combine_to_fma_fsub_0_f64: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v0 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_lshlrev_b32_e32 v10, 3, v0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: global_load_b64 v[0:1], v10, s[2:3] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[2:3], v10, s[2:3] offset:8 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[4:5], v10, s[2:3] offset:16 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[6:7], v10, s[2:3] offset:24 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[8:9], v10, s[2:3] offset:32 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_mul_f64 v[6:7], v[6:7], v[8:9] +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-NEXT: v_fma_f64 v[0:1], v[0:1], v[2:3], v[6:7] +; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], -v[4:5] +; GFX11-NEXT: global_store_b64 v10, v[0:1], s[0:1] +; GFX11-NEXT: s_endpgm %tid = tail call i32 @llvm.amdgcn.workitem.id.x() #0 %gep.0 = getelementptr double, ptr addrspace(1) %in, i32 %tid %gep.1 = getelementptr double, ptr addrspace(1) %gep.0, i32 1 @@ -792,53 +792,53 @@ define amdgpu_kernel void @no_aggressive_combine_to_fma_fsub_0_f64(ptr addrspace ; fold (fsub x, (fma y, z, (fmul u, v))) ; -> (fma (fneg y), z, (fma (fneg u), v, x)) define amdgpu_kernel void @aggressive_combine_to_fma_fsub_1_f64(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #1 { -; SI-FMA-LABEL: aggressive_combine_to_fma_fsub_1_f64: -; SI-FMA: ; %bb.0: -; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 -; SI-FMA-NEXT: s_mov_b32 s7, 0xf000 -; SI-FMA-NEXT: s_mov_b32 s6, 0 -; SI-FMA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 -; SI-FMA-NEXT: v_mov_b32_e32 v1, 0 -; SI-FMA-NEXT: s_waitcnt lgkmcnt(0) -; SI-FMA-NEXT: s_mov_b64 s[4:5], s[2:3] -; SI-FMA-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:8 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:16 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[4:7], 0 addr64 offset:24 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: buffer_load_dwordx2 v[10:11], v[0:1], s[4:7], 0 addr64 offset:32 glc -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: s_mov_b64 s[2:3], s[6:7] -; SI-FMA-NEXT: v_fma_f64 v[2:3], -v[8:9], v[10:11], v[2:3] -; SI-FMA-NEXT: v_fma_f64 v[2:3], -v[4:5], v[6:7], v[2:3] -; SI-FMA-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 -; SI-FMA-NEXT: s_endpgm +; SI-LABEL: aggressive_combine_to_fma_fsub_1_f64: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, 0 +; SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 +; SI-NEXT: v_mov_b32_e32 v1, 0 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b64 s[4:5], s[2:3] +; SI-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:8 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:16 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[4:7], 0 addr64 offset:24 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[10:11], v[0:1], s[4:7], 0 addr64 offset:32 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: s_mov_b64 s[2:3], s[6:7] +; SI-NEXT: v_fma_f64 v[2:3], -v[8:9], v[10:11], v[2:3] +; SI-NEXT: v_fma_f64 v[2:3], -v[4:5], v[6:7], v[2:3] +; SI-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: aggressive_combine_to_fma_fsub_1_f64: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-FMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FMA-NEXT: v_lshlrev_b32_e32 v10, 3, v0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[0:1], v10, s[2:3] glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[2:3], v10, s[2:3] offset:8 glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[4:5], v10, s[2:3] offset:16 glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[6:7], v10, s[2:3] offset:24 glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b64 v[8:9], v10, s[2:3] offset:32 glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f64 v[0:1], -v[6:7], v[8:9], v[0:1] -; GFX11-FMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FMA-NEXT: v_fma_f64 v[0:1], -v[2:3], v[4:5], v[0:1] -; GFX11-FMA-NEXT: global_store_b64 v10, v[0:1], s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: aggressive_combine_to_fma_fsub_1_f64: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v0 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_lshlrev_b32_e32 v10, 3, v0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: global_load_b64 v[0:1], v10, s[2:3] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[2:3], v10, s[2:3] offset:8 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[4:5], v10, s[2:3] offset:16 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[6:7], v10, s[2:3] offset:24 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[8:9], v10, s[2:3] offset:32 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f64 v[0:1], -v[6:7], v[8:9], v[0:1] +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_fma_f64 v[0:1], -v[2:3], v[4:5], v[0:1] +; GFX11-NEXT: global_store_b64 v10, v[0:1], s[0:1] +; GFX11-NEXT: s_endpgm %tid = tail call i32 @llvm.amdgcn.workitem.id.x() #0 %gep.0 = getelementptr double, ptr addrspace(1) %in, i32 %tid %gep.1 = getelementptr double, ptr addrspace(1) %gep.0, i32 1 @@ -862,55 +862,55 @@ define amdgpu_kernel void @aggressive_combine_to_fma_fsub_1_f64(ptr addrspace(1) ret void } define amdgpu_kernel void @no_aggressive_combine_to_fma_fsub_1_f64(ptr addrspace(1) noalias %out, ptr addrspace(1) noalias %in) #1 { -; SI-NOFMA-LABEL: no_aggressive_combine_to_fma_fsub_1_f64: -; SI-NOFMA: ; %bb.0: -; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 -; SI-NOFMA-NEXT: s_mov_b32 s7, 0xf000 -; SI-NOFMA-NEXT: s_mov_b32 s6, 0 -; SI-NOFMA-NEXT: v_lshlrev_b32_e32 v0, 3, v0 -; SI-NOFMA-NEXT: v_mov_b32_e32 v1, 0 -; SI-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; SI-NOFMA-NEXT: s_mov_b64 s[4:5], s[2:3] -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:8 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:16 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[4:7], 0 addr64 offset:24 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: buffer_load_dwordx2 v[10:11], v[0:1], s[4:7], 0 addr64 offset:32 glc -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: s_mov_b64 s[2:3], s[6:7] -; SI-NOFMA-NEXT: v_mul_f64 v[8:9], v[8:9], v[10:11] -; SI-NOFMA-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9] -; SI-NOFMA-NEXT: v_add_f64 v[2:3], v[2:3], -v[4:5] -; SI-NOFMA-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 -; SI-NOFMA-NEXT: s_endpgm +; SI-LABEL: no_aggressive_combine_to_fma_fsub_1_f64: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, 0 +; SI-NEXT: v_lshlrev_b32_e32 v0, 3, v0 +; SI-NEXT: v_mov_b32_e32 v1, 0 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b64 s[4:5], s[2:3] +; SI-NEXT: buffer_load_dwordx2 v[2:3], v[0:1], s[4:7], 0 addr64 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[4:5], v[0:1], s[4:7], 0 addr64 offset:8 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[6:7], v[0:1], s[4:7], 0 addr64 offset:16 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[8:9], v[0:1], s[4:7], 0 addr64 offset:24 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dwordx2 v[10:11], v[0:1], s[4:7], 0 addr64 offset:32 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: s_mov_b64 s[2:3], s[6:7] +; SI-NEXT: v_mul_f64 v[8:9], v[8:9], v[10:11] +; SI-NEXT: v_fma_f64 v[4:5], v[4:5], v[6:7], v[8:9] +; SI-NEXT: v_add_f64 v[2:3], v[2:3], -v[4:5] +; SI-NEXT: buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64 +; SI-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: no_aggressive_combine_to_fma_fsub_1_f64: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: v_and_b32_e32 v0, 0x3ff, v0 -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_lshlrev_b32_e32 v10, 3, v0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[0:1], v10, s[2:3] glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[2:3], v10, s[2:3] offset:8 glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[4:5], v10, s[2:3] offset:16 glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[6:7], v10, s[2:3] offset:24 glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b64 v[8:9], v10, s[2:3] offset:32 glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: v_mul_f64 v[6:7], v[6:7], v[8:9] -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[6:7] -; GFX11-NOFMA-NEXT: v_add_f64 v[0:1], v[0:1], -v[2:3] -; GFX11-NOFMA-NEXT: global_store_b64 v10, v[0:1], s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; GFX11-LABEL: no_aggressive_combine_to_fma_fsub_1_f64: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: v_and_b32_e32 v0, 0x3ff, v0 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_lshlrev_b32_e32 v10, 3, v0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: global_load_b64 v[0:1], v10, s[2:3] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[2:3], v10, s[2:3] offset:8 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[4:5], v10, s[2:3] offset:16 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[6:7], v10, s[2:3] offset:24 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b64 v[8:9], v10, s[2:3] offset:32 glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_mul_f64 v[6:7], v[6:7], v[8:9] +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-NEXT: v_fma_f64 v[2:3], v[2:3], v[4:5], v[6:7] +; GFX11-NEXT: v_add_f64 v[0:1], v[0:1], -v[2:3] +; GFX11-NEXT: global_store_b64 v10, v[0:1], s[0:1] +; GFX11-NEXT: s_endpgm %tid = tail call i32 @llvm.amdgcn.workitem.id.x() #0 %gep.0 = getelementptr double, ptr addrspace(1) %in, i32 %tid %gep.1 = getelementptr double, ptr addrspace(1) %gep.0, i32 1 @@ -939,7 +939,58 @@ define amdgpu_kernel void @no_aggressive_combine_to_fma_fsub_1_f64(ptr addrspace ; define amdgpu_kernel void @test_f32_mul_add_x_one_y(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_add_x_one_y: +; SI-LABEL: test_f32_mul_add_x_one_y: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 +; SI-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm +; +; GFX11-LABEL: test_f32_mul_add_x_one_y: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load volatile float, ptr addrspace(1) %in1 + %y = load volatile float, ptr addrspace(1) %in2 + %a = fadd contract float %x, 1.0 + %m = fmul contract float %a, %y + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_add_x_one_y_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_add_x_one_y_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -958,12 +1009,11 @@ define amdgpu_kernel void @test_f32_mul_add_x_one_y(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: v_add_f32_e32 v0, 1.0, v0 -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v0, v1 -; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NOFMA-NEXT: v_mac_f32_e32 v1, v0, v1 +; SI-NOFMA-NEXT: buffer_store_dword v1, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_add_x_one_y: +; SI-FMA-LABEL: test_f32_mul_add_x_one_y_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -986,49 +1036,83 @@ define amdgpu_kernel void @test_f32_mul_add_x_one_y(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_add_x_one_y: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: v_add_f32_e32 v1, 1.0, v1 -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v1, v2 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_add_x_one_y_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fmac_f32_e32 v2, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v2, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load volatile float, ptr addrspace(1) %in1 + %y = load volatile float, ptr addrspace(1) %in2 + %a = fadd contract ninf float %x, 1.0 + %m = fmul contract ninf float %a, %y + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_y_add_x_one(ptr addrspace(1) %out, +; SI-LABEL: test_f32_mul_y_add_x_one: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 glc +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 +; SI-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: test_f32_mul_add_x_one_y: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fmac_f32_e32 v2, v1, v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v2, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_add_x_one: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v2, v1 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load volatile float, ptr addrspace(1) %in1 %y = load volatile float, ptr addrspace(1) %in2 %a = fadd contract float %x, 1.0 - %m = fmul contract float %a, %y + %m = fmul contract float %y, %a store float %m, ptr addrspace(1) %out ret void } -define amdgpu_kernel void @test_f32_mul_y_add_x_one(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_y_add_x_one: +define amdgpu_kernel void @test_f32_mul_y_add_x_one_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_y_add_x_one_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1047,12 +1131,11 @@ define amdgpu_kernel void @test_f32_mul_y_add_x_one(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: v_add_f32_e32 v0, 1.0, v0 -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v1, v0 -; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NOFMA-NEXT: v_mac_f32_e32 v1, v0, v1 +; SI-NOFMA-NEXT: buffer_store_dword v1, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_y_add_x_one: +; SI-FMA-LABEL: test_f32_mul_y_add_x_one_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1075,126 +1158,72 @@ define amdgpu_kernel void @test_f32_mul_y_add_x_one(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_y_add_x_one: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] glc dlc -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: v_add_f32_e32 v1, 1.0, v1 -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v2, v1 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm -; -; GFX11-FMA-LABEL: test_f32_mul_y_add_x_one: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] glc dlc -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fmac_f32_e32 v2, v1, v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v2, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_add_x_one_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] glc dlc +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fmac_f32_e32 v2, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v2, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load volatile float, ptr addrspace(1) %in1 %y = load volatile float, ptr addrspace(1) %in2 - %a = fadd contract float %x, 1.0 - %m = fmul contract float %y, %a + %a = fadd contract ninf float %x, 1.0 + %m = fmul contract ninf float %y, %a store float %m, ptr addrspace(1) %out ret void } define amdgpu_kernel void @test_f32_mul_add_x_negone_y(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_add_x_negone_y: -; SI-NOFMA: ; %bb.0: -; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 -; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd -; SI-NOFMA-NEXT: s_mov_b32 s7, 0xf000 -; SI-NOFMA-NEXT: s_mov_b32 s6, -1 -; SI-NOFMA-NEXT: s_mov_b32 s14, s6 -; SI-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; SI-NOFMA-NEXT: s_mov_b32 s12, s2 -; SI-NOFMA-NEXT: s_mov_b32 s13, s3 -; SI-NOFMA-NEXT: s_mov_b32 s15, s7 -; SI-NOFMA-NEXT: s_mov_b32 s10, s6 -; SI-NOFMA-NEXT: s_mov_b32 s11, s7 -; SI-NOFMA-NEXT: buffer_load_dword v0, off, s[12:15], 0 -; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 -; SI-NOFMA-NEXT: s_mov_b32 s4, s0 -; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_add_f32_e32 v0, -1.0, v0 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v0, v1 -; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-NOFMA-NEXT: s_endpgm -; -; SI-FMA-LABEL: test_f32_mul_add_x_negone_y: -; SI-FMA: ; %bb.0: -; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 -; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd -; SI-FMA-NEXT: s_mov_b32 s7, 0xf000 -; SI-FMA-NEXT: s_mov_b32 s6, -1 -; SI-FMA-NEXT: s_mov_b32 s14, s6 -; SI-FMA-NEXT: s_waitcnt lgkmcnt(0) -; SI-FMA-NEXT: s_mov_b32 s12, s2 -; SI-FMA-NEXT: s_mov_b32 s13, s3 -; SI-FMA-NEXT: s_mov_b32 s15, s7 -; SI-FMA-NEXT: s_mov_b32 s10, s6 -; SI-FMA-NEXT: s_mov_b32 s11, s7 -; SI-FMA-NEXT: buffer_load_dword v0, off, s[12:15], 0 -; SI-FMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 -; SI-FMA-NEXT: s_mov_b32 s4, s0 -; SI-FMA-NEXT: s_mov_b32 s5, s1 -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: v_fma_f32 v0, v0, v1, -v1 -; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-FMA-NEXT: s_endpgm -; -; GFX11-NOFMA-LABEL: test_f32_mul_add_x_negone_y: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_add_f32_e32 v1, -1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v1, v2 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; SI-LABEL: test_f32_mul_add_x_negone_y: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_add_f32_e32 v0, -1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: test_f32_mul_add_x_negone_y: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f32 v1, v1, v2, -v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_add_x_negone_y: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_add_f32_e32 v1, -1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 @@ -1205,8 +1234,8 @@ define amdgpu_kernel void @test_f32_mul_add_x_negone_y(ptr addrspace(1) %out, ret void } -define amdgpu_kernel void @test_f32_mul_y_add_x_negone(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_y_add_x_negone: +define amdgpu_kernel void @test_f32_mul_add_x_negone_y_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_add_x_negone_y_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1223,14 +1252,12 @@ define amdgpu_kernel void @test_f32_mul_y_add_x_negone(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_add_f32_e32 v0, -1.0, v0 ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NOFMA-NEXT: v_mad_f32 v0, v0, v1, -v1 ; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_y_add_x_negone: +; SI-FMA-LABEL: test_f32_mul_add_x_negone_y_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1252,38 +1279,72 @@ define amdgpu_kernel void @test_f32_mul_y_add_x_negone(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_y_add_x_negone: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_add_f32_e32 v1, -1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v2, v1 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_add_x_negone_y_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f32 v1, v1, v2, -v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %a = fadd contract ninf float %x, -1.0 + %m = fmul contract ninf float %a, %y + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_y_add_x_negone(ptr addrspace(1) %out, +; SI-LABEL: test_f32_mul_y_add_x_negone: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_add_f32_e32 v0, -1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: test_f32_mul_y_add_x_negone: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f32 v1, v1, v2, -v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_add_x_negone: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_add_f32_e32 v1, -1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v2, v1 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 @@ -1294,8 +1355,8 @@ define amdgpu_kernel void @test_f32_mul_y_add_x_negone(ptr addrspace(1) %out, ret void } -define amdgpu_kernel void @test_f32_mul_sub_one_x_y(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_sub_one_x_y: +define amdgpu_kernel void @test_f32_mul_y_add_x_negone_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_y_add_x_negone_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1312,14 +1373,12 @@ define amdgpu_kernel void @test_f32_mul_sub_one_x_y(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_sub_f32_e32 v0, 1.0, v0 ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NOFMA-NEXT: v_mad_f32 v0, v0, v1, -v1 ; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_sub_one_x_y: +; SI-FMA-LABEL: test_f32_mul_y_add_x_negone_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1337,42 +1396,76 @@ define amdgpu_kernel void @test_f32_mul_sub_one_x_y(ptr addrspace(1) %out, ; SI-FMA-NEXT: s_mov_b32 s4, s0 ; SI-FMA-NEXT: s_mov_b32 s5, s1 ; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: v_fma_f32 v0, -v0, v1, v1 +; SI-FMA-NEXT: v_fma_f32 v0, v0, v1, -v1 ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_sub_one_x_y: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_sub_f32_e32 v1, 1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v1, v2 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_add_x_negone_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f32 v1, v1, v2, -v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %a = fadd contract ninf float %x, -1.0 + %m = fmul contract ninf float %y, %a + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_sub_one_x_y(ptr addrspace(1) %out, +; SI-LABEL: test_f32_mul_sub_one_x_y: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_sub_f32_e32 v0, 1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: test_f32_mul_sub_one_x_y: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f32 v1, -v1, v2, v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_sub_one_x_y: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_sub_f32_e32 v1, 1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 @@ -1383,85 +1476,100 @@ define amdgpu_kernel void @test_f32_mul_sub_one_x_y(ptr addrspace(1) %out, ret void } -define amdgpu_kernel void @test_f32_mul_y_sub_one_x(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_y_sub_one_x: -; SI-NOFMA: ; %bb.0: -; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 -; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd -; SI-NOFMA-NEXT: s_mov_b32 s7, 0xf000 -; SI-NOFMA-NEXT: s_mov_b32 s6, -1 -; SI-NOFMA-NEXT: s_mov_b32 s14, s6 -; SI-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; SI-NOFMA-NEXT: s_mov_b32 s12, s2 -; SI-NOFMA-NEXT: s_mov_b32 s13, s3 -; SI-NOFMA-NEXT: s_mov_b32 s15, s7 -; SI-NOFMA-NEXT: s_mov_b32 s10, s6 -; SI-NOFMA-NEXT: s_mov_b32 s11, s7 -; SI-NOFMA-NEXT: buffer_load_dword v0, off, s[12:15], 0 -; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 -; SI-NOFMA-NEXT: s_mov_b32 s4, s0 -; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_sub_f32_e32 v0, 1.0, v0 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v1, v0 -; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-NOFMA-NEXT: s_endpgm -; -; SI-FMA-LABEL: test_f32_mul_y_sub_one_x: -; SI-FMA: ; %bb.0: -; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 -; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd -; SI-FMA-NEXT: s_mov_b32 s7, 0xf000 -; SI-FMA-NEXT: s_mov_b32 s6, -1 -; SI-FMA-NEXT: s_mov_b32 s14, s6 -; SI-FMA-NEXT: s_waitcnt lgkmcnt(0) -; SI-FMA-NEXT: s_mov_b32 s12, s2 -; SI-FMA-NEXT: s_mov_b32 s13, s3 -; SI-FMA-NEXT: s_mov_b32 s15, s7 -; SI-FMA-NEXT: s_mov_b32 s10, s6 -; SI-FMA-NEXT: s_mov_b32 s11, s7 -; SI-FMA-NEXT: buffer_load_dword v0, off, s[12:15], 0 -; SI-FMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 -; SI-FMA-NEXT: s_mov_b32 s4, s0 -; SI-FMA-NEXT: s_mov_b32 s5, s1 -; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: v_fma_f32 v0, -v0, v1, v1 -; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 -; SI-FMA-NEXT: s_endpgm +define amdgpu_kernel void @test_f32_mul_sub_one_x_y_ninf(ptr addrspace(1) %out, +; SI-LABEL: test_f32_mul_sub_one_x_y_ninf: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_sub_f32_e32 v0, 1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_y_sub_one_x: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_sub_f32_e32 v1, 1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v2, v1 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_sub_one_x_y_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_sub_f32_e32 v1, 1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %s = fsub contract ninf float 1.0, %x + %m = fmul contract ninf float %s, %y + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_y_sub_one_x(ptr addrspace(1) %out, +; SI-LABEL: test_f32_mul_y_sub_one_x: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_sub_f32_e32 v0, 1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: test_f32_mul_y_sub_one_x: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f32 v1, -v1, v2, v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_sub_one_x: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_sub_f32_e32 v1, 1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v2, v1 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 @@ -1472,8 +1580,8 @@ define amdgpu_kernel void @test_f32_mul_y_sub_one_x(ptr addrspace(1) %out, ret void } -define amdgpu_kernel void @test_f32_mul_sub_negone_x_y(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_sub_negone_x_y: +define amdgpu_kernel void @test_f32_mul_y_sub_one_x_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_y_sub_one_x_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1490,14 +1598,12 @@ define amdgpu_kernel void @test_f32_mul_sub_negone_x_y(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_sub_f32_e32 v0, -1.0, v0 ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NOFMA-NEXT: v_mad_f32 v0, -v0, v1, v1 ; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_sub_negone_x_y: +; SI-FMA-LABEL: test_f32_mul_y_sub_one_x_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1515,42 +1621,76 @@ define amdgpu_kernel void @test_f32_mul_sub_negone_x_y(ptr addrspace(1) %out, ; SI-FMA-NEXT: s_mov_b32 s4, s0 ; SI-FMA-NEXT: s_mov_b32 s5, s1 ; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: v_fma_f32 v0, -v0, v1, -v1 +; SI-FMA-NEXT: v_fma_f32 v0, -v0, v1, v1 ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_sub_negone_x_y: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_sub_f32_e32 v1, -1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v1, v2 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_sub_one_x_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f32 v1, -v1, v2, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %s = fsub contract ninf float 1.0, %x + %m = fmul contract ninf float %y, %s + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_sub_negone_x_y(ptr addrspace(1) %out, +; SI-LABEL: test_f32_mul_sub_negone_x_y: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_sub_f32_e32 v0, -1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: test_f32_mul_sub_negone_x_y: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f32 v1, -v1, v2, -v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_sub_negone_x_y: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_sub_f32_e32 v1, -1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 @@ -1561,8 +1701,112 @@ define amdgpu_kernel void @test_f32_mul_sub_negone_x_y(ptr addrspace(1) %out, ret void } +define amdgpu_kernel void @test_f32_mul_sub_negone_x_y_ninf(ptr addrspace(1) %out, +; SI-LABEL: test_f32_mul_sub_negone_x_y_ninf: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_sub_f32_e32 v0, -1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm +; +; GFX11-LABEL: test_f32_mul_sub_negone_x_y_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_sub_f32_e32 v1, -1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %s = fsub contract ninf float -1.0, %x + %m = fmul contract ninf float %s, %y + store float %m, ptr addrspace(1) %out + ret void +} + define amdgpu_kernel void @test_f32_mul_y_sub_negone_x(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_y_sub_negone_x: +; SI-LABEL: test_f32_mul_y_sub_negone_x: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_sub_f32_e32 v0, -1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm +; +; GFX11-LABEL: test_f32_mul_y_sub_negone_x: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_sub_f32_e32 v1, -1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v2, v1 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %s = fsub contract float -1.0, %x + %m = fmul contract float %y, %s + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_y_sub_negone_x_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_y_sub_negone_x_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1579,14 +1823,12 @@ define amdgpu_kernel void @test_f32_mul_y_sub_negone_x(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_sub_f32_e32 v0, -1.0, v0 ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NOFMA-NEXT: v_mad_f32 v0, -v0, v1, -v1 ; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_y_sub_negone_x: +; SI-FMA-LABEL: test_f32_mul_y_sub_negone_x_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1608,50 +1850,84 @@ define amdgpu_kernel void @test_f32_mul_y_sub_negone_x(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_y_sub_negone_x: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_sub_f32_e32 v1, -1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v2, v1 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm -; -; GFX11-FMA-LABEL: test_f32_mul_y_sub_negone_x: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f32 v1, -v1, v2, -v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_sub_negone_x_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f32 v1, -v1, v2, -v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 %y = load float, ptr addrspace(1) %in2 - %s = fsub contract float -1.0, %x - %m = fmul contract float %y, %s + %s = fsub contract ninf float -1.0, %x + %m = fmul contract ninf float %y, %s store float %m, ptr addrspace(1) %out ret void } define amdgpu_kernel void @test_f32_mul_sub_x_one_y(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_sub_x_one_y: +; SI-LABEL: test_f32_mul_sub_x_one_y: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_add_f32_e32 v0, -1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm +; +; GFX11-LABEL: test_f32_mul_sub_x_one_y: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_add_f32_e32 v1, -1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %s = fsub contract float %x, 1.0 + %m = fmul contract float %s, %y + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_sub_x_one_y_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_sub_x_one_y_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1668,14 +1944,12 @@ define amdgpu_kernel void @test_f32_mul_sub_x_one_y(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_add_f32_e32 v0, -1.0, v0 ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NOFMA-NEXT: v_mad_f32 v0, v0, v1, -v1 ; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_sub_x_one_y: +; SI-FMA-LABEL: test_f32_mul_sub_x_one_y_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1697,50 +1971,84 @@ define amdgpu_kernel void @test_f32_mul_sub_x_one_y(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_sub_x_one_y: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_add_f32_e32 v1, -1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v1, v2 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm -; -; GFX11-FMA-LABEL: test_f32_mul_sub_x_one_y: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f32 v1, v1, v2, -v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_sub_x_one_y_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f32 v1, v1, v2, -v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 %y = load float, ptr addrspace(1) %in2 - %s = fsub contract float %x, 1.0 - %m = fmul contract float %s, %y + %s = fsub contract ninf float %x, 1.0 + %m = fmul contract ninf float %s, %y store float %m, ptr addrspace(1) %out ret void } define amdgpu_kernel void @test_f32_mul_y_sub_x_one(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_y_sub_x_one: +; SI-LABEL: test_f32_mul_y_sub_x_one: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_add_f32_e32 v0, -1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm +; +; GFX11-LABEL: test_f32_mul_y_sub_x_one: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_add_f32_e32 v1, -1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v2, v1 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %s = fsub contract float %x, 1.0 + %m = fmul contract float %y, %s + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_y_sub_x_one_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_y_sub_x_one_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1757,14 +2065,12 @@ define amdgpu_kernel void @test_f32_mul_y_sub_x_one(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_add_f32_e32 v0, -1.0, v0 ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NOFMA-NEXT: v_mad_f32 v0, v0, v1, -v1 ; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_y_sub_x_one: +; SI-FMA-LABEL: test_f32_mul_y_sub_x_one_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1786,50 +2092,84 @@ define amdgpu_kernel void @test_f32_mul_y_sub_x_one(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_y_sub_x_one: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_add_f32_e32 v1, -1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v2, v1 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm -; -; GFX11-FMA-LABEL: test_f32_mul_y_sub_x_one: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fma_f32 v1, v1, v2, -v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_sub_x_one_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f32 v1, v1, v2, -v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 %y = load float, ptr addrspace(1) %in2 - %s = fsub contract float %x, 1.0 - %m = fmul contract float %y, %s + %s = fsub contract ninf float %x, 1.0 + %m = fmul contract ninf float %y, %s store float %m, ptr addrspace(1) %out ret void } define amdgpu_kernel void @test_f32_mul_sub_x_negone_y(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_sub_x_negone_y: +; SI-LABEL: test_f32_mul_sub_x_negone_y: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm +; +; GFX11-LABEL: test_f32_mul_sub_x_negone_y: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %s = fsub contract float %x, -1.0 + %m = fmul contract float %s, %y + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_sub_x_negone_y_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_sub_x_negone_y_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1846,14 +2186,12 @@ define amdgpu_kernel void @test_f32_mul_sub_x_negone_y(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_add_f32_e32 v0, 1.0, v0 ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v0, v1 -; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NOFMA-NEXT: v_mac_f32_e32 v1, v0, v1 +; SI-NOFMA-NEXT: buffer_store_dword v1, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_sub_x_negone_y: +; SI-FMA-LABEL: test_f32_mul_sub_x_negone_y_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1875,50 +2213,84 @@ define amdgpu_kernel void @test_f32_mul_sub_x_negone_y(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_sub_x_negone_y: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_add_f32_e32 v1, 1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v1, v2 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_sub_x_negone_y_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fmac_f32_e32 v2, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v2, s[0:1] +; GFX11-NEXT: s_endpgm + ptr addrspace(1) %in1, + ptr addrspace(1) %in2) { + %x = load float, ptr addrspace(1) %in1 + %y = load float, ptr addrspace(1) %in2 + %s = fsub contract ninf float %x, -1.0 + %m = fmul contract ninf float %s, %y + store float %m, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @test_f32_mul_y_sub_x_negone(ptr addrspace(1) %out, +; SI-LABEL: test_f32_mul_y_sub_x_negone: +; SI: ; %bb.0: +; SI-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; SI-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd +; SI-NEXT: s_mov_b32 s7, 0xf000 +; SI-NEXT: s_mov_b32 s6, -1 +; SI-NEXT: s_mov_b32 s14, s6 +; SI-NEXT: s_waitcnt lgkmcnt(0) +; SI-NEXT: s_mov_b32 s12, s2 +; SI-NEXT: s_mov_b32 s13, s3 +; SI-NEXT: s_mov_b32 s15, s7 +; SI-NEXT: s_mov_b32 s10, s6 +; SI-NEXT: s_mov_b32 s11, s7 +; SI-NEXT: buffer_load_dword v0, off, s[12:15], 0 +; SI-NEXT: buffer_load_dword v1, off, s[8:11], 0 +; SI-NEXT: s_mov_b32 s4, s0 +; SI-NEXT: s_mov_b32 s5, s1 +; SI-NEXT: s_waitcnt vmcnt(1) +; SI-NEXT: v_add_f32_e32 v0, 1.0, v0 +; SI-NEXT: s_waitcnt vmcnt(0) +; SI-NEXT: v_mul_f32_e32 v0, v1, v0 +; SI-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NEXT: s_endpgm ; -; GFX11-FMA-LABEL: test_f32_mul_sub_x_negone_y: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fmac_f32_e32 v2, v1, v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v2, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_sub_x_negone: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: v_add_f32_e32 v1, 1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v1, v2, v1 +; GFX11-NEXT: global_store_b32 v0, v1, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 %y = load float, ptr addrspace(1) %in2 %s = fsub contract float %x, -1.0 - %m = fmul contract float %s, %y + %m = fmul contract float %y, %s store float %m, ptr addrspace(1) %out ret void } -define amdgpu_kernel void @test_f32_mul_y_sub_x_negone(ptr addrspace(1) %out, -; SI-NOFMA-LABEL: test_f32_mul_y_sub_x_negone: +define amdgpu_kernel void @test_f32_mul_y_sub_x_negone_ninf(ptr addrspace(1) %out, +; SI-NOFMA-LABEL: test_f32_mul_y_sub_x_negone_ninf: ; SI-NOFMA: ; %bb.0: ; SI-NOFMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-NOFMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1935,14 +2307,12 @@ define amdgpu_kernel void @test_f32_mul_y_sub_x_negone(ptr addrspace(1) %out, ; SI-NOFMA-NEXT: buffer_load_dword v1, off, s[8:11], 0 ; SI-NOFMA-NEXT: s_mov_b32 s4, s0 ; SI-NOFMA-NEXT: s_mov_b32 s5, s1 -; SI-NOFMA-NEXT: s_waitcnt vmcnt(1) -; SI-NOFMA-NEXT: v_add_f32_e32 v0, 1.0, v0 ; SI-NOFMA-NEXT: s_waitcnt vmcnt(0) -; SI-NOFMA-NEXT: v_mul_f32_e32 v0, v1, v0 -; SI-NOFMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 +; SI-NOFMA-NEXT: v_mac_f32_e32 v1, v0, v1 +; SI-NOFMA-NEXT: buffer_store_dword v1, off, s[4:7], 0 ; SI-NOFMA-NEXT: s_endpgm ; -; SI-FMA-LABEL: test_f32_mul_y_sub_x_negone: +; SI-FMA-LABEL: test_f32_mul_y_sub_x_negone_ninf: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 ; SI-FMA-NEXT: s_load_dwordx2 s[8:9], s[4:5], 0xd @@ -1964,44 +2334,26 @@ define amdgpu_kernel void @test_f32_mul_y_sub_x_negone(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_store_dword v0, off, s[4:7], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_mul_y_sub_x_negone: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x1 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: v_add_f32_e32 v1, 1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v1, v2, v1 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm -; -; GFX11-FMA-LABEL: test_f32_mul_y_sub_x_negone: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 -; GFX11-FMA-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x1 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[2:3] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: v_fmac_f32_e32 v2, v1, v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v2, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_mul_y_sub_x_negone_ninf: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: s_load_b128 s[0:3], s[4:5], 0x24 +; GFX11-NEXT: s_load_b64 s[4:5], s[4:5], 0x34 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x1 +; GFX11-NEXT: global_load_b32 v1, v0, s[2:3] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fmac_f32_e32 v2, v1, v2 +; GFX11-NEXT: global_store_b32 v0, v2, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2) { %x = load float, ptr addrspace(1) %in1 %y = load float, ptr addrspace(1) %in2 - %s = fsub contract float %x, -1.0 - %m = fmul contract float %y, %s + %s = fsub contract ninf float %x, -1.0 + %m = fmul contract ninf float %y, %s store float %m, ptr addrspace(1) %out ret void } @@ -2048,66 +2400,51 @@ define amdgpu_kernel void @test_f32_interp(ptr addrspace(1) %out, ; SI-FMA-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9 ; SI-FMA-NEXT: s_mov_b32 s11, 0xf000 ; SI-FMA-NEXT: s_mov_b32 s10, -1 -; SI-FMA-NEXT: s_mov_b32 s18, s10 -; SI-FMA-NEXT: s_mov_b32 s19, s11 +; SI-FMA-NEXT: s_mov_b32 s14, s10 +; SI-FMA-NEXT: s_mov_b32 s15, s11 ; SI-FMA-NEXT: s_waitcnt lgkmcnt(0) ; SI-FMA-NEXT: s_mov_b32 s16, s4 ; SI-FMA-NEXT: s_mov_b32 s17, s5 -; SI-FMA-NEXT: s_mov_b32 s14, s10 -; SI-FMA-NEXT: s_mov_b32 s12, s2 -; SI-FMA-NEXT: s_mov_b32 s13, s3 -; SI-FMA-NEXT: s_mov_b32 s15, s11 ; SI-FMA-NEXT: s_mov_b32 s4, s6 ; SI-FMA-NEXT: s_mov_b32 s5, s7 ; SI-FMA-NEXT: s_mov_b32 s6, s10 ; SI-FMA-NEXT: s_mov_b32 s7, s11 -; SI-FMA-NEXT: buffer_load_dword v0, off, s[16:19], 0 -; SI-FMA-NEXT: buffer_load_dword v1, off, s[4:7], 0 +; SI-FMA-NEXT: s_mov_b32 s12, s2 +; SI-FMA-NEXT: s_mov_b32 s13, s3 +; SI-FMA-NEXT: s_mov_b32 s18, s10 +; SI-FMA-NEXT: s_mov_b32 s19, s11 +; SI-FMA-NEXT: buffer_load_dword v0, off, s[4:7], 0 +; SI-FMA-NEXT: buffer_load_dword v1, off, s[16:19], 0 ; SI-FMA-NEXT: buffer_load_dword v2, off, s[12:15], 0 ; SI-FMA-NEXT: s_mov_b32 s8, s0 ; SI-FMA-NEXT: s_mov_b32 s9, s1 +; SI-FMA-NEXT: s_waitcnt vmcnt(2) +; SI-FMA-NEXT: v_sub_f32_e32 v3, 1.0, v0 ; SI-FMA-NEXT: s_waitcnt vmcnt(1) -; SI-FMA-NEXT: v_fma_f32 v0, -v1, v0, v0 +; SI-FMA-NEXT: v_mul_f32_e32 v1, v1, v3 ; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: v_fma_f32 v0, v2, v1, v0 +; SI-FMA-NEXT: v_fma_f32 v0, v2, v0, v1 ; SI-FMA-NEXT: buffer_store_dword v0, off, s[8:11], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f32_interp: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_load_b256 s[0:7], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x2 -; GFX11-NOFMA-NEXT: global_load_b32 v1, v0, s[6:7] -; GFX11-NOFMA-NEXT: global_load_b32 v2, v0, s[4:5] -; GFX11-NOFMA-NEXT: global_load_b32 v3, v0, s[2:3] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(2) -; GFX11-NOFMA-NEXT: v_sub_f32_e32 v4, 1.0, v1 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f32_e32 v2, v2, v4 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: v_fmac_f32_e32 v2, v3, v1 -; GFX11-NOFMA-NEXT: global_store_b32 v0, v2, s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm -; -; GFX11-FMA-LABEL: test_f32_interp: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_load_b256 s[0:7], s[4:5], 0x24 -; GFX11-FMA-NEXT: v_mov_b32_e32 v0, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x2 -; GFX11-FMA-NEXT: global_load_b32 v1, v0, s[4:5] -; GFX11-FMA-NEXT: global_load_b32 v2, v0, s[6:7] -; GFX11-FMA-NEXT: global_load_b32 v3, v0, s[2:3] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-FMA-NEXT: v_fma_f32 v1, -v2, v1, v1 -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FMA-NEXT: v_fmac_f32_e32 v1, v3, v2 -; GFX11-FMA-NEXT: global_store_b32 v0, v1, s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f32_interp: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24 +; GFX11-NEXT: v_mov_b32_e32 v0, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x2 +; GFX11-NEXT: global_load_b32 v1, v0, s[6:7] +; GFX11-NEXT: global_load_b32 v2, v0, s[4:5] +; GFX11-NEXT: global_load_b32 v3, v0, s[2:3] +; GFX11-NEXT: s_waitcnt vmcnt(2) +; GFX11-NEXT: v_sub_f32_e32 v4, 1.0, v1 +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-NEXT: v_mul_f32_e32 v2, v2, v4 +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fmac_f32_e32 v2, v3, v1 +; GFX11-NEXT: global_store_b32 v0, v2, s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2, ptr addrspace(1) %in3) { @@ -2123,6 +2460,66 @@ define amdgpu_kernel void @test_f32_interp(ptr addrspace(1) %out, } define amdgpu_kernel void @test_f64_interp(ptr addrspace(1) %out, +; TAHITI-LABEL: test_f64_interp: +; TAHITI: ; %bb.0: +; TAHITI-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9 +; TAHITI-NEXT: s_mov_b32 s11, 0xf000 +; TAHITI-NEXT: s_mov_b32 s10, -1 +; TAHITI-NEXT: s_mov_b32 s18, s10 +; TAHITI-NEXT: s_mov_b32 s19, s11 +; TAHITI-NEXT: s_waitcnt lgkmcnt(0) +; TAHITI-NEXT: s_mov_b32 s16, s4 +; TAHITI-NEXT: s_mov_b32 s17, s5 +; TAHITI-NEXT: s_mov_b32 s4, s6 +; TAHITI-NEXT: s_mov_b32 s5, s7 +; TAHITI-NEXT: s_mov_b32 s6, s10 +; TAHITI-NEXT: s_mov_b32 s7, s11 +; TAHITI-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0 +; TAHITI-NEXT: buffer_load_dwordx2 v[2:3], off, s[16:19], 0 +; TAHITI-NEXT: s_mov_b32 s14, s10 +; TAHITI-NEXT: s_mov_b32 s12, s2 +; TAHITI-NEXT: s_mov_b32 s13, s3 +; TAHITI-NEXT: s_mov_b32 s15, s11 +; TAHITI-NEXT: buffer_load_dwordx2 v[4:5], off, s[12:15], 0 +; TAHITI-NEXT: s_mov_b32 s8, s0 +; TAHITI-NEXT: s_mov_b32 s9, s1 +; TAHITI-NEXT: s_waitcnt vmcnt(2) +; TAHITI-NEXT: v_add_f64 v[6:7], -v[0:1], 1.0 +; TAHITI-NEXT: s_waitcnt vmcnt(1) +; TAHITI-NEXT: v_mul_f64 v[2:3], v[2:3], v[6:7] +; TAHITI-NEXT: s_waitcnt vmcnt(0) +; TAHITI-NEXT: v_fma_f64 v[0:1], v[4:5], v[0:1], v[2:3] +; TAHITI-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 +; TAHITI-NEXT: s_endpgm +; +; VERDE-LABEL: test_f64_interp: +; VERDE: ; %bb.0: +; VERDE-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9 +; VERDE-NEXT: s_mov_b32 s11, 0xf000 +; VERDE-NEXT: s_mov_b32 s10, -1 +; VERDE-NEXT: s_mov_b32 s14, s10 +; VERDE-NEXT: s_mov_b32 s15, s11 +; VERDE-NEXT: s_waitcnt lgkmcnt(0) +; VERDE-NEXT: s_mov_b32 s12, s6 +; VERDE-NEXT: s_mov_b32 s13, s7 +; VERDE-NEXT: buffer_load_dwordx2 v[0:1], off, s[12:15], 0 +; VERDE-NEXT: s_mov_b32 s6, s10 +; VERDE-NEXT: s_mov_b32 s7, s11 +; VERDE-NEXT: buffer_load_dwordx2 v[2:3], off, s[4:7], 0 +; VERDE-NEXT: s_mov_b32 s4, s2 +; VERDE-NEXT: s_mov_b32 s5, s3 +; VERDE-NEXT: buffer_load_dwordx2 v[4:5], off, s[4:7], 0 +; VERDE-NEXT: s_mov_b32 s8, s0 +; VERDE-NEXT: s_mov_b32 s9, s1 +; VERDE-NEXT: s_waitcnt vmcnt(2) +; VERDE-NEXT: v_add_f64 v[6:7], -v[0:1], 1.0 +; VERDE-NEXT: s_waitcnt vmcnt(1) +; VERDE-NEXT: v_mul_f64 v[2:3], v[2:3], v[6:7] +; VERDE-NEXT: s_waitcnt vmcnt(0) +; VERDE-NEXT: v_fma_f64 v[0:1], v[4:5], v[0:1], v[2:3] +; VERDE-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 +; VERDE-NEXT: s_endpgm +; ; SI-FMA-LABEL: test_f64_interp: ; SI-FMA: ; %bb.0: ; SI-FMA-NEXT: s_load_dwordx8 s[0:7], s[4:5], 0x9 @@ -2137,8 +2534,8 @@ define amdgpu_kernel void @test_f64_interp(ptr addrspace(1) %out, ; SI-FMA-NEXT: s_mov_b32 s5, s7 ; SI-FMA-NEXT: s_mov_b32 s6, s10 ; SI-FMA-NEXT: s_mov_b32 s7, s11 -; SI-FMA-NEXT: buffer_load_dwordx2 v[0:1], off, s[16:19], 0 -; SI-FMA-NEXT: buffer_load_dwordx2 v[2:3], off, s[4:7], 0 +; SI-FMA-NEXT: buffer_load_dwordx2 v[0:1], off, s[4:7], 0 +; SI-FMA-NEXT: buffer_load_dwordx2 v[2:3], off, s[16:19], 0 ; SI-FMA-NEXT: s_mov_b32 s14, s10 ; SI-FMA-NEXT: s_mov_b32 s12, s2 ; SI-FMA-NEXT: s_mov_b32 s13, s3 @@ -2146,48 +2543,33 @@ define amdgpu_kernel void @test_f64_interp(ptr addrspace(1) %out, ; SI-FMA-NEXT: buffer_load_dwordx2 v[4:5], off, s[12:15], 0 ; SI-FMA-NEXT: s_mov_b32 s8, s0 ; SI-FMA-NEXT: s_mov_b32 s9, s1 +; SI-FMA-NEXT: s_waitcnt vmcnt(2) +; SI-FMA-NEXT: v_add_f64 v[6:7], -v[0:1], 1.0 ; SI-FMA-NEXT: s_waitcnt vmcnt(1) -; SI-FMA-NEXT: v_fma_f64 v[0:1], -v[2:3], v[0:1], v[0:1] +; SI-FMA-NEXT: v_mul_f64 v[2:3], v[2:3], v[6:7] ; SI-FMA-NEXT: s_waitcnt vmcnt(0) -; SI-FMA-NEXT: v_fma_f64 v[0:1], v[4:5], v[2:3], v[0:1] +; SI-FMA-NEXT: v_fma_f64 v[0:1], v[4:5], v[0:1], v[2:3] ; SI-FMA-NEXT: buffer_store_dwordx2 v[0:1], off, s[8:11], 0 ; SI-FMA-NEXT: s_endpgm ; -; GFX11-NOFMA-LABEL: test_f64_interp: -; GFX11-NOFMA: ; %bb.0: -; GFX11-NOFMA-NEXT: s_load_b256 s[0:7], s[4:5], 0x24 -; GFX11-NOFMA-NEXT: v_mov_b32_e32 v8, 0 -; GFX11-NOFMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NOFMA-NEXT: s_clause 0x2 -; GFX11-NOFMA-NEXT: global_load_b64 v[0:1], v8, s[6:7] -; GFX11-NOFMA-NEXT: global_load_b64 v[2:3], v8, s[4:5] -; GFX11-NOFMA-NEXT: global_load_b64 v[4:5], v8, s[2:3] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(2) -; GFX11-NOFMA-NEXT: v_add_f64 v[6:7], -v[0:1], 1.0 -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-NOFMA-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-NOFMA-NEXT: v_mul_f64 v[2:3], v[2:3], v[6:7] -; GFX11-NOFMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-NOFMA-NEXT: v_fma_f64 v[0:1], v[4:5], v[0:1], v[2:3] -; GFX11-NOFMA-NEXT: global_store_b64 v8, v[0:1], s[0:1] -; GFX11-NOFMA-NEXT: s_endpgm -; -; GFX11-FMA-LABEL: test_f64_interp: -; GFX11-FMA: ; %bb.0: -; GFX11-FMA-NEXT: s_load_b256 s[0:7], s[4:5], 0x24 -; GFX11-FMA-NEXT: v_mov_b32_e32 v6, 0 -; GFX11-FMA-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FMA-NEXT: s_clause 0x2 -; GFX11-FMA-NEXT: global_load_b64 v[0:1], v6, s[4:5] -; GFX11-FMA-NEXT: global_load_b64 v[2:3], v6, s[6:7] -; GFX11-FMA-NEXT: global_load_b64 v[4:5], v6, s[2:3] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(1) -; GFX11-FMA-NEXT: v_fma_f64 v[0:1], -v[2:3], v[0:1], v[0:1] -; GFX11-FMA-NEXT: s_waitcnt vmcnt(0) -; GFX11-FMA-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FMA-NEXT: v_fma_f64 v[0:1], v[4:5], v[2:3], v[0:1] -; GFX11-FMA-NEXT: global_store_b64 v6, v[0:1], s[0:1] -; GFX11-FMA-NEXT: s_endpgm +; GFX11-LABEL: test_f64_interp: +; GFX11: ; %bb.0: +; GFX11-NEXT: s_load_b256 s[0:7], s[4:5], 0x24 +; GFX11-NEXT: v_mov_b32_e32 v8, 0 +; GFX11-NEXT: s_waitcnt lgkmcnt(0) +; GFX11-NEXT: s_clause 0x2 +; GFX11-NEXT: global_load_b64 v[0:1], v8, s[6:7] +; GFX11-NEXT: global_load_b64 v[2:3], v8, s[4:5] +; GFX11-NEXT: global_load_b64 v[4:5], v8, s[2:3] +; GFX11-NEXT: s_waitcnt vmcnt(2) +; GFX11-NEXT: v_add_f64 v[6:7], -v[0:1], 1.0 +; GFX11-NEXT: s_waitcnt vmcnt(1) +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX11-NEXT: v_mul_f64 v[2:3], v[2:3], v[6:7] +; GFX11-NEXT: s_waitcnt vmcnt(0) +; GFX11-NEXT: v_fma_f64 v[0:1], v[4:5], v[0:1], v[2:3] +; GFX11-NEXT: global_store_b64 v8, v[0:1], s[0:1] +; GFX11-NEXT: s_endpgm ptr addrspace(1) %in1, ptr addrspace(1) %in2, ptr addrspace(1) %in3) { @@ -2356,3 +2738,6 @@ define amdgpu_kernel void @fma_neg_b_c_v4f32(ptr addrspace(1) %out, ptr addrspac attributes #0 = { nounwind readnone } attributes #1 = { nounwind } attributes #2 = { nounwind "no-signed-zeros-fp-math"="true" } +;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line: +; GFX11-FMA: {{.*}} +; GFX11-NOFMA: {{.*}} diff --git a/llvm/test/CodeGen/AMDGPU/fmed3.ll b/llvm/test/CodeGen/AMDGPU/fmed3.ll index 9233f8059a20..9e152253bb6c 100644 --- a/llvm/test/CodeGen/AMDGPU/fmed3.ll +++ b/llvm/test/CodeGen/AMDGPU/fmed3.ll @@ -7464,18 +7464,15 @@ define amdgpu_kernel void @v_test_nnan_input_fmed3_r_i_i_f16(ptr addrspace(1) %o ; SI-GISEL-NEXT: s_waitcnt lgkmcnt(0) ; SI-GISEL-NEXT: s_mov_b64 s[4:5], s[2:3] ; SI-GISEL-NEXT: buffer_load_ushort v2, v[0:1], s[4:7], 0 addr64 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v3, 1.0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v4, 2.0 ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v2 -; SI-GISEL-NEXT: v_add_f32_e32 v2, v2, v3 +; SI-GISEL-NEXT: v_add_f32_e32 v2, 1.0, v2 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v2, v2 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v2 -; SI-GISEL-NEXT: v_max_f32_e32 v2, v2, v4 +; SI-GISEL-NEXT: v_max_f32_e32 v2, 2.0, v2 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v2, v2 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v2 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v3, 4.0 -; SI-GISEL-NEXT: v_min_f32_e32 v2, v2, v3 +; SI-GISEL-NEXT: v_min_f32_e32 v2, 4.0, v2 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v2, v2 ; SI-GISEL-NEXT: s_mov_b64 s[2:3], s[6:7] ; SI-GISEL-NEXT: buffer_store_short v2, v[0:1], s[0:3], 0 addr64 @@ -7639,27 +7636,24 @@ define amdgpu_kernel void @v_nnan_inputs_med3_f16_pat0(ptr addrspace(1) %out, pt ; SI-GISEL-NEXT: v_mov_b32_e32 v1, 0 ; SI-GISEL-NEXT: s_mov_b32 s10, 0 ; SI-GISEL-NEXT: s_mov_b32 s11, 0xf000 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, 1.0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v3, 2.0 ; SI-GISEL-NEXT: s_waitcnt lgkmcnt(0) ; SI-GISEL-NEXT: s_mov_b64 s[8:9], s[2:3] -; SI-GISEL-NEXT: buffer_load_ushort v4, v[0:1], s[8:11], 0 addr64 glc +; SI-GISEL-NEXT: buffer_load_ushort v2, v[0:1], s[8:11], 0 addr64 glc ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v5, 4.0 ; SI-GISEL-NEXT: s_mov_b64 s[8:9], s[4:5] -; SI-GISEL-NEXT: buffer_load_ushort v6, v[0:1], s[8:11], 0 addr64 glc +; SI-GISEL-NEXT: buffer_load_ushort v3, v[0:1], s[8:11], 0 addr64 glc ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) ; SI-GISEL-NEXT: s_mov_b64 s[8:9], s[6:7] -; SI-GISEL-NEXT: buffer_load_ushort v7, v[0:1], s[8:11], 0 addr64 glc +; SI-GISEL-NEXT: buffer_load_ushort v4, v[0:1], s[8:11], 0 addr64 glc ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v4, v4 -; SI-GISEL-NEXT: v_add_f32_e32 v2, v4, v2 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v4, v6 +; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v2 +; SI-GISEL-NEXT: v_add_f32_e32 v2, 1.0, v2 +; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v3, v3 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v2, v2 -; SI-GISEL-NEXT: v_add_f32_e32 v3, v4, v3 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v4, v7 +; SI-GISEL-NEXT: v_add_f32_e32 v3, 2.0, v3 +; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v4, v4 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v3, v3 -; SI-GISEL-NEXT: v_add_f32_e32 v4, v4, v5 +; SI-GISEL-NEXT: v_add_f32_e32 v4, 4.0, v4 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v2 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v4, v4 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v3, v3 @@ -8712,12 +8706,10 @@ define half @v_test_fmed3_r_i_i_f16_minimumnum_maximumnum(half %a) #1 { ; SI-GISEL: ; %bb.0: ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, 2.0 -; SI-GISEL-NEXT: v_max_f32_e32 v0, v0, v1 +; SI-GISEL-NEXT: v_max_f32_e32 v0, 2.0, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, 4.0 -; SI-GISEL-NEXT: v_min_f32_e32 v0, v0, v1 +; SI-GISEL-NEXT: v_min_f32_e32 v0, 4.0, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -8796,17 +8788,15 @@ define <2 x half> @v_test_fmed3_r_i_i_v2f16_minimumnum_maximumnum(<2 x half> %a) ; SI-GISEL: ; %bb.0: ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, 2.0 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, v1 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v3, 4.0 -; SI-GISEL-NEXT: v_max_f32_e32 v0, v0, v2 -; SI-GISEL-NEXT: v_max_f32_e32 v1, v1, v2 +; SI-GISEL-NEXT: v_max_f32_e32 v0, 2.0, v0 +; SI-GISEL-NEXT: v_max_f32_e32 v1, 2.0, v1 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v1, v1 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, v1 -; SI-GISEL-NEXT: v_min_f32_e32 v0, v0, v3 -; SI-GISEL-NEXT: v_min_f32_e32 v1, v1, v3 +; SI-GISEL-NEXT: v_min_f32_e32 v0, 4.0, v0 +; SI-GISEL-NEXT: v_min_f32_e32 v1, 4.0, v1 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v1, v1 ; SI-GISEL-NEXT: s_setpc_b64 s[30:31] diff --git a/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll b/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll index 64a9727330cf..76da0aaf251b 100644 --- a/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/fneg-fabs.bf16.ll @@ -107,12 +107,10 @@ define amdgpu_kernel void @fneg_fabs_fmul_bf16(ptr addrspace(1) %out, bfloat %x, ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_and_b32 s3, s2, 0x7fff -; CI-NEXT: s_lshl_b32 s3, s3, 16 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s3 +; CI-NEXT: s_lshl_b32 s3, s2, 16 ; CI-NEXT: s_and_b32 s2, s2, 0xffff0000 -; CI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; CI-NEXT: v_mul_f32_e32 v0, s2, v0 +; CI-NEXT: v_mov_b32_e32 v0, s3 +; CI-NEXT: v_mul_f32_e64 v0, s2, -|v0| ; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v0 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 @@ -204,12 +202,10 @@ define amdgpu_kernel void @fneg_fabs_free_bf16(ptr addrspace(1) %out, i16 %in) { ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_and_b32 s2, s2, 0x7fff -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s2 -; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; CI-NEXT: s_bitset1_b32 s2, 15 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: v_mov_b32_e32 v2, s2 ; CI-NEXT: flat_store_short v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -279,12 +275,10 @@ define amdgpu_kernel void @fneg_fabs_bf16(ptr addrspace(1) %out, bfloat %in) { ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_and_b32 s2, s2, 0x7fff -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s2 -; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; CI-NEXT: s_bitset1_b32 s2, 15 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: v_mov_b32_e32 v2, s2 ; CI-NEXT: flat_store_short v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -345,43 +339,22 @@ define amdgpu_kernel void @fneg_fabs_bf16(ptr addrspace(1) %out, bfloat %in) { } define amdgpu_kernel void @v_fneg_fabs_bf16(ptr addrspace(1) %out, ptr addrspace(1) %in) { -; CI-LABEL: v_fneg_fabs_bf16: -; CI: ; %bb.0: -; CI-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 -; CI-NEXT: s_add_i32 s12, s12, s17 -; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 -; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 -; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: v_mov_b32_e32 v0, s2 -; CI-NEXT: v_mov_b32_e32 v1, s3 -; CI-NEXT: flat_load_ushort v2, v[0:1] -; CI-NEXT: v_mov_b32_e32 v0, s0 -; CI-NEXT: v_mov_b32_e32 v1, s1 -; CI-NEXT: s_waitcnt vmcnt(0) -; CI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; CI-NEXT: v_mul_f32_e64 v2, 1.0, |v2| -; CI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 -; CI-NEXT: v_xor_b32_e32 v2, 0x80000000, v2 -; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; CI-NEXT: flat_store_short v[0:1], v2 -; CI-NEXT: s_endpgm -; -; VI-LABEL: v_fneg_fabs_bf16: -; VI: ; %bb.0: -; VI-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 -; VI-NEXT: s_add_i32 s12, s12, s17 -; VI-NEXT: s_mov_b32 flat_scratch_lo, s13 -; VI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 -; VI-NEXT: s_waitcnt lgkmcnt(0) -; VI-NEXT: v_mov_b32_e32 v0, s2 -; VI-NEXT: v_mov_b32_e32 v1, s3 -; VI-NEXT: flat_load_ushort v2, v[0:1] -; VI-NEXT: v_mov_b32_e32 v0, s0 -; VI-NEXT: v_mov_b32_e32 v1, s1 -; VI-NEXT: s_waitcnt vmcnt(0) -; VI-NEXT: v_or_b32_e32 v2, 0x8000, v2 -; VI-NEXT: flat_store_short v[0:1], v2 -; VI-NEXT: s_endpgm +; CIVI-LABEL: v_fneg_fabs_bf16: +; CIVI: ; %bb.0: +; CIVI-NEXT: s_load_dwordx4 s[0:3], s[8:9], 0x0 +; CIVI-NEXT: s_add_i32 s12, s12, s17 +; CIVI-NEXT: s_mov_b32 flat_scratch_lo, s13 +; CIVI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 +; CIVI-NEXT: s_waitcnt lgkmcnt(0) +; CIVI-NEXT: v_mov_b32_e32 v0, s2 +; CIVI-NEXT: v_mov_b32_e32 v1, s3 +; CIVI-NEXT: flat_load_ushort v2, v[0:1] +; CIVI-NEXT: v_mov_b32_e32 v0, s0 +; CIVI-NEXT: v_mov_b32_e32 v1, s1 +; CIVI-NEXT: s_waitcnt vmcnt(0) +; CIVI-NEXT: v_or_b32_e32 v2, 0x8000, v2 +; CIVI-NEXT: flat_store_short v[0:1], v2 +; CIVI-NEXT: s_endpgm ; ; GFX9-LABEL: v_fneg_fabs_bf16: ; GFX9: ; %bb.0: @@ -431,21 +404,13 @@ define amdgpu_kernel void @s_fneg_fabs_v2bf16_non_bc_src(ptr addrspace(1) %out, ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_and_b32 s3, s2, 0xffff0000 -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_add_f32_e64 v0, s3, 2.0 -; CI-NEXT: v_add_f32_e64 v1, s2, 1.0 -; CI-NEXT: v_readfirstlane_b32 s2, v0 +; CI-NEXT: s_lshl_b32 s3, s2, 16 ; CI-NEXT: s_and_b32 s2, s2, 0xffff0000 -; CI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 -; CI-NEXT: s_bitset0_b32 s2, 31 -; CI-NEXT: v_and_b32_e32 v0, 0x7fffffff, v1 -; CI-NEXT: s_and_b32 s2, s2, 0xffff0000 -; CI-NEXT: s_xor_b32 s2, s2, 0x80000000 -; CI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; CI-NEXT: s_lshr_b32 s2, s2, 16 -; CI-NEXT: v_xor_b32_e32 v0, 0x80000000, v0 -; CI-NEXT: v_alignbit_b32 v2, s2, v0, 16 +; CI-NEXT: v_add_f32_e64 v1, s2, 2.0 +; CI-NEXT: v_add_f32_e64 v0, s3, 1.0 +; CI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; CI-NEXT: v_alignbit_b32 v0, v1, v0, 16 +; CI-NEXT: v_or_b32_e32 v2, 0x80008000, v0 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 ; CI-NEXT: flat_store_dword v[0:1], v2 @@ -566,15 +531,10 @@ define amdgpu_kernel void @s_fneg_fabs_v2bf16_bc_src(ptr addrspace(1) %out, <2 x ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_and_b32 s3, s2, 0x7fff -; CI-NEXT: s_and_b32 s2, s2, 0x7fff0000 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s2 -; CI-NEXT: s_lshl_b32 s2, s3, 16 -; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; CI-NEXT: v_mul_f32_e64 v1, -1.0, s2 -; CI-NEXT: v_alignbit_b32 v2, v0, v1, 16 +; CI-NEXT: s_or_b32 s2, s2, 0x80008000 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: v_mov_b32_e32 v2, s2 ; CI-NEXT: flat_store_dword v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -629,27 +589,11 @@ define amdgpu_kernel void @fneg_fabs_v4bf16(ptr addrspace(1) %out, <4 x bfloat> ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_lshl_b32 s4, s2, 16 -; CI-NEXT: s_and_b32 s2, s2, 0xffff0000 -; CI-NEXT: v_mul_f32_e64 v2, 1.0, |s2| -; CI-NEXT: s_and_b32 s2, s3, 0xffff0000 -; CI-NEXT: s_lshl_b32 s5, s3, 16 -; CI-NEXT: v_mul_f32_e64 v3, 1.0, |s2| -; CI-NEXT: v_mul_f32_e64 v0, 1.0, |s4| -; CI-NEXT: v_mul_f32_e64 v1, 1.0, |s5| -; CI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 -; CI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 -; CI-NEXT: v_xor_b32_e32 v3, 0x80000000, v3 -; CI-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 -; CI-NEXT: v_xor_b32_e32 v2, 0x80000000, v2 -; CI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; CI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 -; CI-NEXT: v_xor_b32_e32 v1, 0x80000000, v1 -; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; CI-NEXT: v_xor_b32_e32 v0, 0x80000000, v0 -; CI-NEXT: v_alignbit_b32 v1, v3, v1, 16 -; CI-NEXT: v_alignbit_b32 v0, v2, v0, 16 +; CI-NEXT: s_or_b32 s3, s3, 0x80008000 +; CI-NEXT: s_or_b32 s2, s2, 0x80008000 ; CI-NEXT: v_mov_b32_e32 v3, s1 +; CI-NEXT: v_mov_b32_e32 v0, s2 +; CI-NEXT: v_mov_b32_e32 v1, s3 ; CI-NEXT: v_mov_b32_e32 v2, s0 ; CI-NEXT: flat_store_dwordx2 v[2:3], v[0:1] ; CI-NEXT: s_endpgm @@ -860,21 +804,20 @@ define amdgpu_kernel void @s_fneg_multi_use_fabs_v2bf16(ptr addrspace(1) %out0, ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: v_mov_b32_e32 v1, s1 -; CI-NEXT: v_mov_b32_e32 v2, s2 -; CI-NEXT: s_and_b32 s1, s4, 0x7fff -; CI-NEXT: s_and_b32 s2, s4, 0x7fff0000 -; CI-NEXT: v_mul_f32_e64 v4, -1.0, s2 -; CI-NEXT: s_lshl_b32 s1, s1, 16 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: s_and_b32 s0, s4, 0x7fff7fff -; CI-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; CI-NEXT: v_mul_f32_e64 v5, -1.0, s1 -; CI-NEXT: v_alignbit_b32 v4, v4, v5, 16 -; CI-NEXT: v_mov_b32_e32 v5, s0 +; CI-NEXT: v_mov_b32_e32 v2, s2 +; CI-NEXT: s_or_b32 s2, s0, 0x8000 +; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: s_and_b32 s1, s4, 0x7fff0000 +; CI-NEXT: s_and_b32 s2, s2, 0xffff +; CI-NEXT: s_or_b32 s1, s1, s2 +; CI-NEXT: s_bitset1_b32 s1, 31 +; CI-NEXT: v_mov_b32_e32 v4, s0 ; CI-NEXT: v_mov_b32_e32 v3, s3 -; CI-NEXT: flat_store_dword v[0:1], v5 -; CI-NEXT: flat_store_dword v[2:3], v4 +; CI-NEXT: flat_store_dword v[0:1], v4 +; CI-NEXT: v_mov_b32_e32 v0, s1 +; CI-NEXT: flat_store_dword v[2:3], v0 ; CI-NEXT: s_endpgm ; ; VI-LABEL: s_fneg_multi_use_fabs_v2bf16: @@ -1086,5 +1029,3 @@ declare <4 x bfloat> @llvm.fabs.v4bf16(<4 x bfloat>) #1 attributes #0 = { nounwind } attributes #1 = { nounwind readnone } -;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line: -; CIVI: {{.*}} diff --git a/llvm/test/CodeGen/AMDGPU/fneg.bf16.ll b/llvm/test/CodeGen/AMDGPU/fneg.bf16.ll index d232693b46ad..98044a72870f 100644 --- a/llvm/test/CodeGen/AMDGPU/fneg.bf16.ll +++ b/llvm/test/CodeGen/AMDGPU/fneg.bf16.ll @@ -14,11 +14,10 @@ define amdgpu_kernel void @s_fneg_bf16(ptr addrspace(1) %out, bfloat %in) #0 { ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s2 -; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; CI-NEXT: s_xor_b32 s2, s2, 0x8000 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: v_mov_b32_e32 v2, s2 ; CI-NEXT: flat_store_short v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -93,9 +92,7 @@ define amdgpu_kernel void @v_fneg_bf16(ptr addrspace(1) %out, ptr addrspace(1) % ; CI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc ; CI-NEXT: flat_load_ushort v2, v[0:1] ; CI-NEXT: s_waitcnt vmcnt(0) -; CI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; CI-NEXT: v_mul_f32_e32 v2, -1.0, v2 -; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 +; CI-NEXT: v_xor_b32_e32 v2, 0x8000, v2 ; CI-NEXT: flat_store_short v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -170,11 +167,10 @@ define amdgpu_kernel void @s_fneg_free_bf16(ptr addrspace(1) %out, i16 %in) #0 { ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s2 -; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v0 +; CI-NEXT: s_xor_b32 s2, s2, 0x8000 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: v_mov_b32_e32 v2, s2 ; CI-NEXT: flat_store_short v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -248,9 +244,9 @@ define amdgpu_kernel void @v_fneg_fold_bf16(ptr addrspace(1) %out, ptr addrspace ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 ; CI-NEXT: s_waitcnt vmcnt(0) +; CI-NEXT: v_xor_b32_e32 v3, 0x8000, v2 ; CI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; CI-NEXT: v_mul_f32_e32 v3, -1.0, v2 -; CI-NEXT: v_and_b32_e32 v3, 0xffff0000, v3 +; CI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 ; CI-NEXT: v_mul_f32_e32 v2, v3, v2 ; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 ; CI-NEXT: flat_store_short v[0:1], v2 @@ -365,13 +361,13 @@ define amdgpu_kernel void @s_fneg_v2bf16(ptr addrspace(1) %out, <2 x bfloat> %in ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) ; CI-NEXT: s_and_b32 s3, s2, 0xffff0000 -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s3 -; CI-NEXT: v_mul_f32_e64 v1, -1.0, s2 -; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; CI-NEXT: v_alignbit_b32 v2, v0, v1, 16 +; CI-NEXT: s_xor_b32 s2, s2, 0x8000 +; CI-NEXT: s_and_b32 s2, s2, 0xffff +; CI-NEXT: s_or_b32 s2, s2, s3 +; CI-NEXT: s_add_i32 s2, s2, 0x80000000 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: v_mov_b32_e32 v2, s2 ; CI-NEXT: flat_store_dword v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -426,16 +422,16 @@ define amdgpu_kernel void @s_fneg_v2bf16_nonload(ptr addrspace(1) %out) #0 { ; CI-NEXT: ; def s2 ; CI-NEXT: ;;#ASMEND ; CI-NEXT: s_and_b32 s3, s2, 0xffff0000 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s3 -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; CI-NEXT: v_mul_f32_e64 v1, -1.0, s2 -; CI-NEXT: v_alignbit_b32 v2, v0, v1, 16 +; CI-NEXT: s_xor_b32 s2, s2, 0x8000 +; CI-NEXT: s_and_b32 s2, s2, 0xffff +; CI-NEXT: s_or_b32 s2, s2, s3 +; CI-NEXT: s_add_i32 s2, s2, 0x80000000 ; CI-NEXT: s_waitcnt lgkmcnt(0) ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: s_mov_b32 flat_scratch_lo, s13 ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: v_mov_b32_e32 v2, s2 ; CI-NEXT: flat_store_dword v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -501,13 +497,11 @@ define amdgpu_kernel void @v_fneg_v2bf16(ptr addrspace(1) %out, ptr addrspace(1) ; CI-NEXT: v_add_i32_e32 v0, vcc, s0, v0 ; CI-NEXT: v_addc_u32_e32 v1, vcc, 0, v1, vcc ; CI-NEXT: flat_load_dword v2, v[0:1] +; CI-NEXT: s_mov_b32 s0, 0xffff ; CI-NEXT: s_waitcnt vmcnt(0) -; CI-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 -; CI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; CI-NEXT: v_mul_f32_e32 v3, -1.0, v3 -; CI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 -; CI-NEXT: v_mul_f32_e32 v2, -1.0, v2 -; CI-NEXT: v_alignbit_b32 v2, v3, v2, 16 +; CI-NEXT: v_xor_b32_e32 v3, 0x8000, v2 +; CI-NEXT: v_bfi_b32 v2, s0, v3, v2 +; CI-NEXT: v_add_i32_e32 v2, vcc, 0x80000000, v2 ; CI-NEXT: flat_store_dword v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -570,13 +564,13 @@ define amdgpu_kernel void @fneg_free_v2bf16(ptr addrspace(1) %out, i32 %in) #0 { ; CI-NEXT: s_lshr_b32 flat_scratch_hi, s12, 8 ; CI-NEXT: s_waitcnt lgkmcnt(0) ; CI-NEXT: s_and_b32 s3, s2, 0xffff0000 -; CI-NEXT: s_lshl_b32 s2, s2, 16 -; CI-NEXT: v_mul_f32_e64 v0, -1.0, s3 -; CI-NEXT: v_mul_f32_e64 v1, -1.0, s2 -; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 -; CI-NEXT: v_alignbit_b32 v2, v0, v1, 16 +; CI-NEXT: s_xor_b32 s2, s2, 0x8000 +; CI-NEXT: s_and_b32 s2, s2, 0xffff +; CI-NEXT: s_or_b32 s2, s2, s3 +; CI-NEXT: s_add_i32 s2, s2, 0x80000000 ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 +; CI-NEXT: v_mov_b32_e32 v2, s2 ; CI-NEXT: flat_store_dword v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -637,16 +631,14 @@ define amdgpu_kernel void @v_fneg_fold_v2bf16(ptr addrspace(1) %out, ptr addrspa ; CI-NEXT: v_mov_b32_e32 v0, s0 ; CI-NEXT: v_mov_b32_e32 v1, s1 ; CI-NEXT: s_waitcnt vmcnt(0) -; CI-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 -; CI-NEXT: v_lshlrev_b32_e32 v2, 16, v2 -; CI-NEXT: v_mul_f32_e32 v4, -1.0, v3 -; CI-NEXT: v_mul_f32_e32 v5, -1.0, v2 -; CI-NEXT: v_and_b32_e32 v4, 0xffff0000, v4 -; CI-NEXT: v_and_b32_e32 v5, 0xffff0000, v5 -; CI-NEXT: v_mul_f32_e32 v3, v4, v3 -; CI-NEXT: v_mul_f32_e32 v2, v5, v2 -; CI-NEXT: v_lshrrev_b32_e32 v3, 16, v3 -; CI-NEXT: v_alignbit_b32 v2, v3, v2, 16 +; CI-NEXT: v_xor_b32_e32 v3, 0x8000, v2 +; CI-NEXT: v_lshlrev_b32_e32 v4, 16, v2 +; CI-NEXT: v_and_b32_e32 v2, 0xffff0000, v2 +; CI-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; CI-NEXT: v_mul_f32_e64 v2, -v2, v2 +; CI-NEXT: v_mul_f32_e32 v3, v3, v4 +; CI-NEXT: v_lshrrev_b32_e32 v2, 16, v2 +; CI-NEXT: v_alignbit_b32 v2, v2, v3, 16 ; CI-NEXT: flat_store_dword v[0:1], v2 ; CI-NEXT: s_endpgm ; @@ -912,12 +904,9 @@ define amdgpu_kernel void @v_extract_fneg_no_fold_v2bf16(ptr addrspace(1) %in) # ; CI-NEXT: v_mov_b32_e32 v1, s1 ; CI-NEXT: flat_load_dword v0, v[0:1] ; CI-NEXT: s_waitcnt vmcnt(0) -; CI-NEXT: v_and_b32_e32 v1, 0xffff0000, v0 -; CI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 -; CI-NEXT: v_mul_f32_e32 v1, -1.0, v1 -; CI-NEXT: v_mul_f32_e32 v0, -1.0, v0 -; CI-NEXT: v_lshrrev_b32_e32 v1, 16, v1 -; CI-NEXT: v_lshrrev_b32_e32 v0, 16, v0 +; CI-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; CI-NEXT: v_xor_b32_e32 v0, 0x8000, v0 +; CI-NEXT: v_xor_b32_e32 v1, 0x8000, v1 ; CI-NEXT: flat_store_short v[0:1], v0 ; CI-NEXT: s_waitcnt vmcnt(0) ; CI-NEXT: flat_store_short v[0:1], v1 diff --git a/llvm/test/CodeGen/AMDGPU/fptoui_uitofp.ll b/llvm/test/CodeGen/AMDGPU/fptoui_uitofp.ll new file mode 100644 index 000000000000..49204f84acb8 --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/fptoui_uitofp.ll @@ -0,0 +1,296 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 +; RUN: llc -mtriple=amdgcn -mcpu=gfx600 < %s | FileCheck -check-prefix=GFX6 %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx900 < %s | FileCheck -check-prefix=GFX9 %s + +define amdgpu_kernel void @fptoui_f32_to_i16_to_f32(ptr addrspace(1) %out, float %x) { +; GFX6-LABEL: fptoui_f32_to_i16_to_f32: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dword s6, s[4:5], 0xb +; GFX6-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; GFX6-NEXT: s_mov_b32 s3, 0xf000 +; GFX6-NEXT: s_mov_b32 s2, -1 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: v_trunc_f32_e64 v0, |s6| +; GFX6-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f32_to_i16_to_f32: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dword s2, s[4:5], 0x2c +; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v0, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f32_e64 v1, |s2| +; GFX9-NEXT: global_store_dword v0, v1, s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui float %x to i16 + %fp = uitofp i16 %ui to float + store float %fp, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptoui_f32_to_i32_to_f32(ptr addrspace(1) %out, float %x) { +; GFX6-LABEL: fptoui_f32_to_i32_to_f32: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dword s6, s[4:5], 0xb +; GFX6-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; GFX6-NEXT: s_mov_b32 s3, 0xf000 +; GFX6-NEXT: s_mov_b32 s2, -1 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: v_trunc_f32_e64 v0, |s6| +; GFX6-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f32_to_i32_to_f32: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dword s2, s[4:5], 0x2c +; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v0, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f32_e64 v1, |s2| +; GFX9-NEXT: global_store_dword v0, v1, s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui float %x to i32 + %fp = uitofp i32 %ui to float + store float %fp, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptoui_f32_to_i64_to_f32(ptr addrspace(1) %out, float %x) { +; GFX6-LABEL: fptoui_f32_to_i64_to_f32: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dword s6, s[4:5], 0xb +; GFX6-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; GFX6-NEXT: s_mov_b32 s3, 0xf000 +; GFX6-NEXT: s_mov_b32 s2, -1 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: v_trunc_f32_e64 v0, |s6| +; GFX6-NEXT: buffer_store_dword v0, off, s[0:3], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f32_to_i64_to_f32: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dword s2, s[4:5], 0x2c +; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v0, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f32_e64 v1, |s2| +; GFX9-NEXT: global_store_dword v0, v1, s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui float %x to i64 + %fp = uitofp i64 %ui to float + store float %fp, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptoui_f16_to_i16_to_f16(ptr addrspace(1) %out, half %x) { +; GFX6-LABEL: fptoui_f16_to_i16_to_f16: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dword s0, s[4:5], 0xb +; GFX6-NEXT: s_mov_b32 s3, 0xf000 +; GFX6-NEXT: s_mov_b32 s2, -1 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: v_cvt_f32_f16_e32 v0, s0 +; GFX6-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; GFX6-NEXT: v_cvt_u32_f32_e32 v0, v0 +; GFX6-NEXT: v_cvt_f32_u32_e32 v0, v0 +; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f16_to_i16_to_f16: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dword s2, s[4:5], 0x2c +; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v0, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f16_e64 v1, |s2| +; GFX9-NEXT: global_store_short v0, v1, s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui half %x to i16 + %fp = uitofp i16 %ui to half + store half %fp, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptoui_f16_to_i32_to_f16(ptr addrspace(1) %out, half %x) { +; GFX6-LABEL: fptoui_f16_to_i32_to_f16: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dword s0, s[4:5], 0xb +; GFX6-NEXT: s_mov_b32 s3, 0xf000 +; GFX6-NEXT: s_mov_b32 s2, -1 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: v_cvt_f32_f16_e64 v0, |s0| +; GFX6-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; GFX6-NEXT: v_trunc_f32_e32 v0, v0 +; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f16_to_i32_to_f16: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dword s2, s[4:5], 0x2c +; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v0, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f16_e64 v1, |s2| +; GFX9-NEXT: global_store_short v0, v1, s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui half %x to i32 + %fp = uitofp i32 %ui to half + store half %fp, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptoui_f16_to_i64_to_f16(ptr addrspace(1) %out, half %x) { +; GFX6-LABEL: fptoui_f16_to_i64_to_f16: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dword s0, s[4:5], 0xb +; GFX6-NEXT: s_mov_b32 s3, 0xf000 +; GFX6-NEXT: s_mov_b32 s2, -1 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: v_cvt_f32_f16_e64 v0, |s0| +; GFX6-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x9 +; GFX6-NEXT: v_trunc_f32_e32 v0, v0 +; GFX6-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: buffer_store_short v0, off, s[0:3], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f16_to_i64_to_f16: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dword s2, s[4:5], 0x2c +; GFX9-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v0, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f16_e64 v1, |s2| +; GFX9-NEXT: global_store_short v0, v1, s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui half %x to i64 + %fp = uitofp i64 %ui to half + store half %fp, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptoui_f64_to_i16_to_f64(ptr addrspace(1) %out, double %x) { +; GFX6-LABEL: fptoui_f64_to_i16_to_f64: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: v_cvt_u32_f64_e32 v0, s[2:3] +; GFX6-NEXT: s_mov_b32 s3, 0xf000 +; GFX6-NEXT: s_mov_b32 s2, -1 +; GFX6-NEXT: v_cvt_f64_u32_e32 v[0:1], v0 +; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f64_to_i16_to_f64: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v2, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f64_e64 v[0:1], |s[2:3]| +; GFX9-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui double %x to i16 + %fp = uitofp i16 %ui to double + store double %fp, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptoui_f64_to_i32_to_f64(ptr addrspace(1) %out, double %x) { +; GFX6-LABEL: fptoui_f64_to_i32_to_f64: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: v_cvt_u32_f64_e32 v0, s[2:3] +; GFX6-NEXT: s_mov_b32 s3, 0xf000 +; GFX6-NEXT: s_mov_b32 s2, -1 +; GFX6-NEXT: v_cvt_f64_u32_e32 v[0:1], v0 +; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[0:3], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f64_to_i32_to_f64: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v2, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f64_e64 v[0:1], |s[2:3]| +; GFX9-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui double %x to i32 + %fp = uitofp i32 %ui to double + store double %fp, ptr addrspace(1) %out + ret void +} + +define amdgpu_kernel void @fptoui_f64_to_i64_to_f64(ptr addrspace(1) %out, double %x) { +; GFX6-LABEL: fptoui_f64_to_i64_to_f64: +; GFX6: ; %bb.0: ; %entry +; GFX6-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x9 +; GFX6-NEXT: s_mov_b32 s6, -1 +; GFX6-NEXT: s_mov_b32 s5, 0xfffff +; GFX6-NEXT: s_mov_b32 s4, s6 +; GFX6-NEXT: v_not_b32_e32 v0, 31 +; GFX6-NEXT: s_waitcnt lgkmcnt(0) +; GFX6-NEXT: s_bfe_u32 s7, s3, 0xb0014 +; GFX6-NEXT: s_addk_i32 s7, 0xfc01 +; GFX6-NEXT: s_lshr_b64 s[4:5], s[4:5], s7 +; GFX6-NEXT: s_and_b32 s8, s3, 0x80000000 +; GFX6-NEXT: s_andn2_b64 s[4:5], s[2:3], s[4:5] +; GFX6-NEXT: s_cmp_lt_i32 s7, 0 +; GFX6-NEXT: s_cselect_b32 s4, 0, s4 +; GFX6-NEXT: s_cselect_b32 s5, s8, s5 +; GFX6-NEXT: s_cmp_gt_i32 s7, 51 +; GFX6-NEXT: s_cselect_b32 s3, s3, s5 +; GFX6-NEXT: s_cselect_b32 s2, s2, s4 +; GFX6-NEXT: v_ldexp_f64 v[0:1], s[2:3], v0 +; GFX6-NEXT: v_mov_b32_e32 v4, -1 +; GFX6-NEXT: v_fract_f64_e32 v[2:3], v[0:1] +; GFX6-NEXT: v_mov_b32_e32 v5, 0x3fefffff +; GFX6-NEXT: v_min_f64 v[2:3], v[2:3], v[4:5] +; GFX6-NEXT: v_cmp_class_f64_e64 vcc, v[0:1], 3 +; GFX6-NEXT: s_mov_b32 s4, 0 +; GFX6-NEXT: v_cndmask_b32_e32 v2, v2, v0, vcc +; GFX6-NEXT: v_cndmask_b32_e32 v3, v3, v1, vcc +; GFX6-NEXT: v_add_f64 v[0:1], v[0:1], -v[2:3] +; GFX6-NEXT: v_mov_b32_e32 v2, s2 +; GFX6-NEXT: s_mov_b32 s5, 0xc1f00000 +; GFX6-NEXT: v_mov_b32_e32 v3, s3 +; GFX6-NEXT: v_fma_f64 v[2:3], v[0:1], s[4:5], v[2:3] +; GFX6-NEXT: v_cvt_u32_f64_e32 v0, v[0:1] +; GFX6-NEXT: v_cvt_u32_f64_e32 v2, v[2:3] +; GFX6-NEXT: s_mov_b32 s7, 0xf000 +; GFX6-NEXT: s_mov_b32 s4, s0 +; GFX6-NEXT: v_cvt_f64_u32_e32 v[0:1], v0 +; GFX6-NEXT: v_cvt_f64_u32_e32 v[2:3], v2 +; GFX6-NEXT: s_mov_b32 s5, s1 +; GFX6-NEXT: v_ldexp_f64 v[0:1], v[0:1], 32 +; GFX6-NEXT: v_add_f64 v[0:1], v[0:1], v[2:3] +; GFX6-NEXT: buffer_store_dwordx2 v[0:1], off, s[4:7], 0 +; GFX6-NEXT: s_endpgm +; +; GFX9-LABEL: fptoui_f64_to_i64_to_f64: +; GFX9: ; %bb.0: ; %entry +; GFX9-NEXT: s_load_dwordx4 s[0:3], s[4:5], 0x24 +; GFX9-NEXT: v_mov_b32_e32 v2, 0 +; GFX9-NEXT: s_waitcnt lgkmcnt(0) +; GFX9-NEXT: v_trunc_f64_e64 v[0:1], |s[2:3]| +; GFX9-NEXT: global_store_dwordx2 v2, v[0:1], s[0:1] +; GFX9-NEXT: s_endpgm +entry: + %ui = fptoui double %x to i64 + %fp = uitofp i64 %ui to double + store double %fp, ptr addrspace(1) %out + ret void +} diff --git a/llvm/test/CodeGen/AMDGPU/frem.ll b/llvm/test/CodeGen/AMDGPU/frem.ll index 78a961ea0da1..415828f32f92 100644 --- a/llvm/test/CodeGen/AMDGPU/frem.ll +++ b/llvm/test/CodeGen/AMDGPU/frem.ll @@ -4858,7 +4858,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_and_b64 vcc, exec, s[2:3] ; SI-NEXT: v_cvt_f16_f32_e32 v4, v2 ; SI-NEXT: s_cbranch_vccz .LBB9_2 -; SI-NEXT: ; %bb.1: ; %frem.else +; SI-NEXT: ; %bb.1: ; %frem.else20 ; SI-NEXT: v_bfi_b32 v7, s0, 0, v2 ; SI-NEXT: v_cvt_f32_f16_e32 v4, v4 ; SI-NEXT: v_cmp_eq_f32_e32 vcc, v5, v6 @@ -4869,7 +4869,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB9_2: ; SI-NEXT: ; implicit-def: $vgpr4 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB9_3: ; %frem.compute +; SI-NEXT: .LBB9_3: ; %frem.compute19 ; SI-NEXT: s_mov_b32 s3, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 vcc, |v5|, s3 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v4, v5 @@ -4905,10 +4905,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; SI-NEXT: s_cmp_lt_i32 s1, 12 ; SI-NEXT: s_cbranch_scc1 .LBB9_7 -; SI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; SI-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; SI-NEXT: s_sub_i32 s1, s2, s3 ; SI-NEXT: s_add_i32 s1, s1, 11 -; SI-NEXT: .LBB9_5: ; %frem.loop_body +; SI-NEXT: .LBB9_5: ; %frem.loop_body27 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v7, v5 ; SI-NEXT: v_mul_f32_e32 v5, v7, v6 @@ -4923,7 +4923,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB9_5 ; SI-NEXT: ; %bb.6: ; %Flow55 ; SI-NEXT: v_mov_b32_e32 v5, v7 -; SI-NEXT: .LBB9_7: ; %frem.loop_exit +; SI-NEXT: .LBB9_7: ; %frem.loop_exit28 ; SI-NEXT: s_add_i32 s1, s1, -10 ; SI-NEXT: v_ldexp_f32_e64 v5, v5, s1 ; SI-NEXT: v_mul_f32_e32 v6, v5, v6 @@ -4944,7 +4944,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cvt_f32_f16_e64 v7, |v7| ; SI-NEXT: v_cmp_ngt_f32_e32 vcc, v6, v7 ; SI-NEXT: s_cbranch_vccz .LBB9_10 -; SI-NEXT: ; %bb.9: ; %frem.else20 +; SI-NEXT: ; %bb.9: ; %frem.else ; SI-NEXT: s_brev_b32 s0, -2 ; SI-NEXT: v_bfi_b32 v8, s0, 0, v0 ; SI-NEXT: v_cvt_f32_f16_e32 v5, v5 @@ -4956,7 +4956,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB9_10: ; SI-NEXT: ; implicit-def: $vgpr5 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB9_11: ; %frem.compute19 +; SI-NEXT: .LBB9_11: ; %frem.compute ; SI-NEXT: s_mov_b32 s3, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 vcc, |v6|, s3 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v5, v6 @@ -4992,10 +4992,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v7, v7, v5, 1.0 ; SI-NEXT: s_cmp_lt_i32 s1, 12 ; SI-NEXT: s_cbranch_scc1 .LBB9_15 -; SI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; SI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; SI-NEXT: s_sub_i32 s1, s2, s3 ; SI-NEXT: s_add_i32 s1, s1, 11 -; SI-NEXT: .LBB9_13: ; %frem.loop_body27 +; SI-NEXT: .LBB9_13: ; %frem.loop_body ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v8, v6 ; SI-NEXT: v_mul_f32_e32 v6, v8, v7 @@ -5010,7 +5010,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB9_13 ; SI-NEXT: ; %bb.14: ; %Flow ; SI-NEXT: v_mov_b32_e32 v6, v8 -; SI-NEXT: .LBB9_15: ; %frem.loop_exit28 +; SI-NEXT: .LBB9_15: ; %frem.loop_exit ; SI-NEXT: s_add_i32 s1, s1, -10 ; SI-NEXT: v_ldexp_f32_e64 v6, v6, s1 ; SI-NEXT: v_mul_f32_e32 v7, v6, v7 @@ -5084,7 +5084,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_and_b32_e32 v5, 0x7fffffff, v3 ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB9_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else20 ; CI-NEXT: v_cvt_f32_f16_e32 v4, v4 ; CI-NEXT: v_bfi_b32 v7, s0, 0, v2 ; CI-NEXT: v_cmp_eq_f32_e32 vcc, v6, v5 @@ -5093,7 +5093,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB9_8 ; CI-NEXT: .LBB9_2: ; CI-NEXT: ; implicit-def: $vgpr4 -; CI-NEXT: .LBB9_3: ; %frem.compute +; CI-NEXT: .LBB9_3: ; %frem.compute19 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v9, v6 ; CI-NEXT: v_frexp_mant_f32_e32 v4, v6 ; CI-NEXT: v_frexp_mant_f32_e32 v6, v5 @@ -5118,10 +5118,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v6 ; CI-NEXT: v_div_fixup_f32 v8, v8, v5, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB9_7 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; CI-NEXT: v_sub_i32_e32 v6, vcc, v9, v10 ; CI-NEXT: v_add_i32_e32 v6, vcc, 11, v6 -; CI-NEXT: .LBB9_5: ; %frem.loop_body +; CI-NEXT: .LBB9_5: ; %frem.loop_body27 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v9, v7 ; CI-NEXT: v_mul_f32_e32 v7, v9, v8 @@ -5136,7 +5136,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB9_5 ; CI-NEXT: ; %bb.6: ; %Flow55 ; CI-NEXT: v_mov_b32_e32 v7, v9 -; CI-NEXT: .LBB9_7: ; %frem.loop_exit +; CI-NEXT: .LBB9_7: ; %frem.loop_exit28 ; CI-NEXT: v_add_i32_e32 v6, vcc, -10, v6 ; CI-NEXT: v_ldexp_f32_e32 v6, v7, v6 ; CI-NEXT: v_mul_f32_e32 v7, v6, v8 @@ -5157,7 +5157,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cvt_f32_f16_e64 v6, |v6| ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v7, v6 ; CI-NEXT: s_cbranch_vccz .LBB9_10 -; CI-NEXT: ; %bb.9: ; %frem.else20 +; CI-NEXT: ; %bb.9: ; %frem.else ; CI-NEXT: v_cvt_f32_f16_e32 v5, v5 ; CI-NEXT: s_brev_b32 s0, -2 ; CI-NEXT: v_bfi_b32 v8, s0, 0, v0 @@ -5167,7 +5167,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB9_16 ; CI-NEXT: .LBB9_10: ; CI-NEXT: ; implicit-def: $vgpr5 -; CI-NEXT: .LBB9_11: ; %frem.compute19 +; CI-NEXT: .LBB9_11: ; %frem.compute ; CI-NEXT: v_frexp_exp_i32_f32_e32 v10, v7 ; CI-NEXT: v_frexp_mant_f32_e32 v5, v7 ; CI-NEXT: v_frexp_mant_f32_e32 v7, v6 @@ -5192,10 +5192,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v7 ; CI-NEXT: v_div_fixup_f32 v9, v9, v6, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB9_15 -; CI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; CI-NEXT: v_sub_i32_e32 v7, vcc, v10, v11 ; CI-NEXT: v_add_i32_e32 v7, vcc, 11, v7 -; CI-NEXT: .LBB9_13: ; %frem.loop_body27 +; CI-NEXT: .LBB9_13: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v10, v8 ; CI-NEXT: v_mul_f32_e32 v8, v10, v9 @@ -5210,7 +5210,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB9_13 ; CI-NEXT: ; %bb.14: ; %Flow ; CI-NEXT: v_mov_b32_e32 v8, v10 -; CI-NEXT: .LBB9_15: ; %frem.loop_exit28 +; CI-NEXT: .LBB9_15: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v7, vcc, -10, v7 ; CI-NEXT: v_ldexp_f32_e32 v7, v8, v7 ; CI-NEXT: v_mul_f32_e32 v8, v7, v9 @@ -5275,7 +5275,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cvt_f32_f16_e64 v3, |v1| ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v4, v3 ; VI-NEXT: s_cbranch_vccz .LBB9_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else20 ; VI-NEXT: s_movk_i32 s2, 0x7fff ; VI-NEXT: v_bfi_b32 v2, s2, 0, v0 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v4, v3 @@ -5284,7 +5284,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB9_8 ; VI-NEXT: .LBB9_2: ; VI-NEXT: ; implicit-def: $vgpr2 -; VI-NEXT: .LBB9_3: ; %frem.compute +; VI-NEXT: .LBB9_3: ; %frem.compute19 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v7, v4 ; VI-NEXT: v_frexp_mant_f32_e32 v2, v4 ; VI-NEXT: v_frexp_mant_f32_e32 v4, v3 @@ -5309,10 +5309,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v4 ; VI-NEXT: v_div_fixup_f32 v6, v6, v3, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB9_7 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; VI-NEXT: v_sub_u32_e32 v4, vcc, v7, v8 ; VI-NEXT: v_add_u32_e32 v4, vcc, 11, v4 -; VI-NEXT: .LBB9_5: ; %frem.loop_body +; VI-NEXT: .LBB9_5: ; %frem.loop_body27 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v7, v5 ; VI-NEXT: v_mul_f32_e32 v5, v7, v6 @@ -5327,7 +5327,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB9_5 ; VI-NEXT: ; %bb.6: ; %Flow55 ; VI-NEXT: v_mov_b32_e32 v5, v7 -; VI-NEXT: .LBB9_7: ; %frem.loop_exit +; VI-NEXT: .LBB9_7: ; %frem.loop_exit28 ; VI-NEXT: v_add_u32_e32 v4, vcc, -10, v4 ; VI-NEXT: v_ldexp_f32 v4, v5, v4 ; VI-NEXT: v_mul_f32_e32 v5, v4, v6 @@ -5347,7 +5347,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cvt_f32_f16_e64 v6, |v4| ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v7, v6 ; VI-NEXT: s_cbranch_vccz .LBB9_10 -; VI-NEXT: ; %bb.9: ; %frem.else20 +; VI-NEXT: ; %bb.9: ; %frem.else ; VI-NEXT: s_movk_i32 s2, 0x7fff ; VI-NEXT: v_bfi_b32 v5, s2, 0, v3 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v7, v6 @@ -5356,7 +5356,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB9_16 ; VI-NEXT: .LBB9_10: ; VI-NEXT: ; implicit-def: $vgpr5 -; VI-NEXT: .LBB9_11: ; %frem.compute19 +; VI-NEXT: .LBB9_11: ; %frem.compute ; VI-NEXT: v_frexp_exp_i32_f32_e32 v10, v7 ; VI-NEXT: v_frexp_mant_f32_e32 v5, v7 ; VI-NEXT: v_frexp_mant_f32_e32 v7, v6 @@ -5381,10 +5381,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v7 ; VI-NEXT: v_div_fixup_f32 v9, v9, v6, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB9_15 -; VI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; VI-NEXT: v_sub_u32_e32 v7, vcc, v10, v11 ; VI-NEXT: v_add_u32_e32 v7, vcc, 11, v7 -; VI-NEXT: .LBB9_13: ; %frem.loop_body27 +; VI-NEXT: .LBB9_13: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v10, v8 ; VI-NEXT: v_mul_f32_e32 v8, v10, v9 @@ -5399,7 +5399,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB9_13 ; VI-NEXT: ; %bb.14: ; %Flow ; VI-NEXT: v_mov_b32_e32 v8, v10 -; VI-NEXT: .LBB9_15: ; %frem.loop_exit28 +; VI-NEXT: .LBB9_15: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v7, vcc, -10, v7 ; VI-NEXT: v_ldexp_f32 v7, v8, v7 ; VI-NEXT: v_mul_f32_e32 v8, v7, v9 @@ -5443,7 +5443,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cvt_f32_f16_e64 v3, |v0| ; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, v4, v3 ; GFX9-NEXT: s_cbranch_vccz .LBB9_2 -; GFX9-NEXT: ; %bb.1: ; %frem.else +; GFX9-NEXT: ; %bb.1: ; %frem.else20 ; GFX9-NEXT: s_movk_i32 s2, 0x7fff ; GFX9-NEXT: v_bfi_b32 v2, s2, 0, v1 ; GFX9-NEXT: v_cmp_eq_f32_e32 vcc, v4, v3 @@ -5452,7 +5452,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB9_8 ; GFX9-NEXT: .LBB9_2: ; GFX9-NEXT: ; implicit-def: $vgpr2 -; GFX9-NEXT: .LBB9_3: ; %frem.compute +; GFX9-NEXT: .LBB9_3: ; %frem.compute19 ; GFX9-NEXT: v_frexp_exp_i32_f32_e32 v7, v4 ; GFX9-NEXT: v_frexp_mant_f32_e32 v2, v4 ; GFX9-NEXT: v_frexp_mant_f32_e32 v4, v3 @@ -5477,10 +5477,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 12, v4 ; GFX9-NEXT: v_div_fixup_f32 v6, v6, v3, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB9_7 -; GFX9-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX9-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; GFX9-NEXT: v_sub_u32_e32 v4, v7, v8 ; GFX9-NEXT: v_add_u32_e32 v4, 11, v4 -; GFX9-NEXT: .LBB9_5: ; %frem.loop_body +; GFX9-NEXT: .LBB9_5: ; %frem.loop_body27 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v7, v5 ; GFX9-NEXT: v_mul_f32_e32 v5, v7, v6 @@ -5495,7 +5495,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB9_5 ; GFX9-NEXT: ; %bb.6: ; %Flow55 ; GFX9-NEXT: v_mov_b32_e32 v5, v7 -; GFX9-NEXT: .LBB9_7: ; %frem.loop_exit +; GFX9-NEXT: .LBB9_7: ; %frem.loop_exit28 ; GFX9-NEXT: v_add_u32_e32 v4, -10, v4 ; GFX9-NEXT: v_ldexp_f32 v4, v5, v4 ; GFX9-NEXT: v_mul_f32_e32 v5, v4, v6 @@ -5514,7 +5514,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cvt_f32_f16_sdwa v5, |v0| dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 ; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, v6, v5 ; GFX9-NEXT: s_cbranch_vccz .LBB9_10 -; GFX9-NEXT: ; %bb.9: ; %frem.else20 +; GFX9-NEXT: ; %bb.9: ; %frem.else ; GFX9-NEXT: s_movk_i32 s2, 0x7fff ; GFX9-NEXT: v_bfi_b32 v4, s2, 0, v3 ; GFX9-NEXT: v_cmp_eq_f32_e32 vcc, v6, v5 @@ -5523,7 +5523,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB9_16 ; GFX9-NEXT: .LBB9_10: ; GFX9-NEXT: ; implicit-def: $vgpr4 -; GFX9-NEXT: .LBB9_11: ; %frem.compute19 +; GFX9-NEXT: .LBB9_11: ; %frem.compute ; GFX9-NEXT: v_frexp_exp_i32_f32_e32 v9, v6 ; GFX9-NEXT: v_frexp_mant_f32_e32 v4, v6 ; GFX9-NEXT: v_frexp_mant_f32_e32 v6, v5 @@ -5548,10 +5548,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 12, v6 ; GFX9-NEXT: v_div_fixup_f32 v8, v8, v5, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB9_15 -; GFX9-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX9-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX9-NEXT: v_sub_u32_e32 v6, v9, v10 ; GFX9-NEXT: v_add_u32_e32 v6, 11, v6 -; GFX9-NEXT: .LBB9_13: ; %frem.loop_body27 +; GFX9-NEXT: .LBB9_13: ; %frem.loop_body ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v9, v7 ; GFX9-NEXT: v_mul_f32_e32 v7, v9, v8 @@ -5566,7 +5566,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB9_13 ; GFX9-NEXT: ; %bb.14: ; %Flow ; GFX9-NEXT: v_mov_b32_e32 v7, v9 -; GFX9-NEXT: .LBB9_15: ; %frem.loop_exit28 +; GFX9-NEXT: .LBB9_15: ; %frem.loop_exit ; GFX9-NEXT: v_add_u32_e32 v6, -10, v6 ; GFX9-NEXT: v_ldexp_f32 v6, v7, v6 ; GFX9-NEXT: v_mul_f32_e32 v7, v6, v8 @@ -5612,7 +5612,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cvt_f32_f16_e64 v3, |v0| ; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v4, v3 ; GFX10-NEXT: s_cbranch_vccz .LBB9_2 -; GFX10-NEXT: ; %bb.1: ; %frem.else +; GFX10-NEXT: ; %bb.1: ; %frem.else20 ; GFX10-NEXT: v_bfi_b32 v2, 0x7fff, 0, v1 ; GFX10-NEXT: v_cmp_eq_f32_e32 vcc_lo, v4, v3 ; GFX10-NEXT: v_cndmask_b32_e32 v2, v1, v2, vcc_lo @@ -5620,7 +5620,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB9_8 ; GFX10-NEXT: .LBB9_2: ; GFX10-NEXT: ; implicit-def: $vgpr2 -; GFX10-NEXT: .LBB9_3: ; %frem.compute +; GFX10-NEXT: .LBB9_3: ; %frem.compute19 ; GFX10-NEXT: v_frexp_mant_f32_e32 v2, v4 ; GFX10-NEXT: v_frexp_mant_f32_e32 v6, v3 ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v5, v4 @@ -5647,10 +5647,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v6 ; GFX10-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB9_7 -; GFX10-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX10-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 11 -; GFX10-NEXT: .LBB9_5: ; %frem.loop_body +; GFX10-NEXT: .LBB9_5: ; %frem.loop_body27 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v7, v4 ; GFX10-NEXT: s_add_i32 s2, s2, -11 @@ -5666,7 +5666,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.6: ; %Flow55 ; GFX10-NEXT: v_mov_b32_e32 v6, s2 ; GFX10-NEXT: v_mov_b32_e32 v4, v7 -; GFX10-NEXT: .LBB9_7: ; %frem.loop_exit +; GFX10-NEXT: .LBB9_7: ; %frem.loop_exit28 ; GFX10-NEXT: v_add_nc_u32_e32 v6, -10, v6 ; GFX10-NEXT: v_ldexp_f32 v4, v4, v6 ; GFX10-NEXT: v_mul_f32_e32 v5, v4, v5 @@ -5684,7 +5684,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cvt_f32_f16_e64 v6, |v3| ; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v6, v4 ; GFX10-NEXT: s_cbranch_vccz .LBB9_10 -; GFX10-NEXT: ; %bb.9: ; %frem.else20 +; GFX10-NEXT: ; %bb.9: ; %frem.else ; GFX10-NEXT: v_bfi_b32 v5, 0x7fff, 0, v3 ; GFX10-NEXT: v_cmp_eq_f32_e32 vcc_lo, v6, v4 ; GFX10-NEXT: v_cndmask_b32_e32 v5, v3, v5, vcc_lo @@ -5692,7 +5692,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB9_16 ; GFX10-NEXT: .LBB9_10: ; GFX10-NEXT: ; implicit-def: $vgpr5 -; GFX10-NEXT: .LBB9_11: ; %frem.compute19 +; GFX10-NEXT: .LBB9_11: ; %frem.compute ; GFX10-NEXT: v_frexp_mant_f32_e32 v5, v6 ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v7, v6 ; GFX10-NEXT: v_ldexp_f32 v6, v5, 11 @@ -5719,10 +5719,10 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v8 ; GFX10-NEXT: v_div_fixup_f32 v7, v7, v5, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB9_15 -; GFX10-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX10-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 11 -; GFX10-NEXT: .LBB9_13: ; %frem.loop_body27 +; GFX10-NEXT: .LBB9_13: ; %frem.loop_body ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v9, v6 ; GFX10-NEXT: s_add_i32 s2, s2, -11 @@ -5738,7 +5738,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.14: ; %Flow ; GFX10-NEXT: v_mov_b32_e32 v8, s2 ; GFX10-NEXT: v_mov_b32_e32 v6, v9 -; GFX10-NEXT: .LBB9_15: ; %frem.loop_exit28 +; GFX10-NEXT: .LBB9_15: ; %frem.loop_exit ; GFX10-NEXT: v_add_nc_u32_e32 v8, -10, v8 ; GFX10-NEXT: v_ldexp_f32 v6, v6, v8 ; GFX10-NEXT: v_mul_f32_e32 v7, v6, v7 @@ -5782,7 +5782,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v4, v3 ; GFX11-TRUE16-NEXT: s_cbranch_vccz .LBB9_2 -; GFX11-TRUE16-NEXT: ; %bb.1: ; %frem.else +; GFX11-TRUE16-NEXT: ; %bb.1: ; %frem.else20 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v0.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, 0 ; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v4, v3 @@ -5793,7 +5793,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_branch .LBB9_8 ; GFX11-TRUE16-NEXT: .LBB9_2: ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr2 -; GFX11-TRUE16-NEXT: .LBB9_3: ; %frem.compute +; GFX11-TRUE16-NEXT: .LBB9_3: ; %frem.compute19 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v2, v4 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v6, v3 ; GFX11-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v5, v4 @@ -5829,11 +5829,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB9_7 -; GFX11-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX11-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; GFX11-TRUE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-TRUE16-NEXT: .LBB9_5: ; %frem.loop_body +; GFX11-TRUE16-NEXT: .LBB9_5: ; %frem.loop_body27 ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v7, v4 @@ -5853,7 +5853,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: ; %bb.6: ; %Flow55 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, s2 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v7 -; GFX11-TRUE16-NEXT: .LBB9_7: ; %frem.loop_exit +; GFX11-TRUE16-NEXT: .LBB9_7: ; %frem.loop_exit28 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v6, -10, v6 ; GFX11-TRUE16-NEXT: v_ldexp_f32 v4, v4, v6 @@ -5880,7 +5880,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v6, v5 ; GFX11-TRUE16-NEXT: s_cbranch_vccz .LBB9_10 -; GFX11-TRUE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX11-TRUE16-NEXT: ; %bb.9: ; %frem.else ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v3.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, 0 ; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v6, v5 @@ -5891,7 +5891,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_branch .LBB9_16 ; GFX11-TRUE16-NEXT: .LBB9_10: ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7 -; GFX11-TRUE16-NEXT: .LBB9_11: ; %frem.compute19 +; GFX11-TRUE16-NEXT: .LBB9_11: ; %frem.compute ; GFX11-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v8, v6 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v6, v6 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) @@ -5927,11 +5927,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_div_fixup_f32 v8, v8, v6, 1.0 ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB9_15 -; GFX11-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX11-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX11-TRUE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-TRUE16-NEXT: .LBB9_13: ; %frem.loop_body27 +; GFX11-TRUE16-NEXT: .LBB9_13: ; %frem.loop_body ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v10, v7 @@ -5951,7 +5951,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: ; %bb.14: ; %Flow ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v9, s2 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v7, v10 -; GFX11-TRUE16-NEXT: .LBB9_15: ; %frem.loop_exit28 +; GFX11-TRUE16-NEXT: .LBB9_15: ; %frem.loop_exit ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v9, -10, v9 ; GFX11-TRUE16-NEXT: v_ldexp_f32 v7, v7, v9 @@ -6002,7 +6002,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v4, v3 ; GFX11-FAKE16-NEXT: s_cbranch_vccz .LBB9_2 -; GFX11-FAKE16-NEXT: ; %bb.1: ; %frem.else +; GFX11-FAKE16-NEXT: ; %bb.1: ; %frem.else20 ; GFX11-FAKE16-NEXT: v_bfi_b32 v2, 0x7fff, 0, v0 ; GFX11-FAKE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v4, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -6011,7 +6011,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_branch .LBB9_8 ; GFX11-FAKE16-NEXT: .LBB9_2: ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr2 -; GFX11-FAKE16-NEXT: .LBB9_3: ; %frem.compute +; GFX11-FAKE16-NEXT: .LBB9_3: ; %frem.compute19 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v2, v4 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v6, v3 ; GFX11-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v5, v4 @@ -6047,11 +6047,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-FAKE16-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB9_7 -; GFX11-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX11-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; GFX11-FAKE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-FAKE16-NEXT: .LBB9_5: ; %frem.loop_body +; GFX11-FAKE16-NEXT: .LBB9_5: ; %frem.loop_body27 ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v7, v4 @@ -6071,7 +6071,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: ; %bb.6: ; %Flow55 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v6, s2 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v7 -; GFX11-FAKE16-NEXT: .LBB9_7: ; %frem.loop_exit +; GFX11-FAKE16-NEXT: .LBB9_7: ; %frem.loop_exit28 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v6, -10, v6 ; GFX11-FAKE16-NEXT: v_ldexp_f32 v4, v4, v6 @@ -6097,7 +6097,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v7, v5 ; GFX11-FAKE16-NEXT: s_cbranch_vccz .LBB9_10 -; GFX11-FAKE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX11-FAKE16-NEXT: ; %bb.9: ; %frem.else ; GFX11-FAKE16-NEXT: v_bfi_b32 v6, 0x7fff, 0, v3 ; GFX11-FAKE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v7, v5 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -6106,7 +6106,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_branch .LBB9_16 ; GFX11-FAKE16-NEXT: .LBB9_10: ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr6 -; GFX11-FAKE16-NEXT: .LBB9_11: ; %frem.compute19 +; GFX11-FAKE16-NEXT: .LBB9_11: ; %frem.compute ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v6, v7 ; GFX11-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v8, v7 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4) @@ -6142,11 +6142,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-FAKE16-NEXT: v_div_fixup_f32 v8, v8, v6, 1.0 ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB9_15 -; GFX11-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX11-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX11-FAKE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-FAKE16-NEXT: .LBB9_13: ; %frem.loop_body27 +; GFX11-FAKE16-NEXT: .LBB9_13: ; %frem.loop_body ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v10, v7 @@ -6166,7 +6166,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: ; %bb.14: ; %Flow ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v9, s2 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v7, v10 -; GFX11-FAKE16-NEXT: .LBB9_15: ; %frem.loop_exit28 +; GFX11-FAKE16-NEXT: .LBB9_15: ; %frem.loop_exit ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v9, -10, v9 ; GFX11-FAKE16-NEXT: v_ldexp_f32 v7, v7, v9 @@ -6220,7 +6220,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-TRUE16-NEXT: s_cmp_ngt_f32 s6, s5 ; GFX1150-TRUE16-NEXT: s_cbranch_scc0 .LBB9_2 -; GFX1150-TRUE16-NEXT: ; %bb.1: ; %frem.else +; GFX1150-TRUE16-NEXT: ; %bb.1: ; %frem.else20 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v0.l, s4 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 ; GFX1150-TRUE16-NEXT: s_cmp_eq_f32 s6, s5 @@ -6232,7 +6232,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_branch .LBB9_8 ; GFX1150-TRUE16-NEXT: .LBB9_2: ; GFX1150-TRUE16-NEXT: ; implicit-def: $vgpr0 -; GFX1150-TRUE16-NEXT: .LBB9_3: ; %frem.compute +; GFX1150-TRUE16-NEXT: .LBB9_3: ; %frem.compute19 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v1, s5 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v0, s6 ; GFX1150-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v3, s6 @@ -6267,11 +6267,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v4 ; GFX1150-TRUE16-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1150-TRUE16-NEXT: s_cbranch_vccnz .LBB9_7 -; GFX1150-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1150-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; GFX1150-TRUE16-NEXT: s_sub_i32 s5, s6, s5 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: s_add_i32 s5, s5, 11 -; GFX1150-TRUE16-NEXT: .LBB9_5: ; %frem.loop_body +; GFX1150-TRUE16-NEXT: .LBB9_5: ; %frem.loop_body27 ; GFX1150-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v5, v2 @@ -6293,7 +6293,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: ; %bb.6: ; %Flow55 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v4, s5 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v2, v5 -; GFX1150-TRUE16-NEXT: .LBB9_7: ; %frem.loop_exit +; GFX1150-TRUE16-NEXT: .LBB9_7: ; %frem.loop_exit28 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-TRUE16-NEXT: v_add_nc_u32_e32 v4, -10, v4 ; GFX1150-TRUE16-NEXT: v_ldexp_f32 v2, v2, v4 @@ -6323,7 +6323,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-TRUE16-NEXT: s_cmp_ngt_f32 s8, s7 ; GFX1150-TRUE16-NEXT: s_cbranch_scc0 .LBB9_10 -; GFX1150-TRUE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX1150-TRUE16-NEXT: ; %bb.9: ; %frem.else ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v1.l, s6 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 ; GFX1150-TRUE16-NEXT: s_cmp_eq_f32 s8, s7 @@ -6335,7 +6335,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_branch .LBB9_16 ; GFX1150-TRUE16-NEXT: .LBB9_10: ; GFX1150-TRUE16-NEXT: ; implicit-def: $vgpr1 -; GFX1150-TRUE16-NEXT: .LBB9_11: ; %frem.compute19 +; GFX1150-TRUE16-NEXT: .LBB9_11: ; %frem.compute ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v2, s7 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v1, s8 ; GFX1150-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v4, s8 @@ -6370,11 +6370,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v5 ; GFX1150-TRUE16-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1150-TRUE16-NEXT: s_cbranch_vccnz .LBB9_15 -; GFX1150-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX1150-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX1150-TRUE16-NEXT: s_sub_i32 s7, s8, s7 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: s_add_i32 s7, s7, 11 -; GFX1150-TRUE16-NEXT: .LBB9_13: ; %frem.loop_body27 +; GFX1150-TRUE16-NEXT: .LBB9_13: ; %frem.loop_body ; GFX1150-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v6, v3 @@ -6396,7 +6396,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: ; %bb.14: ; %Flow ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v5, s7 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v3, v6 -; GFX1150-TRUE16-NEXT: .LBB9_15: ; %frem.loop_exit28 +; GFX1150-TRUE16-NEXT: .LBB9_15: ; %frem.loop_exit ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-TRUE16-NEXT: v_add_nc_u32_e32 v5, -10, v5 ; GFX1150-TRUE16-NEXT: v_ldexp_f32 v3, v3, v5 @@ -6459,7 +6459,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-FAKE16-NEXT: s_cmp_ngt_f32 s6, s5 ; GFX1150-FAKE16-NEXT: s_cbranch_scc0 .LBB9_2 -; GFX1150-FAKE16-NEXT: ; %bb.1: ; %frem.else +; GFX1150-FAKE16-NEXT: ; %bb.1: ; %frem.else20 ; GFX1150-FAKE16-NEXT: s_cmp_eq_f32 s6, s5 ; GFX1150-FAKE16-NEXT: v_bfi_b32 v0, 0x7fff, 0, s4 ; GFX1150-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -6469,7 +6469,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_branch .LBB9_8 ; GFX1150-FAKE16-NEXT: .LBB9_2: ; GFX1150-FAKE16-NEXT: ; implicit-def: $vgpr0 -; GFX1150-FAKE16-NEXT: .LBB9_3: ; %frem.compute +; GFX1150-FAKE16-NEXT: .LBB9_3: ; %frem.compute19 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v1, s5 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v0, s6 ; GFX1150-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v3, s6 @@ -6504,11 +6504,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v4 ; GFX1150-FAKE16-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1150-FAKE16-NEXT: s_cbranch_vccnz .LBB9_7 -; GFX1150-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1150-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; GFX1150-FAKE16-NEXT: s_sub_i32 s5, s6, s5 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: s_add_i32 s5, s5, 11 -; GFX1150-FAKE16-NEXT: .LBB9_5: ; %frem.loop_body +; GFX1150-FAKE16-NEXT: .LBB9_5: ; %frem.loop_body27 ; GFX1150-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v5, v2 @@ -6530,7 +6530,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: ; %bb.6: ; %Flow55 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v4, s5 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v2, v5 -; GFX1150-FAKE16-NEXT: .LBB9_7: ; %frem.loop_exit +; GFX1150-FAKE16-NEXT: .LBB9_7: ; %frem.loop_exit28 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-FAKE16-NEXT: v_add_nc_u32_e32 v4, -10, v4 ; GFX1150-FAKE16-NEXT: v_ldexp_f32 v2, v2, v4 @@ -6559,7 +6559,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-FAKE16-NEXT: s_cmp_ngt_f32 s8, s7 ; GFX1150-FAKE16-NEXT: s_cbranch_scc0 .LBB9_10 -; GFX1150-FAKE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX1150-FAKE16-NEXT: ; %bb.9: ; %frem.else ; GFX1150-FAKE16-NEXT: s_cmp_eq_f32 s8, s7 ; GFX1150-FAKE16-NEXT: v_bfi_b32 v1, 0x7fff, 0, s6 ; GFX1150-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -6569,7 +6569,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_branch .LBB9_16 ; GFX1150-FAKE16-NEXT: .LBB9_10: ; GFX1150-FAKE16-NEXT: ; implicit-def: $vgpr1 -; GFX1150-FAKE16-NEXT: .LBB9_11: ; %frem.compute19 +; GFX1150-FAKE16-NEXT: .LBB9_11: ; %frem.compute ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v2, s7 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v1, s8 ; GFX1150-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v4, s8 @@ -6604,11 +6604,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v5 ; GFX1150-FAKE16-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1150-FAKE16-NEXT: s_cbranch_vccnz .LBB9_15 -; GFX1150-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX1150-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX1150-FAKE16-NEXT: s_sub_i32 s7, s8, s7 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: s_add_i32 s7, s7, 11 -; GFX1150-FAKE16-NEXT: .LBB9_13: ; %frem.loop_body27 +; GFX1150-FAKE16-NEXT: .LBB9_13: ; %frem.loop_body ; GFX1150-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v6, v3 @@ -6630,7 +6630,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: ; %bb.14: ; %Flow ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v5, s7 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v3, v6 -; GFX1150-FAKE16-NEXT: .LBB9_15: ; %frem.loop_exit28 +; GFX1150-FAKE16-NEXT: .LBB9_15: ; %frem.loop_exit ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-FAKE16-NEXT: v_add_nc_u32_e32 v5, -10, v5 ; GFX1150-FAKE16-NEXT: v_ldexp_f32 v3, v3, v5 @@ -6690,7 +6690,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1200-TRUE16-NEXT: s_cmp_ngt_f32 s6, s5 ; GFX1200-TRUE16-NEXT: s_cbranch_scc0 .LBB9_2 -; GFX1200-TRUE16-NEXT: ; %bb.1: ; %frem.else +; GFX1200-TRUE16-NEXT: ; %bb.1: ; %frem.else20 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v0.l, s4 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 ; GFX1200-TRUE16-NEXT: s_cmp_eq_f32 s6, s5 @@ -6702,7 +6702,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_branch .LBB9_8 ; GFX1200-TRUE16-NEXT: .LBB9_2: ; GFX1200-TRUE16-NEXT: ; implicit-def: $vgpr0 -; GFX1200-TRUE16-NEXT: .LBB9_3: ; %frem.compute +; GFX1200-TRUE16-NEXT: .LBB9_3: ; %frem.compute19 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v1, s5 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v0, s6 ; GFX1200-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v3, s6 @@ -6737,11 +6737,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v4 ; GFX1200-TRUE16-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1200-TRUE16-NEXT: s_cbranch_vccnz .LBB9_7 -; GFX1200-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1200-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; GFX1200-TRUE16-NEXT: s_sub_co_i32 s5, s6, s5 ; GFX1200-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX1200-TRUE16-NEXT: s_add_co_i32 s5, s5, 11 -; GFX1200-TRUE16-NEXT: .LBB9_5: ; %frem.loop_body +; GFX1200-TRUE16-NEXT: .LBB9_5: ; %frem.loop_body27 ; GFX1200-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v5, v2 @@ -6765,7 +6765,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: ; %bb.6: ; %Flow55 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v4, s5 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v2, v5 -; GFX1200-TRUE16-NEXT: .LBB9_7: ; %frem.loop_exit +; GFX1200-TRUE16-NEXT: .LBB9_7: ; %frem.loop_exit28 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_add_nc_u32_e32 v4, -10, v4 ; GFX1200-TRUE16-NEXT: v_ldexp_f32 v2, v2, v4 @@ -6799,7 +6799,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_2) ; GFX1200-TRUE16-NEXT: s_cmp_ngt_f32 s8, s7 ; GFX1200-TRUE16-NEXT: s_cbranch_scc0 .LBB9_10 -; GFX1200-TRUE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX1200-TRUE16-NEXT: ; %bb.9: ; %frem.else ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v1.l, s6 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 ; GFX1200-TRUE16-NEXT: s_cmp_eq_f32 s8, s7 @@ -6811,7 +6811,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_branch .LBB9_16 ; GFX1200-TRUE16-NEXT: .LBB9_10: ; GFX1200-TRUE16-NEXT: ; implicit-def: $vgpr1 -; GFX1200-TRUE16-NEXT: .LBB9_11: ; %frem.compute19 +; GFX1200-TRUE16-NEXT: .LBB9_11: ; %frem.compute ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v2, s7 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v1, s8 ; GFX1200-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v4, s8 @@ -6847,11 +6847,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v5 ; GFX1200-TRUE16-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1200-TRUE16-NEXT: s_cbranch_vccnz .LBB9_15 -; GFX1200-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX1200-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX1200-TRUE16-NEXT: s_sub_co_i32 s7, s8, s7 ; GFX1200-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX1200-TRUE16-NEXT: s_add_co_i32 s7, s7, 11 -; GFX1200-TRUE16-NEXT: .LBB9_13: ; %frem.loop_body27 +; GFX1200-TRUE16-NEXT: .LBB9_13: ; %frem.loop_body ; GFX1200-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v6, v3 @@ -6875,7 +6875,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: ; %bb.14: ; %Flow ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v5, s7 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v3, v6 -; GFX1200-TRUE16-NEXT: .LBB9_15: ; %frem.loop_exit28 +; GFX1200-TRUE16-NEXT: .LBB9_15: ; %frem.loop_exit ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_add_nc_u32_e32 v5, -10, v5 ; GFX1200-TRUE16-NEXT: v_ldexp_f32 v3, v3, v5 @@ -6940,7 +6940,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1200-FAKE16-NEXT: s_cmp_ngt_f32 s6, s5 ; GFX1200-FAKE16-NEXT: s_cbranch_scc0 .LBB9_2 -; GFX1200-FAKE16-NEXT: ; %bb.1: ; %frem.else +; GFX1200-FAKE16-NEXT: ; %bb.1: ; %frem.else20 ; GFX1200-FAKE16-NEXT: s_cmp_eq_f32 s6, s5 ; GFX1200-FAKE16-NEXT: v_bfi_b32 v0, 0x7fff, 0, s4 ; GFX1200-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -6950,7 +6950,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_branch .LBB9_8 ; GFX1200-FAKE16-NEXT: .LBB9_2: ; GFX1200-FAKE16-NEXT: ; implicit-def: $vgpr0 -; GFX1200-FAKE16-NEXT: .LBB9_3: ; %frem.compute +; GFX1200-FAKE16-NEXT: .LBB9_3: ; %frem.compute19 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v1, s5 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v0, s6 ; GFX1200-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v3, s6 @@ -6986,11 +6986,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v4 ; GFX1200-FAKE16-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1200-FAKE16-NEXT: s_cbranch_vccnz .LBB9_7 -; GFX1200-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1200-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body27.preheader ; GFX1200-FAKE16-NEXT: s_sub_co_i32 s5, s6, s5 ; GFX1200-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX1200-FAKE16-NEXT: s_add_co_i32 s5, s5, 11 -; GFX1200-FAKE16-NEXT: .LBB9_5: ; %frem.loop_body +; GFX1200-FAKE16-NEXT: .LBB9_5: ; %frem.loop_body27 ; GFX1200-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v5, v2 @@ -7014,7 +7014,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: ; %bb.6: ; %Flow55 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v4, s5 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v2, v5 -; GFX1200-FAKE16-NEXT: .LBB9_7: ; %frem.loop_exit +; GFX1200-FAKE16-NEXT: .LBB9_7: ; %frem.loop_exit28 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_add_nc_u32_e32 v4, -10, v4 ; GFX1200-FAKE16-NEXT: v_ldexp_f32 v2, v2, v4 @@ -7047,7 +7047,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_2) ; GFX1200-FAKE16-NEXT: s_cmp_ngt_f32 s8, s7 ; GFX1200-FAKE16-NEXT: s_cbranch_scc0 .LBB9_10 -; GFX1200-FAKE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX1200-FAKE16-NEXT: ; %bb.9: ; %frem.else ; GFX1200-FAKE16-NEXT: s_cmp_eq_f32 s8, s7 ; GFX1200-FAKE16-NEXT: v_bfi_b32 v1, 0x7fff, 0, s6 ; GFX1200-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -7058,7 +7058,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_branch .LBB9_16 ; GFX1200-FAKE16-NEXT: .LBB9_10: ; GFX1200-FAKE16-NEXT: ; implicit-def: $vgpr1 -; GFX1200-FAKE16-NEXT: .LBB9_11: ; %frem.compute19 +; GFX1200-FAKE16-NEXT: .LBB9_11: ; %frem.compute ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v2, s7 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v1, s8 ; GFX1200-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v4, s8 @@ -7094,11 +7094,11 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v5 ; GFX1200-FAKE16-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1200-FAKE16-NEXT: s_cbranch_vccnz .LBB9_15 -; GFX1200-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX1200-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX1200-FAKE16-NEXT: s_sub_co_i32 s7, s8, s7 ; GFX1200-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX1200-FAKE16-NEXT: s_add_co_i32 s7, s7, 11 -; GFX1200-FAKE16-NEXT: .LBB9_13: ; %frem.loop_body27 +; GFX1200-FAKE16-NEXT: .LBB9_13: ; %frem.loop_body ; GFX1200-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v6, v3 @@ -7122,7 +7122,7 @@ define amdgpu_kernel void @frem_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: ; %bb.14: ; %Flow ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v5, s7 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v3, v6 -; GFX1200-FAKE16-NEXT: .LBB9_15: ; %frem.loop_exit28 +; GFX1200-FAKE16-NEXT: .LBB9_15: ; %frem.loop_exit ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_add_nc_u32_e32 v5, -10, v5 ; GFX1200-FAKE16-NEXT: v_ldexp_f32 v3, v3, v5 @@ -7208,7 +7208,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_and_b64 vcc, exec, s[2:3] ; SI-NEXT: v_cvt_f16_f32_e32 v8, v6 ; SI-NEXT: s_cbranch_vccz .LBB10_2 -; SI-NEXT: ; %bb.1: ; %frem.else +; SI-NEXT: ; %bb.1: ; %frem.else86 ; SI-NEXT: v_bfi_b32 v11, s0, 0, v6 ; SI-NEXT: v_cvt_f32_f16_e32 v8, v8 ; SI-NEXT: v_cmp_eq_f32_e32 vcc, v9, v10 @@ -7219,7 +7219,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB10_2: ; SI-NEXT: ; implicit-def: $vgpr8 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB10_3: ; %frem.compute +; SI-NEXT: .LBB10_3: ; %frem.compute85 ; SI-NEXT: s_mov_b32 s3, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 vcc, |v9|, s3 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v8, v9 @@ -7255,10 +7255,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v10, v10, v8, 1.0 ; SI-NEXT: s_cmp_lt_i32 s1, 12 ; SI-NEXT: s_cbranch_scc1 .LBB10_7 -; SI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; SI-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; SI-NEXT: s_sub_i32 s1, s2, s3 ; SI-NEXT: s_add_i32 s1, s1, 11 -; SI-NEXT: .LBB10_5: ; %frem.loop_body +; SI-NEXT: .LBB10_5: ; %frem.loop_body93 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v11, v9 ; SI-NEXT: v_mul_f32_e32 v9, v11, v10 @@ -7273,7 +7273,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB10_5 ; SI-NEXT: ; %bb.6: ; %Flow133 ; SI-NEXT: v_mov_b32_e32 v9, v11 -; SI-NEXT: .LBB10_7: ; %frem.loop_exit +; SI-NEXT: .LBB10_7: ; %frem.loop_exit94 ; SI-NEXT: s_add_i32 s1, s1, -10 ; SI-NEXT: v_ldexp_f32_e64 v9, v9, s1 ; SI-NEXT: v_mul_f32_e32 v10, v9, v10 @@ -7294,7 +7294,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cvt_f32_f16_e64 v11, |v11| ; SI-NEXT: v_cmp_ngt_f32_e32 vcc, v10, v11 ; SI-NEXT: s_cbranch_vccz .LBB10_10 -; SI-NEXT: ; %bb.9: ; %frem.else20 +; SI-NEXT: ; %bb.9: ; %frem.else53 ; SI-NEXT: s_brev_b32 s0, -2 ; SI-NEXT: v_bfi_b32 v12, s0, 0, v4 ; SI-NEXT: v_cvt_f32_f16_e32 v9, v9 @@ -7306,7 +7306,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB10_10: ; SI-NEXT: ; implicit-def: $vgpr9 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB10_11: ; %frem.compute19 +; SI-NEXT: .LBB10_11: ; %frem.compute52 ; SI-NEXT: s_mov_b32 s3, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 vcc, |v10|, s3 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v9, v10 @@ -7342,10 +7342,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v11, v11, v9, 1.0 ; SI-NEXT: s_cmp_lt_i32 s1, 12 ; SI-NEXT: s_cbranch_scc1 .LBB10_15 -; SI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; SI-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; SI-NEXT: s_sub_i32 s1, s2, s3 ; SI-NEXT: s_add_i32 s1, s1, 11 -; SI-NEXT: .LBB10_13: ; %frem.loop_body27 +; SI-NEXT: .LBB10_13: ; %frem.loop_body60 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v12, v10 ; SI-NEXT: v_mul_f32_e32 v10, v12, v11 @@ -7360,7 +7360,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB10_13 ; SI-NEXT: ; %bb.14: ; %Flow129 ; SI-NEXT: v_mov_b32_e32 v10, v12 -; SI-NEXT: .LBB10_15: ; %frem.loop_exit28 +; SI-NEXT: .LBB10_15: ; %frem.loop_exit61 ; SI-NEXT: s_add_i32 s1, s1, -10 ; SI-NEXT: v_ldexp_f32_e64 v10, v10, s1 ; SI-NEXT: v_mul_f32_e32 v11, v10, v11 @@ -7381,7 +7381,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cvt_f32_f16_e64 v12, |v12| ; SI-NEXT: v_cmp_ngt_f32_e32 vcc, v11, v12 ; SI-NEXT: s_cbranch_vccz .LBB10_18 -; SI-NEXT: ; %bb.17: ; %frem.else53 +; SI-NEXT: ; %bb.17: ; %frem.else20 ; SI-NEXT: s_brev_b32 s0, -2 ; SI-NEXT: v_bfi_b32 v13, s0, 0, v2 ; SI-NEXT: v_cvt_f32_f16_e32 v10, v10 @@ -7393,7 +7393,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB10_18: ; SI-NEXT: ; implicit-def: $vgpr10 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB10_19: ; %frem.compute52 +; SI-NEXT: .LBB10_19: ; %frem.compute19 ; SI-NEXT: s_mov_b32 s3, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 vcc, |v11|, s3 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v10, v11 @@ -7429,10 +7429,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v12, v12, v10, 1.0 ; SI-NEXT: s_cmp_lt_i32 s1, 12 ; SI-NEXT: s_cbranch_scc1 .LBB10_23 -; SI-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; SI-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; SI-NEXT: s_sub_i32 s1, s2, s3 ; SI-NEXT: s_add_i32 s1, s1, 11 -; SI-NEXT: .LBB10_21: ; %frem.loop_body60 +; SI-NEXT: .LBB10_21: ; %frem.loop_body27 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v13, v11 ; SI-NEXT: v_mul_f32_e32 v11, v13, v12 @@ -7447,7 +7447,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB10_21 ; SI-NEXT: ; %bb.22: ; %Flow125 ; SI-NEXT: v_mov_b32_e32 v11, v13 -; SI-NEXT: .LBB10_23: ; %frem.loop_exit61 +; SI-NEXT: .LBB10_23: ; %frem.loop_exit28 ; SI-NEXT: s_add_i32 s1, s1, -10 ; SI-NEXT: v_ldexp_f32_e64 v11, v11, s1 ; SI-NEXT: v_mul_f32_e32 v12, v11, v12 @@ -7468,7 +7468,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cvt_f32_f16_e64 v13, |v13| ; SI-NEXT: v_cmp_ngt_f32_e32 vcc, v12, v13 ; SI-NEXT: s_cbranch_vccz .LBB10_26 -; SI-NEXT: ; %bb.25: ; %frem.else86 +; SI-NEXT: ; %bb.25: ; %frem.else ; SI-NEXT: s_brev_b32 s0, -2 ; SI-NEXT: v_bfi_b32 v14, s0, 0, v0 ; SI-NEXT: v_cvt_f32_f16_e32 v11, v11 @@ -7480,7 +7480,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB10_26: ; SI-NEXT: ; implicit-def: $vgpr11 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB10_27: ; %frem.compute85 +; SI-NEXT: .LBB10_27: ; %frem.compute ; SI-NEXT: s_mov_b32 s3, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 vcc, |v12|, s3 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v11, v12 @@ -7516,10 +7516,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v13, v13, v11, 1.0 ; SI-NEXT: s_cmp_lt_i32 s1, 12 ; SI-NEXT: s_cbranch_scc1 .LBB10_31 -; SI-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; SI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; SI-NEXT: s_sub_i32 s1, s2, s3 ; SI-NEXT: s_add_i32 s1, s1, 11 -; SI-NEXT: .LBB10_29: ; %frem.loop_body93 +; SI-NEXT: .LBB10_29: ; %frem.loop_body ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v14, v12 ; SI-NEXT: v_mul_f32_e32 v12, v14, v13 @@ -7534,7 +7534,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB10_29 ; SI-NEXT: ; %bb.30: ; %Flow ; SI-NEXT: v_mov_b32_e32 v12, v14 -; SI-NEXT: .LBB10_31: ; %frem.loop_exit94 +; SI-NEXT: .LBB10_31: ; %frem.loop_exit ; SI-NEXT: s_add_i32 s1, s1, -10 ; SI-NEXT: v_ldexp_f32_e64 v12, v12, s1 ; SI-NEXT: v_mul_f32_e32 v13, v12, v13 @@ -7638,7 +7638,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_and_b32_e32 v9, 0x7fffffff, v7 ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB10_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else86 ; CI-NEXT: v_cvt_f32_f16_e32 v8, v8 ; CI-NEXT: v_bfi_b32 v11, s0, 0, v6 ; CI-NEXT: v_cmp_eq_f32_e32 vcc, v10, v9 @@ -7647,7 +7647,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB10_8 ; CI-NEXT: .LBB10_2: ; CI-NEXT: ; implicit-def: $vgpr8 -; CI-NEXT: .LBB10_3: ; %frem.compute +; CI-NEXT: .LBB10_3: ; %frem.compute85 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v13, v10 ; CI-NEXT: v_frexp_mant_f32_e32 v8, v10 ; CI-NEXT: v_frexp_mant_f32_e32 v10, v9 @@ -7672,10 +7672,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v10 ; CI-NEXT: v_div_fixup_f32 v12, v12, v9, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB10_7 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; CI-NEXT: v_sub_i32_e32 v10, vcc, v13, v14 ; CI-NEXT: v_add_i32_e32 v10, vcc, 11, v10 -; CI-NEXT: .LBB10_5: ; %frem.loop_body +; CI-NEXT: .LBB10_5: ; %frem.loop_body93 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v13, v11 ; CI-NEXT: v_mul_f32_e32 v11, v13, v12 @@ -7690,7 +7690,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB10_5 ; CI-NEXT: ; %bb.6: ; %Flow133 ; CI-NEXT: v_mov_b32_e32 v11, v13 -; CI-NEXT: .LBB10_7: ; %frem.loop_exit +; CI-NEXT: .LBB10_7: ; %frem.loop_exit94 ; CI-NEXT: v_add_i32_e32 v10, vcc, -10, v10 ; CI-NEXT: v_ldexp_f32_e32 v10, v11, v10 ; CI-NEXT: v_mul_f32_e32 v11, v10, v12 @@ -7711,7 +7711,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cvt_f32_f16_e64 v10, |v10| ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v11, v10 ; CI-NEXT: s_cbranch_vccz .LBB10_10 -; CI-NEXT: ; %bb.9: ; %frem.else20 +; CI-NEXT: ; %bb.9: ; %frem.else53 ; CI-NEXT: v_cvt_f32_f16_e32 v9, v9 ; CI-NEXT: s_brev_b32 s0, -2 ; CI-NEXT: v_bfi_b32 v12, s0, 0, v4 @@ -7721,7 +7721,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB10_16 ; CI-NEXT: .LBB10_10: ; CI-NEXT: ; implicit-def: $vgpr9 -; CI-NEXT: .LBB10_11: ; %frem.compute19 +; CI-NEXT: .LBB10_11: ; %frem.compute52 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v14, v11 ; CI-NEXT: v_frexp_mant_f32_e32 v9, v11 ; CI-NEXT: v_frexp_mant_f32_e32 v11, v10 @@ -7746,10 +7746,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v11 ; CI-NEXT: v_div_fixup_f32 v13, v13, v10, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB10_15 -; CI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; CI-NEXT: v_sub_i32_e32 v11, vcc, v14, v15 ; CI-NEXT: v_add_i32_e32 v11, vcc, 11, v11 -; CI-NEXT: .LBB10_13: ; %frem.loop_body27 +; CI-NEXT: .LBB10_13: ; %frem.loop_body60 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v14, v12 ; CI-NEXT: v_mul_f32_e32 v12, v14, v13 @@ -7764,7 +7764,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB10_13 ; CI-NEXT: ; %bb.14: ; %Flow129 ; CI-NEXT: v_mov_b32_e32 v12, v14 -; CI-NEXT: .LBB10_15: ; %frem.loop_exit28 +; CI-NEXT: .LBB10_15: ; %frem.loop_exit61 ; CI-NEXT: v_add_i32_e32 v11, vcc, -10, v11 ; CI-NEXT: v_ldexp_f32_e32 v11, v12, v11 ; CI-NEXT: v_mul_f32_e32 v12, v11, v13 @@ -7785,7 +7785,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cvt_f32_f16_e64 v11, |v11| ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v12, v11 ; CI-NEXT: s_cbranch_vccz .LBB10_18 -; CI-NEXT: ; %bb.17: ; %frem.else53 +; CI-NEXT: ; %bb.17: ; %frem.else20 ; CI-NEXT: v_cvt_f32_f16_e32 v10, v10 ; CI-NEXT: s_brev_b32 s0, -2 ; CI-NEXT: v_bfi_b32 v13, s0, 0, v2 @@ -7795,7 +7795,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB10_24 ; CI-NEXT: .LBB10_18: ; CI-NEXT: ; implicit-def: $vgpr10 -; CI-NEXT: .LBB10_19: ; %frem.compute52 +; CI-NEXT: .LBB10_19: ; %frem.compute19 ; CI-NEXT: v_frexp_exp_i32_f32_e32 v15, v12 ; CI-NEXT: v_frexp_mant_f32_e32 v10, v12 ; CI-NEXT: v_frexp_mant_f32_e32 v12, v11 @@ -7820,10 +7820,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v12 ; CI-NEXT: v_div_fixup_f32 v14, v14, v11, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB10_23 -; CI-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; CI-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; CI-NEXT: v_sub_i32_e32 v12, vcc, v15, v16 ; CI-NEXT: v_add_i32_e32 v12, vcc, 11, v12 -; CI-NEXT: .LBB10_21: ; %frem.loop_body60 +; CI-NEXT: .LBB10_21: ; %frem.loop_body27 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v15, v13 ; CI-NEXT: v_mul_f32_e32 v13, v15, v14 @@ -7838,7 +7838,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB10_21 ; CI-NEXT: ; %bb.22: ; %Flow125 ; CI-NEXT: v_mov_b32_e32 v13, v15 -; CI-NEXT: .LBB10_23: ; %frem.loop_exit61 +; CI-NEXT: .LBB10_23: ; %frem.loop_exit28 ; CI-NEXT: v_add_i32_e32 v12, vcc, -10, v12 ; CI-NEXT: v_ldexp_f32_e32 v12, v13, v12 ; CI-NEXT: v_mul_f32_e32 v13, v12, v14 @@ -7859,7 +7859,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cvt_f32_f16_e64 v12, |v12| ; CI-NEXT: v_cmp_ngt_f32_e32 vcc, v13, v12 ; CI-NEXT: s_cbranch_vccz .LBB10_26 -; CI-NEXT: ; %bb.25: ; %frem.else86 +; CI-NEXT: ; %bb.25: ; %frem.else ; CI-NEXT: v_cvt_f32_f16_e32 v11, v11 ; CI-NEXT: s_brev_b32 s0, -2 ; CI-NEXT: v_bfi_b32 v14, s0, 0, v0 @@ -7869,7 +7869,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB10_32 ; CI-NEXT: .LBB10_26: ; CI-NEXT: ; implicit-def: $vgpr11 -; CI-NEXT: .LBB10_27: ; %frem.compute85 +; CI-NEXT: .LBB10_27: ; %frem.compute ; CI-NEXT: v_frexp_exp_i32_f32_e32 v16, v13 ; CI-NEXT: v_frexp_mant_f32_e32 v11, v13 ; CI-NEXT: v_frexp_mant_f32_e32 v13, v12 @@ -7894,10 +7894,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v13 ; CI-NEXT: v_div_fixup_f32 v15, v15, v12, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB10_31 -; CI-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; CI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; CI-NEXT: v_sub_i32_e32 v13, vcc, v16, v17 ; CI-NEXT: v_add_i32_e32 v13, vcc, 11, v13 -; CI-NEXT: .LBB10_29: ; %frem.loop_body93 +; CI-NEXT: .LBB10_29: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v16, v14 ; CI-NEXT: v_mul_f32_e32 v14, v16, v15 @@ -7912,7 +7912,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB10_29 ; CI-NEXT: ; %bb.30: ; %Flow ; CI-NEXT: v_mov_b32_e32 v14, v16 -; CI-NEXT: .LBB10_31: ; %frem.loop_exit94 +; CI-NEXT: .LBB10_31: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v13, vcc, -10, v13 ; CI-NEXT: v_ldexp_f32_e32 v13, v14, v13 ; CI-NEXT: v_mul_f32_e32 v14, v13, v15 @@ -8001,7 +8001,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cvt_f32_f16_e64 v5, |v2| ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v6, v5 ; VI-NEXT: s_cbranch_vccz .LBB10_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else86 ; VI-NEXT: s_movk_i32 s2, 0x7fff ; VI-NEXT: v_bfi_b32 v4, s2, 0, v0 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v6, v5 @@ -8010,7 +8010,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB10_8 ; VI-NEXT: .LBB10_2: ; VI-NEXT: ; implicit-def: $vgpr4 -; VI-NEXT: .LBB10_3: ; %frem.compute +; VI-NEXT: .LBB10_3: ; %frem.compute85 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v9, v6 ; VI-NEXT: v_frexp_mant_f32_e32 v4, v6 ; VI-NEXT: v_frexp_mant_f32_e32 v6, v5 @@ -8035,10 +8035,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v6 ; VI-NEXT: v_div_fixup_f32 v8, v8, v5, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB10_7 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; VI-NEXT: v_sub_u32_e32 v6, vcc, v9, v10 ; VI-NEXT: v_add_u32_e32 v6, vcc, 11, v6 -; VI-NEXT: .LBB10_5: ; %frem.loop_body +; VI-NEXT: .LBB10_5: ; %frem.loop_body93 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v9, v7 ; VI-NEXT: v_mul_f32_e32 v7, v9, v8 @@ -8053,7 +8053,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB10_5 ; VI-NEXT: ; %bb.6: ; %Flow133 ; VI-NEXT: v_mov_b32_e32 v7, v9 -; VI-NEXT: .LBB10_7: ; %frem.loop_exit +; VI-NEXT: .LBB10_7: ; %frem.loop_exit94 ; VI-NEXT: v_add_u32_e32 v6, vcc, -10, v6 ; VI-NEXT: v_ldexp_f32 v6, v7, v6 ; VI-NEXT: v_mul_f32_e32 v7, v6, v8 @@ -8073,7 +8073,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cvt_f32_f16_e64 v8, |v6| ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v9, v8 ; VI-NEXT: s_cbranch_vccz .LBB10_10 -; VI-NEXT: ; %bb.9: ; %frem.else20 +; VI-NEXT: ; %bb.9: ; %frem.else53 ; VI-NEXT: s_movk_i32 s2, 0x7fff ; VI-NEXT: v_bfi_b32 v7, s2, 0, v5 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v9, v8 @@ -8082,7 +8082,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB10_16 ; VI-NEXT: .LBB10_10: ; VI-NEXT: ; implicit-def: $vgpr7 -; VI-NEXT: .LBB10_11: ; %frem.compute19 +; VI-NEXT: .LBB10_11: ; %frem.compute52 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v12, v9 ; VI-NEXT: v_frexp_mant_f32_e32 v7, v9 ; VI-NEXT: v_frexp_mant_f32_e32 v9, v8 @@ -8107,10 +8107,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v9 ; VI-NEXT: v_div_fixup_f32 v11, v11, v8, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB10_15 -; VI-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; VI-NEXT: v_sub_u32_e32 v9, vcc, v12, v13 ; VI-NEXT: v_add_u32_e32 v9, vcc, 11, v9 -; VI-NEXT: .LBB10_13: ; %frem.loop_body27 +; VI-NEXT: .LBB10_13: ; %frem.loop_body60 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v12, v10 ; VI-NEXT: v_mul_f32_e32 v10, v12, v11 @@ -8125,7 +8125,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB10_13 ; VI-NEXT: ; %bb.14: ; %Flow129 ; VI-NEXT: v_mov_b32_e32 v10, v12 -; VI-NEXT: .LBB10_15: ; %frem.loop_exit28 +; VI-NEXT: .LBB10_15: ; %frem.loop_exit61 ; VI-NEXT: v_add_u32_e32 v9, vcc, -10, v9 ; VI-NEXT: v_ldexp_f32 v9, v10, v9 ; VI-NEXT: v_mul_f32_e32 v10, v9, v11 @@ -8143,7 +8143,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cvt_f32_f16_e64 v9, |v3| ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v10, v9 ; VI-NEXT: s_cbranch_vccz .LBB10_18 -; VI-NEXT: ; %bb.17: ; %frem.else53 +; VI-NEXT: ; %bb.17: ; %frem.else20 ; VI-NEXT: s_movk_i32 s2, 0x7fff ; VI-NEXT: v_bfi_b32 v8, s2, 0, v1 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v10, v9 @@ -8152,7 +8152,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB10_24 ; VI-NEXT: .LBB10_18: ; VI-NEXT: ; implicit-def: $vgpr8 -; VI-NEXT: .LBB10_19: ; %frem.compute52 +; VI-NEXT: .LBB10_19: ; %frem.compute19 ; VI-NEXT: v_frexp_exp_i32_f32_e32 v13, v10 ; VI-NEXT: v_frexp_mant_f32_e32 v8, v10 ; VI-NEXT: v_frexp_mant_f32_e32 v10, v9 @@ -8177,10 +8177,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v10 ; VI-NEXT: v_div_fixup_f32 v12, v12, v9, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB10_23 -; VI-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; VI-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; VI-NEXT: v_sub_u32_e32 v10, vcc, v13, v14 ; VI-NEXT: v_add_u32_e32 v10, vcc, 11, v10 -; VI-NEXT: .LBB10_21: ; %frem.loop_body60 +; VI-NEXT: .LBB10_21: ; %frem.loop_body27 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v13, v11 ; VI-NEXT: v_mul_f32_e32 v11, v13, v12 @@ -8195,7 +8195,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB10_21 ; VI-NEXT: ; %bb.22: ; %Flow125 ; VI-NEXT: v_mov_b32_e32 v11, v13 -; VI-NEXT: .LBB10_23: ; %frem.loop_exit61 +; VI-NEXT: .LBB10_23: ; %frem.loop_exit28 ; VI-NEXT: v_add_u32_e32 v10, vcc, -10, v10 ; VI-NEXT: v_ldexp_f32 v10, v11, v10 ; VI-NEXT: v_mul_f32_e32 v11, v10, v12 @@ -8215,7 +8215,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cvt_f32_f16_e64 v12, |v10| ; VI-NEXT: v_cmp_ngt_f32_e32 vcc, v13, v12 ; VI-NEXT: s_cbranch_vccz .LBB10_26 -; VI-NEXT: ; %bb.25: ; %frem.else86 +; VI-NEXT: ; %bb.25: ; %frem.else ; VI-NEXT: s_movk_i32 s2, 0x7fff ; VI-NEXT: v_bfi_b32 v11, s2, 0, v9 ; VI-NEXT: v_cmp_eq_f32_e32 vcc, v13, v12 @@ -8224,7 +8224,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB10_32 ; VI-NEXT: .LBB10_26: ; VI-NEXT: ; implicit-def: $vgpr11 -; VI-NEXT: .LBB10_27: ; %frem.compute85 +; VI-NEXT: .LBB10_27: ; %frem.compute ; VI-NEXT: v_frexp_exp_i32_f32_e32 v16, v13 ; VI-NEXT: v_frexp_mant_f32_e32 v11, v13 ; VI-NEXT: v_frexp_mant_f32_e32 v13, v12 @@ -8249,10 +8249,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 12, v13 ; VI-NEXT: v_div_fixup_f32 v15, v15, v12, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB10_31 -; VI-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; VI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; VI-NEXT: v_sub_u32_e32 v13, vcc, v16, v17 ; VI-NEXT: v_add_u32_e32 v13, vcc, 11, v13 -; VI-NEXT: .LBB10_29: ; %frem.loop_body93 +; VI-NEXT: .LBB10_29: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v16, v14 ; VI-NEXT: v_mul_f32_e32 v14, v16, v15 @@ -8267,7 +8267,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB10_29 ; VI-NEXT: ; %bb.30: ; %Flow ; VI-NEXT: v_mov_b32_e32 v14, v16 -; VI-NEXT: .LBB10_31: ; %frem.loop_exit94 +; VI-NEXT: .LBB10_31: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v13, vcc, -10, v13 ; VI-NEXT: v_ldexp_f32 v13, v14, v13 ; VI-NEXT: v_mul_f32_e32 v14, v13, v15 @@ -8320,7 +8320,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cvt_f32_f16_e64 v5, |v0| ; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, v6, v5 ; GFX9-NEXT: s_cbranch_vccz .LBB10_2 -; GFX9-NEXT: ; %bb.1: ; %frem.else +; GFX9-NEXT: ; %bb.1: ; %frem.else86 ; GFX9-NEXT: s_movk_i32 s2, 0x7fff ; GFX9-NEXT: v_bfi_b32 v4, s2, 0, v2 ; GFX9-NEXT: v_cmp_eq_f32_e32 vcc, v6, v5 @@ -8329,7 +8329,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB10_8 ; GFX9-NEXT: .LBB10_2: ; GFX9-NEXT: ; implicit-def: $vgpr4 -; GFX9-NEXT: .LBB10_3: ; %frem.compute +; GFX9-NEXT: .LBB10_3: ; %frem.compute85 ; GFX9-NEXT: v_frexp_exp_i32_f32_e32 v9, v6 ; GFX9-NEXT: v_frexp_mant_f32_e32 v4, v6 ; GFX9-NEXT: v_frexp_mant_f32_e32 v6, v5 @@ -8354,10 +8354,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 12, v6 ; GFX9-NEXT: v_div_fixup_f32 v8, v8, v5, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB10_7 -; GFX9-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX9-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; GFX9-NEXT: v_sub_u32_e32 v6, v9, v10 ; GFX9-NEXT: v_add_u32_e32 v6, 11, v6 -; GFX9-NEXT: .LBB10_5: ; %frem.loop_body +; GFX9-NEXT: .LBB10_5: ; %frem.loop_body93 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v9, v7 ; GFX9-NEXT: v_mul_f32_e32 v7, v9, v8 @@ -8372,7 +8372,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB10_5 ; GFX9-NEXT: ; %bb.6: ; %Flow133 ; GFX9-NEXT: v_mov_b32_e32 v7, v9 -; GFX9-NEXT: .LBB10_7: ; %frem.loop_exit +; GFX9-NEXT: .LBB10_7: ; %frem.loop_exit94 ; GFX9-NEXT: v_add_u32_e32 v6, -10, v6 ; GFX9-NEXT: v_ldexp_f32 v6, v7, v6 ; GFX9-NEXT: v_mul_f32_e32 v7, v6, v8 @@ -8391,7 +8391,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cvt_f32_f16_sdwa v7, |v0| dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 ; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, v8, v7 ; GFX9-NEXT: s_cbranch_vccz .LBB10_10 -; GFX9-NEXT: ; %bb.9: ; %frem.else20 +; GFX9-NEXT: ; %bb.9: ; %frem.else53 ; GFX9-NEXT: s_movk_i32 s2, 0x7fff ; GFX9-NEXT: v_bfi_b32 v6, s2, 0, v5 ; GFX9-NEXT: v_cmp_eq_f32_e32 vcc, v8, v7 @@ -8400,7 +8400,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB10_16 ; GFX9-NEXT: .LBB10_10: ; GFX9-NEXT: ; implicit-def: $vgpr6 -; GFX9-NEXT: .LBB10_11: ; %frem.compute19 +; GFX9-NEXT: .LBB10_11: ; %frem.compute52 ; GFX9-NEXT: v_frexp_exp_i32_f32_e32 v11, v8 ; GFX9-NEXT: v_frexp_mant_f32_e32 v6, v8 ; GFX9-NEXT: v_frexp_mant_f32_e32 v8, v7 @@ -8425,10 +8425,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 12, v8 ; GFX9-NEXT: v_div_fixup_f32 v10, v10, v7, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB10_15 -; GFX9-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX9-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; GFX9-NEXT: v_sub_u32_e32 v8, v11, v12 ; GFX9-NEXT: v_add_u32_e32 v8, 11, v8 -; GFX9-NEXT: .LBB10_13: ; %frem.loop_body27 +; GFX9-NEXT: .LBB10_13: ; %frem.loop_body60 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v11, v9 ; GFX9-NEXT: v_mul_f32_e32 v9, v11, v10 @@ -8443,7 +8443,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB10_13 ; GFX9-NEXT: ; %bb.14: ; %Flow129 ; GFX9-NEXT: v_mov_b32_e32 v9, v11 -; GFX9-NEXT: .LBB10_15: ; %frem.loop_exit28 +; GFX9-NEXT: .LBB10_15: ; %frem.loop_exit61 ; GFX9-NEXT: v_add_u32_e32 v8, -10, v8 ; GFX9-NEXT: v_ldexp_f32 v8, v9, v8 ; GFX9-NEXT: v_mul_f32_e32 v9, v8, v10 @@ -8461,7 +8461,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cvt_f32_f16_e64 v8, |v1| ; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, v9, v8 ; GFX9-NEXT: s_cbranch_vccz .LBB10_18 -; GFX9-NEXT: ; %bb.17: ; %frem.else53 +; GFX9-NEXT: ; %bb.17: ; %frem.else20 ; GFX9-NEXT: s_movk_i32 s2, 0x7fff ; GFX9-NEXT: v_bfi_b32 v7, s2, 0, v3 ; GFX9-NEXT: v_cmp_eq_f32_e32 vcc, v9, v8 @@ -8470,7 +8470,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB10_24 ; GFX9-NEXT: .LBB10_18: ; GFX9-NEXT: ; implicit-def: $vgpr7 -; GFX9-NEXT: .LBB10_19: ; %frem.compute52 +; GFX9-NEXT: .LBB10_19: ; %frem.compute19 ; GFX9-NEXT: v_frexp_exp_i32_f32_e32 v12, v9 ; GFX9-NEXT: v_frexp_mant_f32_e32 v7, v9 ; GFX9-NEXT: v_frexp_mant_f32_e32 v9, v8 @@ -8495,10 +8495,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 12, v9 ; GFX9-NEXT: v_div_fixup_f32 v11, v11, v8, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB10_23 -; GFX9-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; GFX9-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; GFX9-NEXT: v_sub_u32_e32 v9, v12, v13 ; GFX9-NEXT: v_add_u32_e32 v9, 11, v9 -; GFX9-NEXT: .LBB10_21: ; %frem.loop_body60 +; GFX9-NEXT: .LBB10_21: ; %frem.loop_body27 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v12, v10 ; GFX9-NEXT: v_mul_f32_e32 v10, v12, v11 @@ -8513,7 +8513,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB10_21 ; GFX9-NEXT: ; %bb.22: ; %Flow125 ; GFX9-NEXT: v_mov_b32_e32 v10, v12 -; GFX9-NEXT: .LBB10_23: ; %frem.loop_exit61 +; GFX9-NEXT: .LBB10_23: ; %frem.loop_exit28 ; GFX9-NEXT: v_add_u32_e32 v9, -10, v9 ; GFX9-NEXT: v_ldexp_f32 v9, v10, v9 ; GFX9-NEXT: v_mul_f32_e32 v10, v9, v11 @@ -8532,7 +8532,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cvt_f32_f16_sdwa v10, |v1| dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 ; GFX9-NEXT: v_cmp_ngt_f32_e32 vcc, v11, v10 ; GFX9-NEXT: s_cbranch_vccz .LBB10_26 -; GFX9-NEXT: ; %bb.25: ; %frem.else86 +; GFX9-NEXT: ; %bb.25: ; %frem.else ; GFX9-NEXT: s_movk_i32 s2, 0x7fff ; GFX9-NEXT: v_bfi_b32 v9, s2, 0, v8 ; GFX9-NEXT: v_cmp_eq_f32_e32 vcc, v11, v10 @@ -8541,7 +8541,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB10_32 ; GFX9-NEXT: .LBB10_26: ; GFX9-NEXT: ; implicit-def: $vgpr9 -; GFX9-NEXT: .LBB10_27: ; %frem.compute85 +; GFX9-NEXT: .LBB10_27: ; %frem.compute ; GFX9-NEXT: v_frexp_exp_i32_f32_e32 v14, v11 ; GFX9-NEXT: v_frexp_mant_f32_e32 v9, v11 ; GFX9-NEXT: v_frexp_mant_f32_e32 v11, v10 @@ -8566,10 +8566,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 12, v11 ; GFX9-NEXT: v_div_fixup_f32 v13, v13, v10, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB10_31 -; GFX9-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; GFX9-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX9-NEXT: v_sub_u32_e32 v11, v14, v15 ; GFX9-NEXT: v_add_u32_e32 v11, 11, v11 -; GFX9-NEXT: .LBB10_29: ; %frem.loop_body93 +; GFX9-NEXT: .LBB10_29: ; %frem.loop_body ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v14, v12 ; GFX9-NEXT: v_mul_f32_e32 v12, v14, v13 @@ -8584,7 +8584,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB10_29 ; GFX9-NEXT: ; %bb.30: ; %Flow ; GFX9-NEXT: v_mov_b32_e32 v12, v14 -; GFX9-NEXT: .LBB10_31: ; %frem.loop_exit94 +; GFX9-NEXT: .LBB10_31: ; %frem.loop_exit ; GFX9-NEXT: v_add_u32_e32 v11, -10, v11 ; GFX9-NEXT: v_ldexp_f32 v11, v12, v11 ; GFX9-NEXT: v_mul_f32_e32 v12, v11, v13 @@ -8640,7 +8640,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cvt_f32_f16_e64 v5, |v0| ; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v6, v5 ; GFX10-NEXT: s_cbranch_vccz .LBB10_2 -; GFX10-NEXT: ; %bb.1: ; %frem.else +; GFX10-NEXT: ; %bb.1: ; %frem.else86 ; GFX10-NEXT: v_bfi_b32 v4, 0x7fff, 0, v2 ; GFX10-NEXT: v_cmp_eq_f32_e32 vcc_lo, v6, v5 ; GFX10-NEXT: v_cndmask_b32_e32 v4, v2, v4, vcc_lo @@ -8648,7 +8648,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB10_8 ; GFX10-NEXT: .LBB10_2: ; GFX10-NEXT: ; implicit-def: $vgpr4 -; GFX10-NEXT: .LBB10_3: ; %frem.compute +; GFX10-NEXT: .LBB10_3: ; %frem.compute85 ; GFX10-NEXT: v_frexp_mant_f32_e32 v4, v6 ; GFX10-NEXT: v_frexp_mant_f32_e32 v8, v5 ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v7, v6 @@ -8675,10 +8675,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v8 ; GFX10-NEXT: v_div_fixup_f32 v7, v7, v5, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB10_7 -; GFX10-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX10-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 11 -; GFX10-NEXT: .LBB10_5: ; %frem.loop_body +; GFX10-NEXT: .LBB10_5: ; %frem.loop_body93 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v9, v6 ; GFX10-NEXT: s_add_i32 s2, s2, -11 @@ -8694,7 +8694,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.6: ; %Flow133 ; GFX10-NEXT: v_mov_b32_e32 v8, s2 ; GFX10-NEXT: v_mov_b32_e32 v6, v9 -; GFX10-NEXT: .LBB10_7: ; %frem.loop_exit +; GFX10-NEXT: .LBB10_7: ; %frem.loop_exit94 ; GFX10-NEXT: v_add_nc_u32_e32 v8, -10, v8 ; GFX10-NEXT: v_ldexp_f32 v6, v6, v8 ; GFX10-NEXT: v_mul_f32_e32 v7, v6, v7 @@ -8712,7 +8712,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cvt_f32_f16_e64 v8, |v5| ; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v8, v7 ; GFX10-NEXT: s_cbranch_vccz .LBB10_10 -; GFX10-NEXT: ; %bb.9: ; %frem.else20 +; GFX10-NEXT: ; %bb.9: ; %frem.else53 ; GFX10-NEXT: v_bfi_b32 v6, 0x7fff, 0, v5 ; GFX10-NEXT: v_cmp_eq_f32_e32 vcc_lo, v8, v7 ; GFX10-NEXT: v_cndmask_b32_e32 v6, v5, v6, vcc_lo @@ -8720,7 +8720,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB10_16 ; GFX10-NEXT: .LBB10_10: ; GFX10-NEXT: ; implicit-def: $vgpr6 -; GFX10-NEXT: .LBB10_11: ; %frem.compute19 +; GFX10-NEXT: .LBB10_11: ; %frem.compute52 ; GFX10-NEXT: v_frexp_mant_f32_e32 v6, v8 ; GFX10-NEXT: v_frexp_mant_f32_e32 v10, v7 ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v9, v8 @@ -8747,10 +8747,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v10 ; GFX10-NEXT: v_div_fixup_f32 v9, v9, v7, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB10_15 -; GFX10-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX10-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 11 -; GFX10-NEXT: .LBB10_13: ; %frem.loop_body27 +; GFX10-NEXT: .LBB10_13: ; %frem.loop_body60 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v11, v8 ; GFX10-NEXT: s_add_i32 s2, s2, -11 @@ -8766,7 +8766,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.14: ; %Flow129 ; GFX10-NEXT: v_mov_b32_e32 v10, s2 ; GFX10-NEXT: v_mov_b32_e32 v8, v11 -; GFX10-NEXT: .LBB10_15: ; %frem.loop_exit28 +; GFX10-NEXT: .LBB10_15: ; %frem.loop_exit61 ; GFX10-NEXT: v_add_nc_u32_e32 v10, -10, v10 ; GFX10-NEXT: v_ldexp_f32 v8, v8, v10 ; GFX10-NEXT: v_mul_f32_e32 v9, v8, v9 @@ -8783,7 +8783,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cvt_f32_f16_e64 v8, |v1| ; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v9, v8 ; GFX10-NEXT: s_cbranch_vccz .LBB10_18 -; GFX10-NEXT: ; %bb.17: ; %frem.else53 +; GFX10-NEXT: ; %bb.17: ; %frem.else20 ; GFX10-NEXT: v_bfi_b32 v7, 0x7fff, 0, v3 ; GFX10-NEXT: v_cmp_eq_f32_e32 vcc_lo, v9, v8 ; GFX10-NEXT: v_cndmask_b32_e32 v7, v3, v7, vcc_lo @@ -8791,7 +8791,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB10_24 ; GFX10-NEXT: .LBB10_18: ; GFX10-NEXT: ; implicit-def: $vgpr7 -; GFX10-NEXT: .LBB10_19: ; %frem.compute52 +; GFX10-NEXT: .LBB10_19: ; %frem.compute19 ; GFX10-NEXT: v_frexp_mant_f32_e32 v7, v9 ; GFX10-NEXT: v_frexp_mant_f32_e32 v11, v8 ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v10, v9 @@ -8818,10 +8818,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v11 ; GFX10-NEXT: v_div_fixup_f32 v10, v10, v8, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB10_23 -; GFX10-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; GFX10-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 11 -; GFX10-NEXT: .LBB10_21: ; %frem.loop_body60 +; GFX10-NEXT: .LBB10_21: ; %frem.loop_body27 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v12, v9 ; GFX10-NEXT: s_add_i32 s2, s2, -11 @@ -8837,7 +8837,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.22: ; %Flow125 ; GFX10-NEXT: v_mov_b32_e32 v11, s2 ; GFX10-NEXT: v_mov_b32_e32 v9, v12 -; GFX10-NEXT: .LBB10_23: ; %frem.loop_exit61 +; GFX10-NEXT: .LBB10_23: ; %frem.loop_exit28 ; GFX10-NEXT: v_add_nc_u32_e32 v11, -10, v11 ; GFX10-NEXT: v_ldexp_f32 v9, v9, v11 ; GFX10-NEXT: v_mul_f32_e32 v10, v9, v10 @@ -8855,7 +8855,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cvt_f32_f16_e64 v11, |v8| ; GFX10-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v11, v10 ; GFX10-NEXT: s_cbranch_vccz .LBB10_26 -; GFX10-NEXT: ; %bb.25: ; %frem.else86 +; GFX10-NEXT: ; %bb.25: ; %frem.else ; GFX10-NEXT: v_bfi_b32 v9, 0x7fff, 0, v8 ; GFX10-NEXT: v_cmp_eq_f32_e32 vcc_lo, v11, v10 ; GFX10-NEXT: v_cndmask_b32_e32 v9, v8, v9, vcc_lo @@ -8863,7 +8863,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB10_32 ; GFX10-NEXT: .LBB10_26: ; GFX10-NEXT: ; implicit-def: $vgpr9 -; GFX10-NEXT: .LBB10_27: ; %frem.compute85 +; GFX10-NEXT: .LBB10_27: ; %frem.compute ; GFX10-NEXT: v_frexp_mant_f32_e32 v9, v11 ; GFX10-NEXT: v_frexp_mant_f32_e32 v13, v10 ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v12, v11 @@ -8890,10 +8890,10 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v13 ; GFX10-NEXT: v_div_fixup_f32 v12, v12, v10, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB10_31 -; GFX10-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; GFX10-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 11 -; GFX10-NEXT: .LBB10_29: ; %frem.loop_body93 +; GFX10-NEXT: .LBB10_29: ; %frem.loop_body ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v14, v11 ; GFX10-NEXT: s_add_i32 s2, s2, -11 @@ -8909,7 +8909,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.30: ; %Flow ; GFX10-NEXT: v_mov_b32_e32 v13, s2 ; GFX10-NEXT: v_mov_b32_e32 v11, v14 -; GFX10-NEXT: .LBB10_31: ; %frem.loop_exit94 +; GFX10-NEXT: .LBB10_31: ; %frem.loop_exit ; GFX10-NEXT: v_add_nc_u32_e32 v13, -10, v13 ; GFX10-NEXT: v_ldexp_f32 v11, v11, v13 ; GFX10-NEXT: v_mul_f32_e32 v12, v11, v12 @@ -8963,7 +8963,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v6, v5 ; GFX11-TRUE16-NEXT: s_cbranch_vccz .LBB10_2 -; GFX11-TRUE16-NEXT: ; %bb.1: ; %frem.else +; GFX11-TRUE16-NEXT: ; %bb.1: ; %frem.else86 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v0.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, 0 ; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v6, v5 @@ -8974,7 +8974,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_branch .LBB10_8 ; GFX11-TRUE16-NEXT: .LBB10_2: ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr4 -; GFX11-TRUE16-NEXT: .LBB10_3: ; %frem.compute +; GFX11-TRUE16-NEXT: .LBB10_3: ; %frem.compute85 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v4, v6 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v8, v5 ; GFX11-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v7, v6 @@ -9010,11 +9010,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_div_fixup_f32 v7, v7, v5, 1.0 ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB10_7 -; GFX11-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX11-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; GFX11-TRUE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-TRUE16-NEXT: .LBB10_5: ; %frem.loop_body +; GFX11-TRUE16-NEXT: .LBB10_5: ; %frem.loop_body93 ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v9, v6 @@ -9034,7 +9034,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: ; %bb.6: ; %Flow133 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v8, s2 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v6, v9 -; GFX11-TRUE16-NEXT: .LBB10_7: ; %frem.loop_exit +; GFX11-TRUE16-NEXT: .LBB10_7: ; %frem.loop_exit94 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v8, -10, v8 ; GFX11-TRUE16-NEXT: v_ldexp_f32 v6, v6, v8 @@ -9061,7 +9061,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v9, v8 ; GFX11-TRUE16-NEXT: s_cbranch_vccz .LBB10_10 -; GFX11-TRUE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX11-TRUE16-NEXT: ; %bb.9: ; %frem.else53 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v5.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, 0 ; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v9, v8 @@ -9072,7 +9072,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_branch .LBB10_16 ; GFX11-TRUE16-NEXT: .LBB10_10: ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7 -; GFX11-TRUE16-NEXT: .LBB10_11: ; %frem.compute19 +; GFX11-TRUE16-NEXT: .LBB10_11: ; %frem.compute52 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v7, v9 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v11, v8 ; GFX11-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v10, v9 @@ -9108,11 +9108,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_div_fixup_f32 v10, v10, v8, 1.0 ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB10_15 -; GFX11-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX11-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; GFX11-TRUE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-TRUE16-NEXT: .LBB10_13: ; %frem.loop_body27 +; GFX11-TRUE16-NEXT: .LBB10_13: ; %frem.loop_body60 ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v12, v9 @@ -9132,7 +9132,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: ; %bb.14: ; %Flow129 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v11, s2 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v9, v12 -; GFX11-TRUE16-NEXT: .LBB10_15: ; %frem.loop_exit28 +; GFX11-TRUE16-NEXT: .LBB10_15: ; %frem.loop_exit61 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v11, -10, v11 ; GFX11-TRUE16-NEXT: v_ldexp_f32 v9, v9, v11 @@ -9156,7 +9156,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v10, v9 ; GFX11-TRUE16-NEXT: s_cbranch_vccz .LBB10_18 -; GFX11-TRUE16-NEXT: ; %bb.17: ; %frem.else53 +; GFX11-TRUE16-NEXT: ; %bb.17: ; %frem.else20 ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v1.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, 0 ; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v10, v9 @@ -9167,7 +9167,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_branch .LBB10_24 ; GFX11-TRUE16-NEXT: .LBB10_18: ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr8 -; GFX11-TRUE16-NEXT: .LBB10_19: ; %frem.compute52 +; GFX11-TRUE16-NEXT: .LBB10_19: ; %frem.compute19 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v8, v10 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v12, v9 ; GFX11-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v11, v10 @@ -9203,11 +9203,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_div_fixup_f32 v11, v11, v9, 1.0 ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB10_23 -; GFX11-TRUE16-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; GFX11-TRUE16-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; GFX11-TRUE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-TRUE16-NEXT: .LBB10_21: ; %frem.loop_body60 +; GFX11-TRUE16-NEXT: .LBB10_21: ; %frem.loop_body27 ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v10 @@ -9227,7 +9227,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: ; %bb.22: ; %Flow125 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v12, s2 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v10, v13 -; GFX11-TRUE16-NEXT: .LBB10_23: ; %frem.loop_exit61 +; GFX11-TRUE16-NEXT: .LBB10_23: ; %frem.loop_exit28 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v12, -10, v12 ; GFX11-TRUE16-NEXT: v_ldexp_f32 v10, v10, v12 @@ -9254,7 +9254,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v13, v12 ; GFX11-TRUE16-NEXT: s_cbranch_vccz .LBB10_26 -; GFX11-TRUE16-NEXT: ; %bb.25: ; %frem.else86 +; GFX11-TRUE16-NEXT: ; %bb.25: ; %frem.else ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v9.l ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, 0 ; GFX11-TRUE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v13, v12 @@ -9265,7 +9265,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_branch .LBB10_32 ; GFX11-TRUE16-NEXT: .LBB10_26: ; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11 -; GFX11-TRUE16-NEXT: .LBB10_27: ; %frem.compute85 +; GFX11-TRUE16-NEXT: .LBB10_27: ; %frem.compute ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v11, v13 ; GFX11-TRUE16-NEXT: v_frexp_mant_f32_e32 v15, v12 ; GFX11-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v14, v13 @@ -9301,11 +9301,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_div_fixup_f32 v14, v14, v12, 1.0 ; GFX11-TRUE16-NEXT: s_cbranch_vccnz .LBB10_31 -; GFX11-TRUE16-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; GFX11-TRUE16-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX11-TRUE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-TRUE16-NEXT: .LBB10_29: ; %frem.loop_body93 +; GFX11-TRUE16-NEXT: .LBB10_29: ; %frem.loop_body ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v16, v13 @@ -9325,7 +9325,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-TRUE16-NEXT: ; %bb.30: ; %Flow ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v15, s2 ; GFX11-TRUE16-NEXT: v_mov_b32_e32 v13, v16 -; GFX11-TRUE16-NEXT: .LBB10_31: ; %frem.loop_exit94 +; GFX11-TRUE16-NEXT: .LBB10_31: ; %frem.loop_exit ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-TRUE16-NEXT: v_add_nc_u32_e32 v15, -10, v15 ; GFX11-TRUE16-NEXT: v_ldexp_f32 v13, v13, v15 @@ -9388,7 +9388,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v6, v5 ; GFX11-FAKE16-NEXT: s_cbranch_vccz .LBB10_2 -; GFX11-FAKE16-NEXT: ; %bb.1: ; %frem.else +; GFX11-FAKE16-NEXT: ; %bb.1: ; %frem.else86 ; GFX11-FAKE16-NEXT: v_bfi_b32 v4, 0x7fff, 0, v0 ; GFX11-FAKE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v6, v5 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -9397,7 +9397,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_branch .LBB10_8 ; GFX11-FAKE16-NEXT: .LBB10_2: ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr4 -; GFX11-FAKE16-NEXT: .LBB10_3: ; %frem.compute +; GFX11-FAKE16-NEXT: .LBB10_3: ; %frem.compute85 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v4, v6 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v8, v5 ; GFX11-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v7, v6 @@ -9433,11 +9433,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-FAKE16-NEXT: v_div_fixup_f32 v7, v7, v5, 1.0 ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB10_7 -; GFX11-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX11-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; GFX11-FAKE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-FAKE16-NEXT: .LBB10_5: ; %frem.loop_body +; GFX11-FAKE16-NEXT: .LBB10_5: ; %frem.loop_body93 ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v9, v6 @@ -9457,7 +9457,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: ; %bb.6: ; %Flow133 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v8, s2 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v6, v9 -; GFX11-FAKE16-NEXT: .LBB10_7: ; %frem.loop_exit +; GFX11-FAKE16-NEXT: .LBB10_7: ; %frem.loop_exit94 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v8, -10, v8 ; GFX11-FAKE16-NEXT: v_ldexp_f32 v6, v6, v8 @@ -9483,7 +9483,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v9, v8 ; GFX11-FAKE16-NEXT: s_cbranch_vccz .LBB10_10 -; GFX11-FAKE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX11-FAKE16-NEXT: ; %bb.9: ; %frem.else53 ; GFX11-FAKE16-NEXT: v_bfi_b32 v7, 0x7fff, 0, v5 ; GFX11-FAKE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v9, v8 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -9492,7 +9492,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_branch .LBB10_16 ; GFX11-FAKE16-NEXT: .LBB10_10: ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr7 -; GFX11-FAKE16-NEXT: .LBB10_11: ; %frem.compute19 +; GFX11-FAKE16-NEXT: .LBB10_11: ; %frem.compute52 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v7, v9 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v11, v8 ; GFX11-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v10, v9 @@ -9528,11 +9528,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-FAKE16-NEXT: v_div_fixup_f32 v10, v10, v8, 1.0 ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB10_15 -; GFX11-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX11-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; GFX11-FAKE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-FAKE16-NEXT: .LBB10_13: ; %frem.loop_body27 +; GFX11-FAKE16-NEXT: .LBB10_13: ; %frem.loop_body60 ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v12, v9 @@ -9552,7 +9552,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: ; %bb.14: ; %Flow129 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v11, s2 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v9, v12 -; GFX11-FAKE16-NEXT: .LBB10_15: ; %frem.loop_exit28 +; GFX11-FAKE16-NEXT: .LBB10_15: ; %frem.loop_exit61 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v11, -10, v11 ; GFX11-FAKE16-NEXT: v_ldexp_f32 v9, v9, v11 @@ -9575,7 +9575,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v10, v9 ; GFX11-FAKE16-NEXT: s_cbranch_vccz .LBB10_18 -; GFX11-FAKE16-NEXT: ; %bb.17: ; %frem.else53 +; GFX11-FAKE16-NEXT: ; %bb.17: ; %frem.else20 ; GFX11-FAKE16-NEXT: v_bfi_b32 v8, 0x7fff, 0, v1 ; GFX11-FAKE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v10, v9 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -9584,7 +9584,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_branch .LBB10_24 ; GFX11-FAKE16-NEXT: .LBB10_18: ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr8 -; GFX11-FAKE16-NEXT: .LBB10_19: ; %frem.compute52 +; GFX11-FAKE16-NEXT: .LBB10_19: ; %frem.compute19 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v8, v10 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v12, v9 ; GFX11-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v11, v10 @@ -9620,11 +9620,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-FAKE16-NEXT: v_div_fixup_f32 v11, v11, v9, 1.0 ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB10_23 -; GFX11-FAKE16-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; GFX11-FAKE16-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; GFX11-FAKE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-FAKE16-NEXT: .LBB10_21: ; %frem.loop_body60 +; GFX11-FAKE16-NEXT: .LBB10_21: ; %frem.loop_body27 ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v13, v10 @@ -9644,7 +9644,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: ; %bb.22: ; %Flow125 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v12, s2 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v10, v13 -; GFX11-FAKE16-NEXT: .LBB10_23: ; %frem.loop_exit61 +; GFX11-FAKE16-NEXT: .LBB10_23: ; %frem.loop_exit28 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v12, -10, v12 ; GFX11-FAKE16-NEXT: v_ldexp_f32 v10, v10, v12 @@ -9670,7 +9670,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_cmp_ngt_f32_e32 vcc_lo, v13, v12 ; GFX11-FAKE16-NEXT: s_cbranch_vccz .LBB10_26 -; GFX11-FAKE16-NEXT: ; %bb.25: ; %frem.else86 +; GFX11-FAKE16-NEXT: ; %bb.25: ; %frem.else ; GFX11-FAKE16-NEXT: v_bfi_b32 v11, 0x7fff, 0, v9 ; GFX11-FAKE16-NEXT: v_cmp_eq_f32_e32 vcc_lo, v13, v12 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -9679,7 +9679,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_branch .LBB10_32 ; GFX11-FAKE16-NEXT: .LBB10_26: ; GFX11-FAKE16-NEXT: ; implicit-def: $vgpr11 -; GFX11-FAKE16-NEXT: .LBB10_27: ; %frem.compute85 +; GFX11-FAKE16-NEXT: .LBB10_27: ; %frem.compute ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v11, v13 ; GFX11-FAKE16-NEXT: v_frexp_mant_f32_e32 v15, v12 ; GFX11-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v14, v13 @@ -9715,11 +9715,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-FAKE16-NEXT: v_div_fixup_f32 v14, v14, v12, 1.0 ; GFX11-FAKE16-NEXT: s_cbranch_vccnz .LBB10_31 -; GFX11-FAKE16-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; GFX11-FAKE16-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX11-FAKE16-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_add_i32 s2, s2, 11 -; GFX11-FAKE16-NEXT: .LBB10_29: ; %frem.loop_body93 +; GFX11-FAKE16-NEXT: .LBB10_29: ; %frem.loop_body ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v16, v13 @@ -9739,7 +9739,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-FAKE16-NEXT: ; %bb.30: ; %Flow ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v15, s2 ; GFX11-FAKE16-NEXT: v_mov_b32_e32 v13, v16 -; GFX11-FAKE16-NEXT: .LBB10_31: ; %frem.loop_exit94 +; GFX11-FAKE16-NEXT: .LBB10_31: ; %frem.loop_exit ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-FAKE16-NEXT: v_add_nc_u32_e32 v15, -10, v15 ; GFX11-FAKE16-NEXT: v_ldexp_f32 v13, v13, v15 @@ -9804,7 +9804,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-TRUE16-NEXT: s_cmp_ngt_f32 s8, s6 ; GFX1150-TRUE16-NEXT: s_cbranch_scc0 .LBB10_2 -; GFX1150-TRUE16-NEXT: ; %bb.1: ; %frem.else +; GFX1150-TRUE16-NEXT: ; %bb.1: ; %frem.else86 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v0.l, s5 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 ; GFX1150-TRUE16-NEXT: s_cmp_eq_f32 s8, s6 @@ -9816,7 +9816,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_branch .LBB10_8 ; GFX1150-TRUE16-NEXT: .LBB10_2: ; GFX1150-TRUE16-NEXT: ; implicit-def: $vgpr0 -; GFX1150-TRUE16-NEXT: .LBB10_3: ; %frem.compute +; GFX1150-TRUE16-NEXT: .LBB10_3: ; %frem.compute85 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v1, s6 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v0, s8 ; GFX1150-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v3, s8 @@ -9851,11 +9851,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v4 ; GFX1150-TRUE16-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1150-TRUE16-NEXT: s_cbranch_vccnz .LBB10_7 -; GFX1150-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1150-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; GFX1150-TRUE16-NEXT: s_sub_i32 s6, s8, s6 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: s_add_i32 s6, s6, 11 -; GFX1150-TRUE16-NEXT: .LBB10_5: ; %frem.loop_body +; GFX1150-TRUE16-NEXT: .LBB10_5: ; %frem.loop_body93 ; GFX1150-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v5, v2 @@ -9877,7 +9877,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: ; %bb.6: ; %Flow133 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v4, s6 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v2, v5 -; GFX1150-TRUE16-NEXT: .LBB10_7: ; %frem.loop_exit +; GFX1150-TRUE16-NEXT: .LBB10_7: ; %frem.loop_exit94 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-TRUE16-NEXT: v_add_nc_u32_e32 v4, -10, v4 ; GFX1150-TRUE16-NEXT: v_ldexp_f32 v2, v2, v4 @@ -9907,7 +9907,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-TRUE16-NEXT: s_cmp_ngt_f32 s10, s9 ; GFX1150-TRUE16-NEXT: s_cbranch_scc0 .LBB10_10 -; GFX1150-TRUE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX1150-TRUE16-NEXT: ; %bb.9: ; %frem.else53 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v1.l, s8 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 ; GFX1150-TRUE16-NEXT: s_cmp_eq_f32 s10, s9 @@ -9919,7 +9919,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_branch .LBB10_16 ; GFX1150-TRUE16-NEXT: .LBB10_10: ; GFX1150-TRUE16-NEXT: ; implicit-def: $vgpr1 -; GFX1150-TRUE16-NEXT: .LBB10_11: ; %frem.compute19 +; GFX1150-TRUE16-NEXT: .LBB10_11: ; %frem.compute52 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v2, s9 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v1, s10 ; GFX1150-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v4, s10 @@ -9954,11 +9954,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v5 ; GFX1150-TRUE16-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1150-TRUE16-NEXT: s_cbranch_vccnz .LBB10_15 -; GFX1150-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX1150-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; GFX1150-TRUE16-NEXT: s_sub_i32 s9, s10, s9 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: s_add_i32 s9, s9, 11 -; GFX1150-TRUE16-NEXT: .LBB10_13: ; %frem.loop_body27 +; GFX1150-TRUE16-NEXT: .LBB10_13: ; %frem.loop_body60 ; GFX1150-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v6, v3 @@ -9980,7 +9980,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: ; %bb.14: ; %Flow129 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v5, s9 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v3, v6 -; GFX1150-TRUE16-NEXT: .LBB10_15: ; %frem.loop_exit28 +; GFX1150-TRUE16-NEXT: .LBB10_15: ; %frem.loop_exit61 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-TRUE16-NEXT: v_add_nc_u32_e32 v5, -10, v5 ; GFX1150-TRUE16-NEXT: v_ldexp_f32 v3, v3, v5 @@ -10008,7 +10008,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-TRUE16-NEXT: s_cmp_ngt_f32 s10, s9 ; GFX1150-TRUE16-NEXT: s_cbranch_scc0 .LBB10_18 -; GFX1150-TRUE16-NEXT: ; %bb.17: ; %frem.else53 +; GFX1150-TRUE16-NEXT: ; %bb.17: ; %frem.else20 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v2.l, s7 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 ; GFX1150-TRUE16-NEXT: s_cmp_eq_f32 s10, s9 @@ -10020,7 +10020,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_branch .LBB10_24 ; GFX1150-TRUE16-NEXT: .LBB10_18: ; GFX1150-TRUE16-NEXT: ; implicit-def: $vgpr2 -; GFX1150-TRUE16-NEXT: .LBB10_19: ; %frem.compute52 +; GFX1150-TRUE16-NEXT: .LBB10_19: ; %frem.compute19 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v3, s9 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v2, s10 ; GFX1150-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v5, s10 @@ -10055,11 +10055,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v6 ; GFX1150-TRUE16-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX1150-TRUE16-NEXT: s_cbranch_vccnz .LBB10_23 -; GFX1150-TRUE16-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; GFX1150-TRUE16-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; GFX1150-TRUE16-NEXT: s_sub_i32 s9, s10, s9 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: s_add_i32 s9, s9, 11 -; GFX1150-TRUE16-NEXT: .LBB10_21: ; %frem.loop_body60 +; GFX1150-TRUE16-NEXT: .LBB10_21: ; %frem.loop_body27 ; GFX1150-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v7, v4 @@ -10081,7 +10081,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: ; %bb.22: ; %Flow125 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v6, s9 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v4, v7 -; GFX1150-TRUE16-NEXT: .LBB10_23: ; %frem.loop_exit61 +; GFX1150-TRUE16-NEXT: .LBB10_23: ; %frem.loop_exit28 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-TRUE16-NEXT: v_add_nc_u32_e32 v6, -10, v6 ; GFX1150-TRUE16-NEXT: v_ldexp_f32 v4, v4, v6 @@ -10111,7 +10111,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-TRUE16-NEXT: s_cmp_ngt_f32 s12, s11 ; GFX1150-TRUE16-NEXT: s_cbranch_scc0 .LBB10_26 -; GFX1150-TRUE16-NEXT: ; %bb.25: ; %frem.else86 +; GFX1150-TRUE16-NEXT: ; %bb.25: ; %frem.else ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v3.l, s10 ; GFX1150-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX1150-TRUE16-NEXT: s_cmp_eq_f32 s12, s11 @@ -10123,7 +10123,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: s_branch .LBB10_32 ; GFX1150-TRUE16-NEXT: .LBB10_26: ; GFX1150-TRUE16-NEXT: ; implicit-def: $vgpr3 -; GFX1150-TRUE16-NEXT: .LBB10_27: ; %frem.compute85 +; GFX1150-TRUE16-NEXT: .LBB10_27: ; %frem.compute ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v4, s11 ; GFX1150-TRUE16-NEXT: v_frexp_mant_f32_e32 v3, s12 ; GFX1150-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v6, s12 @@ -10158,11 +10158,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v7 ; GFX1150-TRUE16-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; GFX1150-TRUE16-NEXT: s_cbranch_vccnz .LBB10_31 -; GFX1150-TRUE16-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; GFX1150-TRUE16-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX1150-TRUE16-NEXT: s_sub_i32 s11, s12, s11 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: s_add_i32 s11, s11, 11 -; GFX1150-TRUE16-NEXT: .LBB10_29: ; %frem.loop_body93 +; GFX1150-TRUE16-NEXT: .LBB10_29: ; %frem.loop_body ; GFX1150-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v8, v5 @@ -10184,7 +10184,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-TRUE16-NEXT: ; %bb.30: ; %Flow ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v7, s11 ; GFX1150-TRUE16-NEXT: v_mov_b32_e32 v5, v8 -; GFX1150-TRUE16-NEXT: .LBB10_31: ; %frem.loop_exit94 +; GFX1150-TRUE16-NEXT: .LBB10_31: ; %frem.loop_exit ; GFX1150-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-TRUE16-NEXT: v_add_nc_u32_e32 v7, -10, v7 ; GFX1150-TRUE16-NEXT: v_ldexp_f32 v5, v5, v7 @@ -10265,7 +10265,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-FAKE16-NEXT: s_cmp_ngt_f32 s8, s6 ; GFX1150-FAKE16-NEXT: s_cbranch_scc0 .LBB10_2 -; GFX1150-FAKE16-NEXT: ; %bb.1: ; %frem.else +; GFX1150-FAKE16-NEXT: ; %bb.1: ; %frem.else86 ; GFX1150-FAKE16-NEXT: s_cmp_eq_f32 s8, s6 ; GFX1150-FAKE16-NEXT: v_bfi_b32 v0, 0x7fff, 0, s5 ; GFX1150-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -10275,7 +10275,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_branch .LBB10_8 ; GFX1150-FAKE16-NEXT: .LBB10_2: ; GFX1150-FAKE16-NEXT: ; implicit-def: $vgpr0 -; GFX1150-FAKE16-NEXT: .LBB10_3: ; %frem.compute +; GFX1150-FAKE16-NEXT: .LBB10_3: ; %frem.compute85 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v1, s6 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v0, s8 ; GFX1150-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v3, s8 @@ -10310,11 +10310,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v4 ; GFX1150-FAKE16-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1150-FAKE16-NEXT: s_cbranch_vccnz .LBB10_7 -; GFX1150-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1150-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; GFX1150-FAKE16-NEXT: s_sub_i32 s6, s8, s6 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: s_add_i32 s6, s6, 11 -; GFX1150-FAKE16-NEXT: .LBB10_5: ; %frem.loop_body +; GFX1150-FAKE16-NEXT: .LBB10_5: ; %frem.loop_body93 ; GFX1150-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v5, v2 @@ -10336,7 +10336,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: ; %bb.6: ; %Flow133 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v4, s6 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v2, v5 -; GFX1150-FAKE16-NEXT: .LBB10_7: ; %frem.loop_exit +; GFX1150-FAKE16-NEXT: .LBB10_7: ; %frem.loop_exit94 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-FAKE16-NEXT: v_add_nc_u32_e32 v4, -10, v4 ; GFX1150-FAKE16-NEXT: v_ldexp_f32 v2, v2, v4 @@ -10365,7 +10365,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-FAKE16-NEXT: s_cmp_ngt_f32 s10, s9 ; GFX1150-FAKE16-NEXT: s_cbranch_scc0 .LBB10_10 -; GFX1150-FAKE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX1150-FAKE16-NEXT: ; %bb.9: ; %frem.else53 ; GFX1150-FAKE16-NEXT: s_cmp_eq_f32 s10, s9 ; GFX1150-FAKE16-NEXT: v_bfi_b32 v1, 0x7fff, 0, s8 ; GFX1150-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -10375,7 +10375,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_branch .LBB10_16 ; GFX1150-FAKE16-NEXT: .LBB10_10: ; GFX1150-FAKE16-NEXT: ; implicit-def: $vgpr1 -; GFX1150-FAKE16-NEXT: .LBB10_11: ; %frem.compute19 +; GFX1150-FAKE16-NEXT: .LBB10_11: ; %frem.compute52 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v2, s9 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v1, s10 ; GFX1150-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v4, s10 @@ -10410,11 +10410,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v5 ; GFX1150-FAKE16-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1150-FAKE16-NEXT: s_cbranch_vccnz .LBB10_15 -; GFX1150-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX1150-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; GFX1150-FAKE16-NEXT: s_sub_i32 s9, s10, s9 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: s_add_i32 s9, s9, 11 -; GFX1150-FAKE16-NEXT: .LBB10_13: ; %frem.loop_body27 +; GFX1150-FAKE16-NEXT: .LBB10_13: ; %frem.loop_body60 ; GFX1150-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v6, v3 @@ -10436,7 +10436,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: ; %bb.14: ; %Flow129 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v5, s9 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v3, v6 -; GFX1150-FAKE16-NEXT: .LBB10_15: ; %frem.loop_exit28 +; GFX1150-FAKE16-NEXT: .LBB10_15: ; %frem.loop_exit61 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-FAKE16-NEXT: v_add_nc_u32_e32 v5, -10, v5 ; GFX1150-FAKE16-NEXT: v_ldexp_f32 v3, v3, v5 @@ -10463,7 +10463,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-FAKE16-NEXT: s_cmp_ngt_f32 s10, s9 ; GFX1150-FAKE16-NEXT: s_cbranch_scc0 .LBB10_18 -; GFX1150-FAKE16-NEXT: ; %bb.17: ; %frem.else53 +; GFX1150-FAKE16-NEXT: ; %bb.17: ; %frem.else20 ; GFX1150-FAKE16-NEXT: s_cmp_eq_f32 s10, s9 ; GFX1150-FAKE16-NEXT: v_bfi_b32 v2, 0x7fff, 0, s7 ; GFX1150-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -10473,7 +10473,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_branch .LBB10_24 ; GFX1150-FAKE16-NEXT: .LBB10_18: ; GFX1150-FAKE16-NEXT: ; implicit-def: $vgpr2 -; GFX1150-FAKE16-NEXT: .LBB10_19: ; %frem.compute52 +; GFX1150-FAKE16-NEXT: .LBB10_19: ; %frem.compute19 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v3, s9 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v2, s10 ; GFX1150-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v5, s10 @@ -10508,11 +10508,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v6 ; GFX1150-FAKE16-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX1150-FAKE16-NEXT: s_cbranch_vccnz .LBB10_23 -; GFX1150-FAKE16-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; GFX1150-FAKE16-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; GFX1150-FAKE16-NEXT: s_sub_i32 s9, s10, s9 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: s_add_i32 s9, s9, 11 -; GFX1150-FAKE16-NEXT: .LBB10_21: ; %frem.loop_body60 +; GFX1150-FAKE16-NEXT: .LBB10_21: ; %frem.loop_body27 ; GFX1150-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v7, v4 @@ -10534,7 +10534,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: ; %bb.22: ; %Flow125 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v6, s9 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v4, v7 -; GFX1150-FAKE16-NEXT: .LBB10_23: ; %frem.loop_exit61 +; GFX1150-FAKE16-NEXT: .LBB10_23: ; %frem.loop_exit28 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-FAKE16-NEXT: v_add_nc_u32_e32 v6, -10, v6 ; GFX1150-FAKE16-NEXT: v_ldexp_f32 v4, v4, v6 @@ -10563,7 +10563,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1150-FAKE16-NEXT: s_cmp_ngt_f32 s12, s11 ; GFX1150-FAKE16-NEXT: s_cbranch_scc0 .LBB10_26 -; GFX1150-FAKE16-NEXT: ; %bb.25: ; %frem.else86 +; GFX1150-FAKE16-NEXT: ; %bb.25: ; %frem.else ; GFX1150-FAKE16-NEXT: s_cmp_eq_f32 s12, s11 ; GFX1150-FAKE16-NEXT: v_bfi_b32 v3, 0x7fff, 0, s10 ; GFX1150-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -10573,7 +10573,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: s_branch .LBB10_32 ; GFX1150-FAKE16-NEXT: .LBB10_26: ; GFX1150-FAKE16-NEXT: ; implicit-def: $vgpr3 -; GFX1150-FAKE16-NEXT: .LBB10_27: ; %frem.compute85 +; GFX1150-FAKE16-NEXT: .LBB10_27: ; %frem.compute ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v4, s11 ; GFX1150-FAKE16-NEXT: v_frexp_mant_f32_e32 v3, s12 ; GFX1150-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v6, s12 @@ -10608,11 +10608,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v7 ; GFX1150-FAKE16-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; GFX1150-FAKE16-NEXT: s_cbranch_vccnz .LBB10_31 -; GFX1150-FAKE16-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; GFX1150-FAKE16-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX1150-FAKE16-NEXT: s_sub_i32 s11, s12, s11 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: s_add_i32 s11, s11, 11 -; GFX1150-FAKE16-NEXT: .LBB10_29: ; %frem.loop_body93 +; GFX1150-FAKE16-NEXT: .LBB10_29: ; %frem.loop_body ; GFX1150-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v8, v5 @@ -10634,7 +10634,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-FAKE16-NEXT: ; %bb.30: ; %Flow ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v7, s11 ; GFX1150-FAKE16-NEXT: v_mov_b32_e32 v5, v8 -; GFX1150-FAKE16-NEXT: .LBB10_31: ; %frem.loop_exit94 +; GFX1150-FAKE16-NEXT: .LBB10_31: ; %frem.loop_exit ; GFX1150-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-FAKE16-NEXT: v_add_nc_u32_e32 v7, -10, v7 ; GFX1150-FAKE16-NEXT: v_ldexp_f32 v5, v5, v7 @@ -10712,7 +10712,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1200-TRUE16-NEXT: s_cmp_ngt_f32 s8, s6 ; GFX1200-TRUE16-NEXT: s_cbranch_scc0 .LBB10_2 -; GFX1200-TRUE16-NEXT: ; %bb.1: ; %frem.else +; GFX1200-TRUE16-NEXT: ; %bb.1: ; %frem.else86 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v0.l, s5 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v1.l, 0 ; GFX1200-TRUE16-NEXT: s_cmp_eq_f32 s8, s6 @@ -10724,7 +10724,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_branch .LBB10_8 ; GFX1200-TRUE16-NEXT: .LBB10_2: ; GFX1200-TRUE16-NEXT: ; implicit-def: $vgpr0 -; GFX1200-TRUE16-NEXT: .LBB10_3: ; %frem.compute +; GFX1200-TRUE16-NEXT: .LBB10_3: ; %frem.compute85 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v1, s6 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v0, s8 ; GFX1200-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v3, s8 @@ -10759,11 +10759,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v4 ; GFX1200-TRUE16-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1200-TRUE16-NEXT: s_cbranch_vccnz .LBB10_7 -; GFX1200-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1200-TRUE16-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; GFX1200-TRUE16-NEXT: s_sub_co_i32 s6, s8, s6 ; GFX1200-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX1200-TRUE16-NEXT: s_add_co_i32 s6, s6, 11 -; GFX1200-TRUE16-NEXT: .LBB10_5: ; %frem.loop_body +; GFX1200-TRUE16-NEXT: .LBB10_5: ; %frem.loop_body93 ; GFX1200-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v5, v2 @@ -10787,7 +10787,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: ; %bb.6: ; %Flow133 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v4, s6 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v2, v5 -; GFX1200-TRUE16-NEXT: .LBB10_7: ; %frem.loop_exit +; GFX1200-TRUE16-NEXT: .LBB10_7: ; %frem.loop_exit94 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_add_nc_u32_e32 v4, -10, v4 ; GFX1200-TRUE16-NEXT: v_ldexp_f32 v2, v2, v4 @@ -10821,7 +10821,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_2) ; GFX1200-TRUE16-NEXT: s_cmp_ngt_f32 s10, s9 ; GFX1200-TRUE16-NEXT: s_cbranch_scc0 .LBB10_10 -; GFX1200-TRUE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX1200-TRUE16-NEXT: ; %bb.9: ; %frem.else53 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v1.l, s8 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 ; GFX1200-TRUE16-NEXT: s_cmp_eq_f32 s10, s9 @@ -10833,7 +10833,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_branch .LBB10_16 ; GFX1200-TRUE16-NEXT: .LBB10_10: ; GFX1200-TRUE16-NEXT: ; implicit-def: $vgpr1 -; GFX1200-TRUE16-NEXT: .LBB10_11: ; %frem.compute19 +; GFX1200-TRUE16-NEXT: .LBB10_11: ; %frem.compute52 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v2, s9 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v1, s10 ; GFX1200-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v4, s10 @@ -10869,11 +10869,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v5 ; GFX1200-TRUE16-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1200-TRUE16-NEXT: s_cbranch_vccnz .LBB10_15 -; GFX1200-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX1200-TRUE16-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; GFX1200-TRUE16-NEXT: s_sub_co_i32 s9, s10, s9 ; GFX1200-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX1200-TRUE16-NEXT: s_add_co_i32 s9, s9, 11 -; GFX1200-TRUE16-NEXT: .LBB10_13: ; %frem.loop_body27 +; GFX1200-TRUE16-NEXT: .LBB10_13: ; %frem.loop_body60 ; GFX1200-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v6, v3 @@ -10897,7 +10897,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: ; %bb.14: ; %Flow129 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v5, s9 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v3, v6 -; GFX1200-TRUE16-NEXT: .LBB10_15: ; %frem.loop_exit28 +; GFX1200-TRUE16-NEXT: .LBB10_15: ; %frem.loop_exit61 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_add_nc_u32_e32 v5, -10, v5 ; GFX1200-TRUE16-NEXT: v_ldexp_f32 v3, v3, v5 @@ -10928,7 +10928,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_2) ; GFX1200-TRUE16-NEXT: s_cmp_ngt_f32 s10, s9 ; GFX1200-TRUE16-NEXT: s_cbranch_scc0 .LBB10_18 -; GFX1200-TRUE16-NEXT: ; %bb.17: ; %frem.else53 +; GFX1200-TRUE16-NEXT: ; %bb.17: ; %frem.else20 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v2.l, s7 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v3.l, 0 ; GFX1200-TRUE16-NEXT: s_cmp_eq_f32 s10, s9 @@ -10941,7 +10941,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_branch .LBB10_24 ; GFX1200-TRUE16-NEXT: .LBB10_18: ; GFX1200-TRUE16-NEXT: ; implicit-def: $vgpr2 -; GFX1200-TRUE16-NEXT: .LBB10_19: ; %frem.compute52 +; GFX1200-TRUE16-NEXT: .LBB10_19: ; %frem.compute19 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v3, s9 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v2, s10 ; GFX1200-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v5, s10 @@ -10977,11 +10977,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v6 ; GFX1200-TRUE16-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX1200-TRUE16-NEXT: s_cbranch_vccnz .LBB10_23 -; GFX1200-TRUE16-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; GFX1200-TRUE16-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; GFX1200-TRUE16-NEXT: s_sub_co_i32 s9, s10, s9 ; GFX1200-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX1200-TRUE16-NEXT: s_add_co_i32 s9, s9, 11 -; GFX1200-TRUE16-NEXT: .LBB10_21: ; %frem.loop_body60 +; GFX1200-TRUE16-NEXT: .LBB10_21: ; %frem.loop_body27 ; GFX1200-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v7, v4 @@ -11005,7 +11005,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: ; %bb.22: ; %Flow125 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v6, s9 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v4, v7 -; GFX1200-TRUE16-NEXT: .LBB10_23: ; %frem.loop_exit61 +; GFX1200-TRUE16-NEXT: .LBB10_23: ; %frem.loop_exit28 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_add_nc_u32_e32 v6, -10, v6 ; GFX1200-TRUE16-NEXT: v_ldexp_f32 v4, v4, v6 @@ -11039,7 +11039,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_2) ; GFX1200-TRUE16-NEXT: s_cmp_ngt_f32 s12, s11 ; GFX1200-TRUE16-NEXT: s_cbranch_scc0 .LBB10_26 -; GFX1200-TRUE16-NEXT: ; %bb.25: ; %frem.else86 +; GFX1200-TRUE16-NEXT: ; %bb.25: ; %frem.else ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v3.l, s10 ; GFX1200-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 ; GFX1200-TRUE16-NEXT: s_cmp_eq_f32 s12, s11 @@ -11051,7 +11051,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: s_branch .LBB10_32 ; GFX1200-TRUE16-NEXT: .LBB10_26: ; GFX1200-TRUE16-NEXT: ; implicit-def: $vgpr3 -; GFX1200-TRUE16-NEXT: .LBB10_27: ; %frem.compute85 +; GFX1200-TRUE16-NEXT: .LBB10_27: ; %frem.compute ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v4, s11 ; GFX1200-TRUE16-NEXT: v_frexp_mant_f32_e32 v3, s12 ; GFX1200-TRUE16-NEXT: v_frexp_exp_i32_f32_e32 v6, s12 @@ -11087,11 +11087,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v7 ; GFX1200-TRUE16-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; GFX1200-TRUE16-NEXT: s_cbranch_vccnz .LBB10_31 -; GFX1200-TRUE16-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; GFX1200-TRUE16-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX1200-TRUE16-NEXT: s_sub_co_i32 s11, s12, s11 ; GFX1200-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX1200-TRUE16-NEXT: s_add_co_i32 s11, s11, 11 -; GFX1200-TRUE16-NEXT: .LBB10_29: ; %frem.loop_body93 +; GFX1200-TRUE16-NEXT: .LBB10_29: ; %frem.loop_body ; GFX1200-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v8, v5 @@ -11115,7 +11115,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-TRUE16-NEXT: ; %bb.30: ; %Flow ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v7, s11 ; GFX1200-TRUE16-NEXT: v_mov_b32_e32 v5, v8 -; GFX1200-TRUE16-NEXT: .LBB10_31: ; %frem.loop_exit94 +; GFX1200-TRUE16-NEXT: .LBB10_31: ; %frem.loop_exit ; GFX1200-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-TRUE16-NEXT: v_add_nc_u32_e32 v7, -10, v7 ; GFX1200-TRUE16-NEXT: v_ldexp_f32 v5, v5, v7 @@ -11203,7 +11203,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3) ; GFX1200-FAKE16-NEXT: s_cmp_ngt_f32 s8, s6 ; GFX1200-FAKE16-NEXT: s_cbranch_scc0 .LBB10_2 -; GFX1200-FAKE16-NEXT: ; %bb.1: ; %frem.else +; GFX1200-FAKE16-NEXT: ; %bb.1: ; %frem.else86 ; GFX1200-FAKE16-NEXT: s_cmp_eq_f32 s8, s6 ; GFX1200-FAKE16-NEXT: v_bfi_b32 v0, 0x7fff, 0, s5 ; GFX1200-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -11213,7 +11213,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_branch .LBB10_8 ; GFX1200-FAKE16-NEXT: .LBB10_2: ; GFX1200-FAKE16-NEXT: ; implicit-def: $vgpr0 -; GFX1200-FAKE16-NEXT: .LBB10_3: ; %frem.compute +; GFX1200-FAKE16-NEXT: .LBB10_3: ; %frem.compute85 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v1, s6 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v0, s8 ; GFX1200-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v3, s8 @@ -11249,11 +11249,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v4 ; GFX1200-FAKE16-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1200-FAKE16-NEXT: s_cbranch_vccnz .LBB10_7 -; GFX1200-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1200-FAKE16-NEXT: ; %bb.4: ; %frem.loop_body93.preheader ; GFX1200-FAKE16-NEXT: s_sub_co_i32 s6, s8, s6 ; GFX1200-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX1200-FAKE16-NEXT: s_add_co_i32 s6, s6, 11 -; GFX1200-FAKE16-NEXT: .LBB10_5: ; %frem.loop_body +; GFX1200-FAKE16-NEXT: .LBB10_5: ; %frem.loop_body93 ; GFX1200-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v5, v2 @@ -11277,7 +11277,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: ; %bb.6: ; %Flow133 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v4, s6 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v2, v5 -; GFX1200-FAKE16-NEXT: .LBB10_7: ; %frem.loop_exit +; GFX1200-FAKE16-NEXT: .LBB10_7: ; %frem.loop_exit94 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_add_nc_u32_e32 v4, -10, v4 ; GFX1200-FAKE16-NEXT: v_ldexp_f32 v2, v2, v4 @@ -11310,7 +11310,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_2) ; GFX1200-FAKE16-NEXT: s_cmp_ngt_f32 s10, s9 ; GFX1200-FAKE16-NEXT: s_cbranch_scc0 .LBB10_10 -; GFX1200-FAKE16-NEXT: ; %bb.9: ; %frem.else20 +; GFX1200-FAKE16-NEXT: ; %bb.9: ; %frem.else53 ; GFX1200-FAKE16-NEXT: s_cmp_eq_f32 s10, s9 ; GFX1200-FAKE16-NEXT: v_bfi_b32 v1, 0x7fff, 0, s8 ; GFX1200-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -11321,7 +11321,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_branch .LBB10_16 ; GFX1200-FAKE16-NEXT: .LBB10_10: ; GFX1200-FAKE16-NEXT: ; implicit-def: $vgpr1 -; GFX1200-FAKE16-NEXT: .LBB10_11: ; %frem.compute19 +; GFX1200-FAKE16-NEXT: .LBB10_11: ; %frem.compute52 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v2, s9 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v1, s10 ; GFX1200-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v4, s10 @@ -11357,11 +11357,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v5 ; GFX1200-FAKE16-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1200-FAKE16-NEXT: s_cbranch_vccnz .LBB10_15 -; GFX1200-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body27.preheader +; GFX1200-FAKE16-NEXT: ; %bb.12: ; %frem.loop_body60.preheader ; GFX1200-FAKE16-NEXT: s_sub_co_i32 s9, s10, s9 ; GFX1200-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX1200-FAKE16-NEXT: s_add_co_i32 s9, s9, 11 -; GFX1200-FAKE16-NEXT: .LBB10_13: ; %frem.loop_body27 +; GFX1200-FAKE16-NEXT: .LBB10_13: ; %frem.loop_body60 ; GFX1200-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v6, v3 @@ -11385,7 +11385,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: ; %bb.14: ; %Flow129 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v5, s9 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v3, v6 -; GFX1200-FAKE16-NEXT: .LBB10_15: ; %frem.loop_exit28 +; GFX1200-FAKE16-NEXT: .LBB10_15: ; %frem.loop_exit61 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_add_nc_u32_e32 v5, -10, v5 ; GFX1200-FAKE16-NEXT: v_ldexp_f32 v3, v3, v5 @@ -11415,7 +11415,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_2) ; GFX1200-FAKE16-NEXT: s_cmp_ngt_f32 s10, s9 ; GFX1200-FAKE16-NEXT: s_cbranch_scc0 .LBB10_18 -; GFX1200-FAKE16-NEXT: ; %bb.17: ; %frem.else53 +; GFX1200-FAKE16-NEXT: ; %bb.17: ; %frem.else20 ; GFX1200-FAKE16-NEXT: s_cmp_eq_f32 s10, s9 ; GFX1200-FAKE16-NEXT: v_bfi_b32 v2, 0x7fff, 0, s7 ; GFX1200-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -11426,7 +11426,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_branch .LBB10_24 ; GFX1200-FAKE16-NEXT: .LBB10_18: ; GFX1200-FAKE16-NEXT: ; implicit-def: $vgpr2 -; GFX1200-FAKE16-NEXT: .LBB10_19: ; %frem.compute52 +; GFX1200-FAKE16-NEXT: .LBB10_19: ; %frem.compute19 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v3, s9 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v2, s10 ; GFX1200-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v5, s10 @@ -11462,11 +11462,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v6 ; GFX1200-FAKE16-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX1200-FAKE16-NEXT: s_cbranch_vccnz .LBB10_23 -; GFX1200-FAKE16-NEXT: ; %bb.20: ; %frem.loop_body60.preheader +; GFX1200-FAKE16-NEXT: ; %bb.20: ; %frem.loop_body27.preheader ; GFX1200-FAKE16-NEXT: s_sub_co_i32 s9, s10, s9 ; GFX1200-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX1200-FAKE16-NEXT: s_add_co_i32 s9, s9, 11 -; GFX1200-FAKE16-NEXT: .LBB10_21: ; %frem.loop_body60 +; GFX1200-FAKE16-NEXT: .LBB10_21: ; %frem.loop_body27 ; GFX1200-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v7, v4 @@ -11490,7 +11490,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: ; %bb.22: ; %Flow125 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v6, s9 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v4, v7 -; GFX1200-FAKE16-NEXT: .LBB10_23: ; %frem.loop_exit61 +; GFX1200-FAKE16-NEXT: .LBB10_23: ; %frem.loop_exit28 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_add_nc_u32_e32 v6, -10, v6 ; GFX1200-FAKE16-NEXT: v_ldexp_f32 v4, v4, v6 @@ -11523,7 +11523,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_2) ; GFX1200-FAKE16-NEXT: s_cmp_ngt_f32 s12, s11 ; GFX1200-FAKE16-NEXT: s_cbranch_scc0 .LBB10_26 -; GFX1200-FAKE16-NEXT: ; %bb.25: ; %frem.else86 +; GFX1200-FAKE16-NEXT: ; %bb.25: ; %frem.else ; GFX1200-FAKE16-NEXT: s_cmp_eq_f32 s12, s11 ; GFX1200-FAKE16-NEXT: v_bfi_b32 v3, 0x7fff, 0, s10 ; GFX1200-FAKE16-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -11534,7 +11534,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: s_branch .LBB10_32 ; GFX1200-FAKE16-NEXT: .LBB10_26: ; GFX1200-FAKE16-NEXT: ; implicit-def: $vgpr3 -; GFX1200-FAKE16-NEXT: .LBB10_27: ; %frem.compute85 +; GFX1200-FAKE16-NEXT: .LBB10_27: ; %frem.compute ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v4, s11 ; GFX1200-FAKE16-NEXT: v_frexp_mant_f32_e32 v3, s12 ; GFX1200-FAKE16-NEXT: v_frexp_exp_i32_f32_e32 v6, s12 @@ -11570,11 +11570,11 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: v_cmp_gt_i32_e32 vcc_lo, 12, v7 ; GFX1200-FAKE16-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; GFX1200-FAKE16-NEXT: s_cbranch_vccnz .LBB10_31 -; GFX1200-FAKE16-NEXT: ; %bb.28: ; %frem.loop_body93.preheader +; GFX1200-FAKE16-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX1200-FAKE16-NEXT: s_sub_co_i32 s11, s12, s11 ; GFX1200-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX1200-FAKE16-NEXT: s_add_co_i32 s11, s11, 11 -; GFX1200-FAKE16-NEXT: .LBB10_29: ; %frem.loop_body93 +; GFX1200-FAKE16-NEXT: .LBB10_29: ; %frem.loop_body ; GFX1200-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v8, v5 @@ -11598,7 +11598,7 @@ define amdgpu_kernel void @frem_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-FAKE16-NEXT: ; %bb.30: ; %Flow ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v7, s11 ; GFX1200-FAKE16-NEXT: v_mov_b32_e32 v5, v8 -; GFX1200-FAKE16-NEXT: .LBB10_31: ; %frem.loop_exit94 +; GFX1200-FAKE16-NEXT: .LBB10_31: ; %frem.loop_exit ; GFX1200-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-FAKE16-NEXT: v_add_nc_u32_e32 v7, -10, v7 ; GFX1200-FAKE16-NEXT: v_ldexp_f32 v5, v5, v7 @@ -11686,7 +11686,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v0|, |v2| ; SI-NEXT: s_and_b64 vcc, exec, s[2:3] ; SI-NEXT: s_cbranch_vccz .LBB11_2 -; SI-NEXT: ; %bb.1: ; %frem.else +; SI-NEXT: ; %bb.1: ; %frem.else16 ; SI-NEXT: s_brev_b32 s2, -2 ; SI-NEXT: v_bfi_b32 v4, s2, 0, v0 ; SI-NEXT: v_cmp_eq_f32_e64 vcc, |v0|, |v2| @@ -11697,7 +11697,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB11_2: ; SI-NEXT: ; implicit-def: $vgpr4 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB11_3: ; %frem.compute +; SI-NEXT: .LBB11_3: ; %frem.compute15 ; SI-NEXT: s_mov_b32 s6, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 s[2:3], |v0|, s6 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v4, v0 @@ -11733,10 +11733,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; SI-NEXT: s_cmp_lt_i32 s3, 13 ; SI-NEXT: s_cbranch_scc1 .LBB11_7 -; SI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; SI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; SI-NEXT: s_sub_i32 s3, s4, s5 ; SI-NEXT: s_add_i32 s3, s3, 12 -; SI-NEXT: .LBB11_5: ; %frem.loop_body +; SI-NEXT: .LBB11_5: ; %frem.loop_body23 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v7, v5 ; SI-NEXT: v_mul_f32_e32 v5, v7, v6 @@ -11751,7 +11751,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB11_5 ; SI-NEXT: ; %bb.6: ; %Flow51 ; SI-NEXT: v_mov_b32_e32 v5, v7 -; SI-NEXT: .LBB11_7: ; %frem.loop_exit +; SI-NEXT: .LBB11_7: ; %frem.loop_exit24 ; SI-NEXT: s_add_i32 s3, s3, -11 ; SI-NEXT: v_ldexp_f32_e64 v5, v5, s3 ; SI-NEXT: v_mul_f32_e32 v6, v5, v6 @@ -11767,7 +11767,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v1|, |v3| ; SI-NEXT: s_and_b64 vcc, exec, s[2:3] ; SI-NEXT: s_cbranch_vccz .LBB11_10 -; SI-NEXT: ; %bb.9: ; %frem.else16 +; SI-NEXT: ; %bb.9: ; %frem.else ; SI-NEXT: s_brev_b32 s2, -2 ; SI-NEXT: v_bfi_b32 v5, s2, 0, v1 ; SI-NEXT: v_cmp_eq_f32_e64 vcc, |v1|, |v3| @@ -11778,7 +11778,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB11_10: ; SI-NEXT: ; implicit-def: $vgpr5 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB11_11: ; %frem.compute15 +; SI-NEXT: .LBB11_11: ; %frem.compute ; SI-NEXT: s_mov_b32 s6, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 s[2:3], |v1|, s6 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v5, v1 @@ -11814,10 +11814,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v7, v7, v5, 1.0 ; SI-NEXT: s_cmp_lt_i32 s3, 13 ; SI-NEXT: s_cbranch_scc1 .LBB11_15 -; SI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; SI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; SI-NEXT: s_sub_i32 s3, s4, s5 ; SI-NEXT: s_add_i32 s3, s3, 12 -; SI-NEXT: .LBB11_13: ; %frem.loop_body23 +; SI-NEXT: .LBB11_13: ; %frem.loop_body ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v8, v6 ; SI-NEXT: v_mul_f32_e32 v6, v8, v7 @@ -11832,7 +11832,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB11_13 ; SI-NEXT: ; %bb.14: ; %Flow ; SI-NEXT: v_mov_b32_e32 v6, v8 -; SI-NEXT: .LBB11_15: ; %frem.loop_exit24 +; SI-NEXT: .LBB11_15: ; %frem.loop_exit ; SI-NEXT: s_add_i32 s3, s3, -11 ; SI-NEXT: v_ldexp_f32_e64 v6, v6, s3 ; SI-NEXT: v_mul_f32_e32 v7, v6, v7 @@ -11877,7 +11877,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v0|, |v2| ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB11_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else16 ; CI-NEXT: s_brev_b32 s2, -2 ; CI-NEXT: v_bfi_b32 v4, s2, 0, v0 ; CI-NEXT: v_cmp_eq_f32_e64 vcc, |v0|, |v2| @@ -11886,7 +11886,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB11_8 ; CI-NEXT: .LBB11_2: ; CI-NEXT: ; implicit-def: $vgpr4 -; CI-NEXT: .LBB11_3: ; %frem.compute +; CI-NEXT: .LBB11_3: ; %frem.compute15 ; CI-NEXT: v_frexp_mant_f32_e64 v5, |v2| ; CI-NEXT: v_ldexp_f32_e64 v5, v5, 1 ; CI-NEXT: v_div_scale_f32 v11, s[2:3], v5, v5, 1.0 @@ -11911,10 +11911,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v6 ; CI-NEXT: v_div_fixup_f32 v8, v8, v5, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB11_7 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; CI-NEXT: v_sub_i32_e32 v6, vcc, v9, v10 ; CI-NEXT: v_add_i32_e32 v6, vcc, 12, v6 -; CI-NEXT: .LBB11_5: ; %frem.loop_body +; CI-NEXT: .LBB11_5: ; %frem.loop_body23 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v9, v7 ; CI-NEXT: v_mul_f32_e32 v7, v9, v8 @@ -11929,7 +11929,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB11_5 ; CI-NEXT: ; %bb.6: ; %Flow51 ; CI-NEXT: v_mov_b32_e32 v7, v9 -; CI-NEXT: .LBB11_7: ; %frem.loop_exit +; CI-NEXT: .LBB11_7: ; %frem.loop_exit24 ; CI-NEXT: v_add_i32_e32 v6, vcc, -11, v6 ; CI-NEXT: v_ldexp_f32_e32 v6, v7, v6 ; CI-NEXT: v_mul_f32_e32 v7, v6, v8 @@ -11945,7 +11945,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v1|, |v3| ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB11_10 -; CI-NEXT: ; %bb.9: ; %frem.else16 +; CI-NEXT: ; %bb.9: ; %frem.else ; CI-NEXT: s_brev_b32 s2, -2 ; CI-NEXT: v_bfi_b32 v5, s2, 0, v1 ; CI-NEXT: v_cmp_eq_f32_e64 vcc, |v1|, |v3| @@ -11954,7 +11954,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB11_16 ; CI-NEXT: .LBB11_10: ; CI-NEXT: ; implicit-def: $vgpr5 -; CI-NEXT: .LBB11_11: ; %frem.compute15 +; CI-NEXT: .LBB11_11: ; %frem.compute ; CI-NEXT: v_frexp_mant_f32_e64 v6, |v3| ; CI-NEXT: v_ldexp_f32_e64 v6, v6, 1 ; CI-NEXT: v_div_scale_f32 v12, s[2:3], v6, v6, 1.0 @@ -11979,10 +11979,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v7 ; CI-NEXT: v_div_fixup_f32 v9, v9, v6, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB11_15 -; CI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; CI-NEXT: v_sub_i32_e32 v7, vcc, v10, v11 ; CI-NEXT: v_add_i32_e32 v7, vcc, 12, v7 -; CI-NEXT: .LBB11_13: ; %frem.loop_body23 +; CI-NEXT: .LBB11_13: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v10, v8 ; CI-NEXT: v_mul_f32_e32 v8, v10, v9 @@ -11997,7 +11997,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB11_13 ; CI-NEXT: ; %bb.14: ; %Flow ; CI-NEXT: v_mov_b32_e32 v8, v10 -; CI-NEXT: .LBB11_15: ; %frem.loop_exit24 +; CI-NEXT: .LBB11_15: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v7, vcc, -11, v7 ; CI-NEXT: v_ldexp_f32_e32 v7, v8, v7 ; CI-NEXT: v_mul_f32_e32 v8, v7, v9 @@ -12042,7 +12042,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v0|, |v2| ; VI-NEXT: s_and_b64 vcc, exec, s[2:3] ; VI-NEXT: s_cbranch_vccz .LBB11_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else16 ; VI-NEXT: s_brev_b32 s2, -2 ; VI-NEXT: v_bfi_b32 v4, s2, 0, v0 ; VI-NEXT: v_cmp_eq_f32_e64 vcc, |v0|, |v2| @@ -12051,7 +12051,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB11_8 ; VI-NEXT: .LBB11_2: ; VI-NEXT: ; implicit-def: $vgpr4 -; VI-NEXT: .LBB11_3: ; %frem.compute +; VI-NEXT: .LBB11_3: ; %frem.compute15 ; VI-NEXT: v_frexp_mant_f32_e64 v5, |v2| ; VI-NEXT: v_ldexp_f32 v5, v5, 1 ; VI-NEXT: v_div_scale_f32 v11, s[2:3], v5, v5, 1.0 @@ -12076,10 +12076,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v6 ; VI-NEXT: v_div_fixup_f32 v8, v8, v5, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB11_7 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; VI-NEXT: v_sub_u32_e32 v6, vcc, v9, v10 ; VI-NEXT: v_add_u32_e32 v6, vcc, 12, v6 -; VI-NEXT: .LBB11_5: ; %frem.loop_body +; VI-NEXT: .LBB11_5: ; %frem.loop_body23 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v9, v7 ; VI-NEXT: v_mul_f32_e32 v7, v9, v8 @@ -12094,7 +12094,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB11_5 ; VI-NEXT: ; %bb.6: ; %Flow51 ; VI-NEXT: v_mov_b32_e32 v7, v9 -; VI-NEXT: .LBB11_7: ; %frem.loop_exit +; VI-NEXT: .LBB11_7: ; %frem.loop_exit24 ; VI-NEXT: v_add_u32_e32 v6, vcc, -11, v6 ; VI-NEXT: v_ldexp_f32 v6, v7, v6 ; VI-NEXT: v_mul_f32_e32 v7, v6, v8 @@ -12110,7 +12110,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v1|, |v3| ; VI-NEXT: s_and_b64 vcc, exec, s[2:3] ; VI-NEXT: s_cbranch_vccz .LBB11_10 -; VI-NEXT: ; %bb.9: ; %frem.else16 +; VI-NEXT: ; %bb.9: ; %frem.else ; VI-NEXT: s_brev_b32 s2, -2 ; VI-NEXT: v_bfi_b32 v5, s2, 0, v1 ; VI-NEXT: v_cmp_eq_f32_e64 vcc, |v1|, |v3| @@ -12119,7 +12119,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB11_16 ; VI-NEXT: .LBB11_10: ; VI-NEXT: ; implicit-def: $vgpr5 -; VI-NEXT: .LBB11_11: ; %frem.compute15 +; VI-NEXT: .LBB11_11: ; %frem.compute ; VI-NEXT: v_frexp_mant_f32_e64 v6, |v3| ; VI-NEXT: v_ldexp_f32 v6, v6, 1 ; VI-NEXT: v_div_scale_f32 v12, s[2:3], v6, v6, 1.0 @@ -12144,10 +12144,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v7 ; VI-NEXT: v_div_fixup_f32 v9, v9, v6, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB11_15 -; VI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; VI-NEXT: v_sub_u32_e32 v7, vcc, v10, v11 ; VI-NEXT: v_add_u32_e32 v7, vcc, 12, v7 -; VI-NEXT: .LBB11_13: ; %frem.loop_body23 +; VI-NEXT: .LBB11_13: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v10, v8 ; VI-NEXT: v_mul_f32_e32 v8, v10, v9 @@ -12162,7 +12162,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB11_13 ; VI-NEXT: ; %bb.14: ; %Flow ; VI-NEXT: v_mov_b32_e32 v8, v10 -; VI-NEXT: .LBB11_15: ; %frem.loop_exit24 +; VI-NEXT: .LBB11_15: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v7, vcc, -11, v7 ; VI-NEXT: v_ldexp_f32 v7, v8, v7 ; VI-NEXT: v_mul_f32_e32 v8, v7, v9 @@ -12202,7 +12202,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v0|, |v2| ; GFX9-NEXT: s_and_b64 vcc, exec, s[2:3] ; GFX9-NEXT: s_cbranch_vccz .LBB11_2 -; GFX9-NEXT: ; %bb.1: ; %frem.else +; GFX9-NEXT: ; %bb.1: ; %frem.else16 ; GFX9-NEXT: s_brev_b32 s2, -2 ; GFX9-NEXT: v_bfi_b32 v4, s2, 0, v0 ; GFX9-NEXT: v_cmp_eq_f32_e64 vcc, |v0|, |v2| @@ -12211,7 +12211,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB11_8 ; GFX9-NEXT: .LBB11_2: ; GFX9-NEXT: ; implicit-def: $vgpr4 -; GFX9-NEXT: .LBB11_3: ; %frem.compute +; GFX9-NEXT: .LBB11_3: ; %frem.compute15 ; GFX9-NEXT: v_frexp_mant_f32_e64 v5, |v2| ; GFX9-NEXT: v_ldexp_f32 v5, v5, 1 ; GFX9-NEXT: v_div_scale_f32 v11, s[2:3], v5, v5, 1.0 @@ -12236,10 +12236,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 13, v6 ; GFX9-NEXT: v_div_fixup_f32 v8, v8, v5, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB11_7 -; GFX9-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX9-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX9-NEXT: v_sub_u32_e32 v6, v9, v10 ; GFX9-NEXT: v_add_u32_e32 v6, 12, v6 -; GFX9-NEXT: .LBB11_5: ; %frem.loop_body +; GFX9-NEXT: .LBB11_5: ; %frem.loop_body23 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v9, v7 ; GFX9-NEXT: v_mul_f32_e32 v7, v9, v8 @@ -12254,7 +12254,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB11_5 ; GFX9-NEXT: ; %bb.6: ; %Flow51 ; GFX9-NEXT: v_mov_b32_e32 v7, v9 -; GFX9-NEXT: .LBB11_7: ; %frem.loop_exit +; GFX9-NEXT: .LBB11_7: ; %frem.loop_exit24 ; GFX9-NEXT: v_add_u32_e32 v6, -11, v6 ; GFX9-NEXT: v_ldexp_f32 v6, v7, v6 ; GFX9-NEXT: v_mul_f32_e32 v7, v6, v8 @@ -12270,7 +12270,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v1|, |v3| ; GFX9-NEXT: s_and_b64 vcc, exec, s[2:3] ; GFX9-NEXT: s_cbranch_vccz .LBB11_10 -; GFX9-NEXT: ; %bb.9: ; %frem.else16 +; GFX9-NEXT: ; %bb.9: ; %frem.else ; GFX9-NEXT: s_brev_b32 s2, -2 ; GFX9-NEXT: v_bfi_b32 v5, s2, 0, v1 ; GFX9-NEXT: v_cmp_eq_f32_e64 vcc, |v1|, |v3| @@ -12279,7 +12279,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB11_16 ; GFX9-NEXT: .LBB11_10: ; GFX9-NEXT: ; implicit-def: $vgpr5 -; GFX9-NEXT: .LBB11_11: ; %frem.compute15 +; GFX9-NEXT: .LBB11_11: ; %frem.compute ; GFX9-NEXT: v_frexp_mant_f32_e64 v6, |v3| ; GFX9-NEXT: v_ldexp_f32 v6, v6, 1 ; GFX9-NEXT: v_div_scale_f32 v12, s[2:3], v6, v6, 1.0 @@ -12304,10 +12304,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 13, v7 ; GFX9-NEXT: v_div_fixup_f32 v9, v9, v6, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB11_15 -; GFX9-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX9-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX9-NEXT: v_sub_u32_e32 v7, v10, v11 ; GFX9-NEXT: v_add_u32_e32 v7, 12, v7 -; GFX9-NEXT: .LBB11_13: ; %frem.loop_body23 +; GFX9-NEXT: .LBB11_13: ; %frem.loop_body ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v10, v8 ; GFX9-NEXT: v_mul_f32_e32 v8, v10, v9 @@ -12322,7 +12322,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB11_13 ; GFX9-NEXT: ; %bb.14: ; %Flow ; GFX9-NEXT: v_mov_b32_e32 v8, v10 -; GFX9-NEXT: .LBB11_15: ; %frem.loop_exit24 +; GFX9-NEXT: .LBB11_15: ; %frem.loop_exit ; GFX9-NEXT: v_add_u32_e32 v7, -11, v7 ; GFX9-NEXT: v_ldexp_f32 v7, v8, v7 ; GFX9-NEXT: v_mul_f32_e32 v8, v7, v9 @@ -12363,7 +12363,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_ngt_f32_e64 s2, |v0|, |v2| ; GFX10-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX10-NEXT: s_cbranch_vccz .LBB11_2 -; GFX10-NEXT: ; %bb.1: ; %frem.else +; GFX10-NEXT: ; %bb.1: ; %frem.else16 ; GFX10-NEXT: v_bfi_b32 v4, 0x7fffffff, 0, v0 ; GFX10-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v0|, |v2| ; GFX10-NEXT: v_cndmask_b32_e32 v4, v0, v4, vcc_lo @@ -12371,7 +12371,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB11_8 ; GFX10-NEXT: .LBB11_2: ; GFX10-NEXT: ; implicit-def: $vgpr4 -; GFX10-NEXT: .LBB11_3: ; %frem.compute +; GFX10-NEXT: .LBB11_3: ; %frem.compute15 ; GFX10-NEXT: v_frexp_mant_f32_e64 v5, |v2| ; GFX10-NEXT: v_frexp_mant_f32_e64 v4, |v0| ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v7, v0 @@ -12398,10 +12398,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v8 ; GFX10-NEXT: v_div_fixup_f32 v7, v7, v5, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB11_7 -; GFX10-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX10-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 12 -; GFX10-NEXT: .LBB11_5: ; %frem.loop_body +; GFX10-NEXT: .LBB11_5: ; %frem.loop_body23 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v9, v6 ; GFX10-NEXT: s_add_i32 s2, s2, -12 @@ -12417,7 +12417,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.6: ; %Flow51 ; GFX10-NEXT: v_mov_b32_e32 v8, s2 ; GFX10-NEXT: v_mov_b32_e32 v6, v9 -; GFX10-NEXT: .LBB11_7: ; %frem.loop_exit +; GFX10-NEXT: .LBB11_7: ; %frem.loop_exit24 ; GFX10-NEXT: v_add_nc_u32_e32 v8, -11, v8 ; GFX10-NEXT: v_ldexp_f32 v6, v6, v8 ; GFX10-NEXT: v_mul_f32_e32 v7, v6, v7 @@ -12432,7 +12432,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_ngt_f32_e64 s2, |v1|, |v3| ; GFX10-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX10-NEXT: s_cbranch_vccz .LBB11_10 -; GFX10-NEXT: ; %bb.9: ; %frem.else16 +; GFX10-NEXT: ; %bb.9: ; %frem.else ; GFX10-NEXT: v_bfi_b32 v5, 0x7fffffff, 0, v1 ; GFX10-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v1|, |v3| ; GFX10-NEXT: v_cndmask_b32_e32 v5, v1, v5, vcc_lo @@ -12440,7 +12440,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB11_16 ; GFX10-NEXT: .LBB11_10: ; GFX10-NEXT: ; implicit-def: $vgpr5 -; GFX10-NEXT: .LBB11_11: ; %frem.compute15 +; GFX10-NEXT: .LBB11_11: ; %frem.compute ; GFX10-NEXT: v_frexp_mant_f32_e64 v6, |v3| ; GFX10-NEXT: v_frexp_mant_f32_e64 v5, |v1| ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v8, v1 @@ -12467,10 +12467,10 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v9 ; GFX10-NEXT: v_div_fixup_f32 v8, v8, v6, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB11_15 -; GFX10-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX10-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 12 -; GFX10-NEXT: .LBB11_13: ; %frem.loop_body23 +; GFX10-NEXT: .LBB11_13: ; %frem.loop_body ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v10, v7 ; GFX10-NEXT: s_add_i32 s2, s2, -12 @@ -12486,7 +12486,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.14: ; %Flow ; GFX10-NEXT: v_mov_b32_e32 v9, s2 ; GFX10-NEXT: v_mov_b32_e32 v7, v10 -; GFX10-NEXT: .LBB11_15: ; %frem.loop_exit24 +; GFX10-NEXT: .LBB11_15: ; %frem.loop_exit ; GFX10-NEXT: v_add_nc_u32_e32 v9, -11, v9 ; GFX10-NEXT: v_ldexp_f32 v7, v7, v9 ; GFX10-NEXT: v_mul_f32_e32 v8, v7, v8 @@ -12524,7 +12524,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: v_cmp_ngt_f32_e64 s2, |v0|, |v2| ; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX11-NEXT: s_cbranch_vccz .LBB11_2 -; GFX11-NEXT: ; %bb.1: ; %frem.else +; GFX11-NEXT: ; %bb.1: ; %frem.else16 ; GFX11-NEXT: v_bfi_b32 v4, 0x7fffffff, 0, v0 ; GFX11-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v0|, |v2| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -12533,7 +12533,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_branch .LBB11_8 ; GFX11-NEXT: .LBB11_2: ; GFX11-NEXT: ; implicit-def: $vgpr4 -; GFX11-NEXT: .LBB11_3: ; %frem.compute +; GFX11-NEXT: .LBB11_3: ; %frem.compute15 ; GFX11-NEXT: v_frexp_mant_f32_e64 v5, |v2| ; GFX11-NEXT: v_frexp_mant_f32_e64 v4, |v0| ; GFX11-NEXT: v_frexp_exp_i32_f32_e32 v7, v0 @@ -12569,11 +12569,11 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_div_fixup_f32 v7, v7, v5, 1.0 ; GFX11-NEXT: s_cbranch_vccnz .LBB11_7 -; GFX11-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX11-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX11-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_add_i32 s2, s2, 12 -; GFX11-NEXT: .LBB11_5: ; %frem.loop_body +; GFX11-NEXT: .LBB11_5: ; %frem.loop_body23 ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-NEXT: v_mov_b32_e32 v9, v6 @@ -12593,7 +12593,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: ; %bb.6: ; %Flow51 ; GFX11-NEXT: v_mov_b32_e32 v8, s2 ; GFX11-NEXT: v_mov_b32_e32 v6, v9 -; GFX11-NEXT: .LBB11_7: ; %frem.loop_exit +; GFX11-NEXT: .LBB11_7: ; %frem.loop_exit24 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-NEXT: v_add_nc_u32_e32 v8, -11, v8 ; GFX11-NEXT: v_ldexp_f32 v6, v6, v8 @@ -12613,7 +12613,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: v_cmp_ngt_f32_e64 s2, |v1|, |v3| ; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX11-NEXT: s_cbranch_vccz .LBB11_10 -; GFX11-NEXT: ; %bb.9: ; %frem.else16 +; GFX11-NEXT: ; %bb.9: ; %frem.else ; GFX11-NEXT: v_bfi_b32 v5, 0x7fffffff, 0, v1 ; GFX11-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v1|, |v3| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -12622,7 +12622,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_branch .LBB11_16 ; GFX11-NEXT: .LBB11_10: ; GFX11-NEXT: ; implicit-def: $vgpr5 -; GFX11-NEXT: .LBB11_11: ; %frem.compute15 +; GFX11-NEXT: .LBB11_11: ; %frem.compute ; GFX11-NEXT: v_frexp_mant_f32_e64 v6, |v3| ; GFX11-NEXT: v_frexp_mant_f32_e64 v5, |v1| ; GFX11-NEXT: v_frexp_exp_i32_f32_e32 v8, v1 @@ -12658,11 +12658,11 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_div_fixup_f32 v8, v8, v6, 1.0 ; GFX11-NEXT: s_cbranch_vccnz .LBB11_15 -; GFX11-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX11-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX11-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_add_i32 s2, s2, 12 -; GFX11-NEXT: .LBB11_13: ; %frem.loop_body23 +; GFX11-NEXT: .LBB11_13: ; %frem.loop_body ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-NEXT: v_mov_b32_e32 v10, v7 @@ -12682,7 +12682,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: ; %bb.14: ; %Flow ; GFX11-NEXT: v_mov_b32_e32 v9, s2 ; GFX11-NEXT: v_mov_b32_e32 v7, v10 -; GFX11-NEXT: .LBB11_15: ; %frem.loop_exit24 +; GFX11-NEXT: .LBB11_15: ; %frem.loop_exit ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-NEXT: v_add_nc_u32_e32 v9, -11, v9 ; GFX11-NEXT: v_ldexp_f32 v7, v7, v9 @@ -12730,7 +12730,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_cmp_ngt_f32 s3, s8 ; GFX1150-NEXT: s_cbranch_scc0 .LBB11_2 -; GFX1150-NEXT: ; %bb.1: ; %frem.else +; GFX1150-NEXT: ; %bb.1: ; %frem.else16 ; GFX1150-NEXT: s_cmp_eq_f32 s3, s8 ; GFX1150-NEXT: v_bfi_b32 v0, 0x7fffffff, 0, s6 ; GFX1150-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -12740,7 +12740,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_branch .LBB11_8 ; GFX1150-NEXT: .LBB11_2: ; GFX1150-NEXT: ; implicit-def: $vgpr0 -; GFX1150-NEXT: .LBB11_3: ; %frem.compute +; GFX1150-NEXT: .LBB11_3: ; %frem.compute15 ; GFX1150-NEXT: v_frexp_mant_f32_e64 v1, |s4| ; GFX1150-NEXT: v_frexp_mant_f32_e64 v0, |s6| ; GFX1150-NEXT: v_frexp_exp_i32_f32_e32 v3, s6 @@ -12775,11 +12775,11 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v4 ; GFX1150-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1150-NEXT: s_cbranch_vccnz .LBB11_7 -; GFX1150-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1150-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX1150-NEXT: s_sub_i32 s7, s7, s8 ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_add_i32 s7, s7, 12 -; GFX1150-NEXT: .LBB11_5: ; %frem.loop_body +; GFX1150-NEXT: .LBB11_5: ; %frem.loop_body23 ; GFX1150-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-NEXT: v_mov_b32_e32 v5, v2 @@ -12801,7 +12801,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: ; %bb.6: ; %Flow51 ; GFX1150-NEXT: v_mov_b32_e32 v4, s7 ; GFX1150-NEXT: v_mov_b32_e32 v2, v5 -; GFX1150-NEXT: .LBB11_7: ; %frem.loop_exit +; GFX1150-NEXT: .LBB11_7: ; %frem.loop_exit24 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-NEXT: v_add_nc_u32_e32 v4, -11, v4 ; GFX1150-NEXT: v_ldexp_f32 v2, v2, v4 @@ -12824,7 +12824,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_cmp_ngt_f32 s6, s8 ; GFX1150-NEXT: s_cbranch_scc0 .LBB11_10 -; GFX1150-NEXT: ; %bb.9: ; %frem.else16 +; GFX1150-NEXT: ; %bb.9: ; %frem.else ; GFX1150-NEXT: s_cmp_eq_f32 s6, s8 ; GFX1150-NEXT: v_bfi_b32 v1, 0x7fffffff, 0, s5 ; GFX1150-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -12834,7 +12834,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_branch .LBB11_16 ; GFX1150-NEXT: .LBB11_10: ; GFX1150-NEXT: ; implicit-def: $vgpr1 -; GFX1150-NEXT: .LBB11_11: ; %frem.compute15 +; GFX1150-NEXT: .LBB11_11: ; %frem.compute ; GFX1150-NEXT: v_frexp_mant_f32_e64 v2, |s2| ; GFX1150-NEXT: v_frexp_mant_f32_e64 v1, |s5| ; GFX1150-NEXT: v_frexp_exp_i32_f32_e32 v4, s5 @@ -12869,11 +12869,11 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v5 ; GFX1150-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1150-NEXT: s_cbranch_vccnz .LBB11_15 -; GFX1150-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX1150-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX1150-NEXT: s_sub_i32 s7, s7, s8 ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_add_i32 s7, s7, 12 -; GFX1150-NEXT: .LBB11_13: ; %frem.loop_body23 +; GFX1150-NEXT: .LBB11_13: ; %frem.loop_body ; GFX1150-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-NEXT: v_mov_b32_e32 v6, v3 @@ -12895,7 +12895,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: ; %bb.14: ; %Flow ; GFX1150-NEXT: v_mov_b32_e32 v5, s7 ; GFX1150-NEXT: v_mov_b32_e32 v3, v6 -; GFX1150-NEXT: .LBB11_15: ; %frem.loop_exit24 +; GFX1150-NEXT: .LBB11_15: ; %frem.loop_exit ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-NEXT: v_add_nc_u32_e32 v5, -11, v5 ; GFX1150-NEXT: v_ldexp_f32 v3, v3, v5 @@ -12950,7 +12950,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1200-NEXT: s_cmp_ngt_f32 s3, s8 ; GFX1200-NEXT: s_cbranch_scc0 .LBB11_2 -; GFX1200-NEXT: ; %bb.1: ; %frem.else +; GFX1200-NEXT: ; %bb.1: ; %frem.else16 ; GFX1200-NEXT: s_cmp_eq_f32 s3, s8 ; GFX1200-NEXT: v_bfi_b32 v0, 0x7fffffff, 0, s6 ; GFX1200-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -12960,7 +12960,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_branch .LBB11_8 ; GFX1200-NEXT: .LBB11_2: ; GFX1200-NEXT: ; implicit-def: $vgpr0 -; GFX1200-NEXT: .LBB11_3: ; %frem.compute +; GFX1200-NEXT: .LBB11_3: ; %frem.compute15 ; GFX1200-NEXT: v_frexp_mant_f32_e64 v1, |s4| ; GFX1200-NEXT: v_frexp_mant_f32_e64 v0, |s6| ; GFX1200-NEXT: v_frexp_exp_i32_f32_e32 v3, s6 @@ -12996,11 +12996,11 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v4 ; GFX1200-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1200-NEXT: s_cbranch_vccnz .LBB11_7 -; GFX1200-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1200-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX1200-NEXT: s_sub_co_i32 s7, s7, s8 ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_add_co_i32 s7, s7, 12 -; GFX1200-NEXT: .LBB11_5: ; %frem.loop_body +; GFX1200-NEXT: .LBB11_5: ; %frem.loop_body23 ; GFX1200-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-NEXT: v_mov_b32_e32 v5, v2 @@ -13024,7 +13024,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: ; %bb.6: ; %Flow51 ; GFX1200-NEXT: v_mov_b32_e32 v4, s7 ; GFX1200-NEXT: v_mov_b32_e32 v2, v5 -; GFX1200-NEXT: .LBB11_7: ; %frem.loop_exit +; GFX1200-NEXT: .LBB11_7: ; %frem.loop_exit24 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-NEXT: v_add_nc_u32_e32 v4, -11, v4 ; GFX1200-NEXT: v_ldexp_f32 v2, v2, v4 @@ -13048,7 +13048,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_cmp_ngt_f32 s6, s8 ; GFX1200-NEXT: s_cbranch_scc0 .LBB11_10 -; GFX1200-NEXT: ; %bb.9: ; %frem.else16 +; GFX1200-NEXT: ; %bb.9: ; %frem.else ; GFX1200-NEXT: s_cmp_eq_f32 s6, s8 ; GFX1200-NEXT: v_bfi_b32 v1, 0x7fffffff, 0, s5 ; GFX1200-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -13059,7 +13059,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_branch .LBB11_16 ; GFX1200-NEXT: .LBB11_10: ; GFX1200-NEXT: ; implicit-def: $vgpr1 -; GFX1200-NEXT: .LBB11_11: ; %frem.compute15 +; GFX1200-NEXT: .LBB11_11: ; %frem.compute ; GFX1200-NEXT: v_frexp_mant_f32_e64 v2, |s2| ; GFX1200-NEXT: v_frexp_mant_f32_e64 v1, |s5| ; GFX1200-NEXT: v_frexp_exp_i32_f32_e32 v4, s5 @@ -13095,11 +13095,11 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v5 ; GFX1200-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1200-NEXT: s_cbranch_vccnz .LBB11_15 -; GFX1200-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX1200-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX1200-NEXT: s_sub_co_i32 s7, s7, s8 ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_add_co_i32 s7, s7, 12 -; GFX1200-NEXT: .LBB11_13: ; %frem.loop_body23 +; GFX1200-NEXT: .LBB11_13: ; %frem.loop_body ; GFX1200-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-NEXT: v_mov_b32_e32 v6, v3 @@ -13123,7 +13123,7 @@ define amdgpu_kernel void @frem_v2f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: ; %bb.14: ; %Flow ; GFX1200-NEXT: v_mov_b32_e32 v5, s7 ; GFX1200-NEXT: v_mov_b32_e32 v3, v6 -; GFX1200-NEXT: .LBB11_15: ; %frem.loop_exit24 +; GFX1200-NEXT: .LBB11_15: ; %frem.loop_exit ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-NEXT: v_add_nc_u32_e32 v5, -11, v5 ; GFX1200-NEXT: v_ldexp_f32 v3, v3, v5 @@ -13187,7 +13187,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v0|, |v4| ; SI-NEXT: s_and_b64 vcc, exec, s[2:3] ; SI-NEXT: s_cbranch_vccz .LBB12_2 -; SI-NEXT: ; %bb.1: ; %frem.else +; SI-NEXT: ; %bb.1: ; %frem.else78 ; SI-NEXT: s_brev_b32 s2, -2 ; SI-NEXT: v_bfi_b32 v8, s2, 0, v0 ; SI-NEXT: v_cmp_eq_f32_e64 vcc, |v0|, |v4| @@ -13198,7 +13198,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB12_2: ; SI-NEXT: ; implicit-def: $vgpr8 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB12_3: ; %frem.compute +; SI-NEXT: .LBB12_3: ; %frem.compute77 ; SI-NEXT: s_mov_b32 s6, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 s[2:3], |v0|, s6 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v8, v0 @@ -13234,10 +13234,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v10, v10, v8, 1.0 ; SI-NEXT: s_cmp_lt_i32 s3, 13 ; SI-NEXT: s_cbranch_scc1 .LBB12_7 -; SI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; SI-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; SI-NEXT: s_sub_i32 s3, s4, s5 ; SI-NEXT: s_add_i32 s3, s3, 12 -; SI-NEXT: .LBB12_5: ; %frem.loop_body +; SI-NEXT: .LBB12_5: ; %frem.loop_body85 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v11, v9 ; SI-NEXT: v_mul_f32_e32 v9, v11, v10 @@ -13252,7 +13252,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB12_5 ; SI-NEXT: ; %bb.6: ; %Flow125 ; SI-NEXT: v_mov_b32_e32 v9, v11 -; SI-NEXT: .LBB12_7: ; %frem.loop_exit +; SI-NEXT: .LBB12_7: ; %frem.loop_exit86 ; SI-NEXT: s_add_i32 s3, s3, -11 ; SI-NEXT: v_ldexp_f32_e64 v9, v9, s3 ; SI-NEXT: v_mul_f32_e32 v10, v9, v10 @@ -13268,7 +13268,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v1|, |v5| ; SI-NEXT: s_and_b64 vcc, exec, s[2:3] ; SI-NEXT: s_cbranch_vccz .LBB12_10 -; SI-NEXT: ; %bb.9: ; %frem.else16 +; SI-NEXT: ; %bb.9: ; %frem.else47 ; SI-NEXT: s_brev_b32 s2, -2 ; SI-NEXT: v_bfi_b32 v9, s2, 0, v1 ; SI-NEXT: v_cmp_eq_f32_e64 vcc, |v1|, |v5| @@ -13279,7 +13279,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB12_10: ; SI-NEXT: ; implicit-def: $vgpr9 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB12_11: ; %frem.compute15 +; SI-NEXT: .LBB12_11: ; %frem.compute46 ; SI-NEXT: s_mov_b32 s6, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 s[2:3], |v1|, s6 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v9, v1 @@ -13315,10 +13315,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v11, v11, v9, 1.0 ; SI-NEXT: s_cmp_lt_i32 s3, 13 ; SI-NEXT: s_cbranch_scc1 .LBB12_15 -; SI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; SI-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; SI-NEXT: s_sub_i32 s3, s4, s5 ; SI-NEXT: s_add_i32 s3, s3, 12 -; SI-NEXT: .LBB12_13: ; %frem.loop_body23 +; SI-NEXT: .LBB12_13: ; %frem.loop_body54 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v12, v10 ; SI-NEXT: v_mul_f32_e32 v10, v12, v11 @@ -13333,7 +13333,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB12_13 ; SI-NEXT: ; %bb.14: ; %Flow121 ; SI-NEXT: v_mov_b32_e32 v10, v12 -; SI-NEXT: .LBB12_15: ; %frem.loop_exit24 +; SI-NEXT: .LBB12_15: ; %frem.loop_exit55 ; SI-NEXT: s_add_i32 s3, s3, -11 ; SI-NEXT: v_ldexp_f32_e64 v10, v10, s3 ; SI-NEXT: v_mul_f32_e32 v11, v10, v11 @@ -13349,7 +13349,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v2|, |v6| ; SI-NEXT: s_and_b64 vcc, exec, s[2:3] ; SI-NEXT: s_cbranch_vccz .LBB12_18 -; SI-NEXT: ; %bb.17: ; %frem.else47 +; SI-NEXT: ; %bb.17: ; %frem.else16 ; SI-NEXT: s_brev_b32 s2, -2 ; SI-NEXT: v_bfi_b32 v10, s2, 0, v2 ; SI-NEXT: v_cmp_eq_f32_e64 vcc, |v2|, |v6| @@ -13360,7 +13360,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB12_18: ; SI-NEXT: ; implicit-def: $vgpr10 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB12_19: ; %frem.compute46 +; SI-NEXT: .LBB12_19: ; %frem.compute15 ; SI-NEXT: s_mov_b32 s6, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 s[2:3], |v2|, s6 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v10, v2 @@ -13396,10 +13396,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v12, v12, v10, 1.0 ; SI-NEXT: s_cmp_lt_i32 s3, 13 ; SI-NEXT: s_cbranch_scc1 .LBB12_23 -; SI-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; SI-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; SI-NEXT: s_sub_i32 s3, s4, s5 ; SI-NEXT: s_add_i32 s3, s3, 12 -; SI-NEXT: .LBB12_21: ; %frem.loop_body54 +; SI-NEXT: .LBB12_21: ; %frem.loop_body23 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v13, v11 ; SI-NEXT: v_mul_f32_e32 v11, v13, v12 @@ -13414,7 +13414,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB12_21 ; SI-NEXT: ; %bb.22: ; %Flow117 ; SI-NEXT: v_mov_b32_e32 v11, v13 -; SI-NEXT: .LBB12_23: ; %frem.loop_exit55 +; SI-NEXT: .LBB12_23: ; %frem.loop_exit24 ; SI-NEXT: s_add_i32 s3, s3, -11 ; SI-NEXT: v_ldexp_f32_e64 v11, v11, s3 ; SI-NEXT: v_mul_f32_e32 v12, v11, v12 @@ -13430,7 +13430,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v3|, |v7| ; SI-NEXT: s_and_b64 vcc, exec, s[2:3] ; SI-NEXT: s_cbranch_vccz .LBB12_26 -; SI-NEXT: ; %bb.25: ; %frem.else78 +; SI-NEXT: ; %bb.25: ; %frem.else ; SI-NEXT: s_brev_b32 s2, -2 ; SI-NEXT: v_bfi_b32 v11, s2, 0, v3 ; SI-NEXT: v_cmp_eq_f32_e64 vcc, |v3|, |v7| @@ -13441,7 +13441,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB12_26: ; SI-NEXT: ; implicit-def: $vgpr11 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB12_27: ; %frem.compute77 +; SI-NEXT: .LBB12_27: ; %frem.compute ; SI-NEXT: s_mov_b32 s6, 0x7f800000 ; SI-NEXT: v_cmp_lt_f32_e64 s[2:3], |v3|, s6 ; SI-NEXT: v_frexp_exp_i32_f32_e32 v11, v3 @@ -13477,10 +13477,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f32 v13, v13, v11, 1.0 ; SI-NEXT: s_cmp_lt_i32 s3, 13 ; SI-NEXT: s_cbranch_scc1 .LBB12_31 -; SI-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; SI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; SI-NEXT: s_sub_i32 s3, s4, s5 ; SI-NEXT: s_add_i32 s3, s3, 12 -; SI-NEXT: .LBB12_29: ; %frem.loop_body85 +; SI-NEXT: .LBB12_29: ; %frem.loop_body ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v14, v12 ; SI-NEXT: v_mul_f32_e32 v12, v14, v13 @@ -13495,7 +13495,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: s_cbranch_scc1 .LBB12_29 ; SI-NEXT: ; %bb.30: ; %Flow ; SI-NEXT: v_mov_b32_e32 v12, v14 -; SI-NEXT: .LBB12_31: ; %frem.loop_exit86 +; SI-NEXT: .LBB12_31: ; %frem.loop_exit ; SI-NEXT: s_add_i32 s3, s3, -11 ; SI-NEXT: v_ldexp_f32_e64 v12, v12, s3 ; SI-NEXT: v_mul_f32_e32 v13, v12, v13 @@ -13548,7 +13548,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v0|, |v4| ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB12_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else78 ; CI-NEXT: s_brev_b32 s2, -2 ; CI-NEXT: v_bfi_b32 v8, s2, 0, v0 ; CI-NEXT: v_cmp_eq_f32_e64 vcc, |v0|, |v4| @@ -13557,7 +13557,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB12_8 ; CI-NEXT: .LBB12_2: ; CI-NEXT: ; implicit-def: $vgpr8 -; CI-NEXT: .LBB12_3: ; %frem.compute +; CI-NEXT: .LBB12_3: ; %frem.compute77 ; CI-NEXT: v_frexp_mant_f32_e64 v9, |v4| ; CI-NEXT: v_ldexp_f32_e64 v9, v9, 1 ; CI-NEXT: v_div_scale_f32 v15, s[2:3], v9, v9, 1.0 @@ -13582,10 +13582,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v10 ; CI-NEXT: v_div_fixup_f32 v12, v12, v9, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB12_7 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; CI-NEXT: v_sub_i32_e32 v10, vcc, v13, v14 ; CI-NEXT: v_add_i32_e32 v10, vcc, 12, v10 -; CI-NEXT: .LBB12_5: ; %frem.loop_body +; CI-NEXT: .LBB12_5: ; %frem.loop_body85 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v13, v11 ; CI-NEXT: v_mul_f32_e32 v11, v13, v12 @@ -13600,7 +13600,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB12_5 ; CI-NEXT: ; %bb.6: ; %Flow125 ; CI-NEXT: v_mov_b32_e32 v11, v13 -; CI-NEXT: .LBB12_7: ; %frem.loop_exit +; CI-NEXT: .LBB12_7: ; %frem.loop_exit86 ; CI-NEXT: v_add_i32_e32 v10, vcc, -11, v10 ; CI-NEXT: v_ldexp_f32_e32 v10, v11, v10 ; CI-NEXT: v_mul_f32_e32 v11, v10, v12 @@ -13616,7 +13616,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v1|, |v5| ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB12_10 -; CI-NEXT: ; %bb.9: ; %frem.else16 +; CI-NEXT: ; %bb.9: ; %frem.else47 ; CI-NEXT: s_brev_b32 s2, -2 ; CI-NEXT: v_bfi_b32 v9, s2, 0, v1 ; CI-NEXT: v_cmp_eq_f32_e64 vcc, |v1|, |v5| @@ -13625,7 +13625,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB12_16 ; CI-NEXT: .LBB12_10: ; CI-NEXT: ; implicit-def: $vgpr9 -; CI-NEXT: .LBB12_11: ; %frem.compute15 +; CI-NEXT: .LBB12_11: ; %frem.compute46 ; CI-NEXT: v_frexp_mant_f32_e64 v10, |v5| ; CI-NEXT: v_ldexp_f32_e64 v10, v10, 1 ; CI-NEXT: v_div_scale_f32 v16, s[2:3], v10, v10, 1.0 @@ -13650,10 +13650,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v11 ; CI-NEXT: v_div_fixup_f32 v13, v13, v10, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB12_15 -; CI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; CI-NEXT: v_sub_i32_e32 v11, vcc, v14, v15 ; CI-NEXT: v_add_i32_e32 v11, vcc, 12, v11 -; CI-NEXT: .LBB12_13: ; %frem.loop_body23 +; CI-NEXT: .LBB12_13: ; %frem.loop_body54 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v14, v12 ; CI-NEXT: v_mul_f32_e32 v12, v14, v13 @@ -13668,7 +13668,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB12_13 ; CI-NEXT: ; %bb.14: ; %Flow121 ; CI-NEXT: v_mov_b32_e32 v12, v14 -; CI-NEXT: .LBB12_15: ; %frem.loop_exit24 +; CI-NEXT: .LBB12_15: ; %frem.loop_exit55 ; CI-NEXT: v_add_i32_e32 v11, vcc, -11, v11 ; CI-NEXT: v_ldexp_f32_e32 v11, v12, v11 ; CI-NEXT: v_mul_f32_e32 v12, v11, v13 @@ -13684,7 +13684,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v2|, |v6| ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB12_18 -; CI-NEXT: ; %bb.17: ; %frem.else47 +; CI-NEXT: ; %bb.17: ; %frem.else16 ; CI-NEXT: s_brev_b32 s2, -2 ; CI-NEXT: v_bfi_b32 v10, s2, 0, v2 ; CI-NEXT: v_cmp_eq_f32_e64 vcc, |v2|, |v6| @@ -13693,7 +13693,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB12_24 ; CI-NEXT: .LBB12_18: ; CI-NEXT: ; implicit-def: $vgpr10 -; CI-NEXT: .LBB12_19: ; %frem.compute46 +; CI-NEXT: .LBB12_19: ; %frem.compute15 ; CI-NEXT: v_frexp_mant_f32_e64 v11, |v6| ; CI-NEXT: v_ldexp_f32_e64 v11, v11, 1 ; CI-NEXT: v_div_scale_f32 v17, s[2:3], v11, v11, 1.0 @@ -13718,10 +13718,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v12 ; CI-NEXT: v_div_fixup_f32 v14, v14, v11, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB12_23 -; CI-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; CI-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; CI-NEXT: v_sub_i32_e32 v12, vcc, v15, v16 ; CI-NEXT: v_add_i32_e32 v12, vcc, 12, v12 -; CI-NEXT: .LBB12_21: ; %frem.loop_body54 +; CI-NEXT: .LBB12_21: ; %frem.loop_body23 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v15, v13 ; CI-NEXT: v_mul_f32_e32 v13, v15, v14 @@ -13736,7 +13736,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB12_21 ; CI-NEXT: ; %bb.22: ; %Flow117 ; CI-NEXT: v_mov_b32_e32 v13, v15 -; CI-NEXT: .LBB12_23: ; %frem.loop_exit55 +; CI-NEXT: .LBB12_23: ; %frem.loop_exit24 ; CI-NEXT: v_add_i32_e32 v12, vcc, -11, v12 ; CI-NEXT: v_ldexp_f32_e32 v12, v13, v12 ; CI-NEXT: v_mul_f32_e32 v13, v12, v14 @@ -13752,7 +13752,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v3|, |v7| ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB12_26 -; CI-NEXT: ; %bb.25: ; %frem.else78 +; CI-NEXT: ; %bb.25: ; %frem.else ; CI-NEXT: s_brev_b32 s2, -2 ; CI-NEXT: v_bfi_b32 v11, s2, 0, v3 ; CI-NEXT: v_cmp_eq_f32_e64 vcc, |v3|, |v7| @@ -13761,7 +13761,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB12_32 ; CI-NEXT: .LBB12_26: ; CI-NEXT: ; implicit-def: $vgpr11 -; CI-NEXT: .LBB12_27: ; %frem.compute77 +; CI-NEXT: .LBB12_27: ; %frem.compute ; CI-NEXT: v_frexp_mant_f32_e64 v12, |v7| ; CI-NEXT: v_ldexp_f32_e64 v12, v12, 1 ; CI-NEXT: v_div_scale_f32 v18, s[2:3], v12, v12, 1.0 @@ -13786,10 +13786,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v13 ; CI-NEXT: v_div_fixup_f32 v15, v15, v12, 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB12_31 -; CI-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; CI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; CI-NEXT: v_sub_i32_e32 v13, vcc, v16, v17 ; CI-NEXT: v_add_i32_e32 v13, vcc, 12, v13 -; CI-NEXT: .LBB12_29: ; %frem.loop_body85 +; CI-NEXT: .LBB12_29: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v16, v14 ; CI-NEXT: v_mul_f32_e32 v14, v16, v15 @@ -13804,7 +13804,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_cbranch_vccnz .LBB12_29 ; CI-NEXT: ; %bb.30: ; %Flow ; CI-NEXT: v_mov_b32_e32 v14, v16 -; CI-NEXT: .LBB12_31: ; %frem.loop_exit86 +; CI-NEXT: .LBB12_31: ; %frem.loop_exit ; CI-NEXT: v_add_i32_e32 v13, vcc, -11, v13 ; CI-NEXT: v_ldexp_f32_e32 v13, v14, v13 ; CI-NEXT: v_mul_f32_e32 v14, v13, v15 @@ -13857,7 +13857,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v0|, |v4| ; VI-NEXT: s_and_b64 vcc, exec, s[2:3] ; VI-NEXT: s_cbranch_vccz .LBB12_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else78 ; VI-NEXT: s_brev_b32 s2, -2 ; VI-NEXT: v_bfi_b32 v8, s2, 0, v0 ; VI-NEXT: v_cmp_eq_f32_e64 vcc, |v0|, |v4| @@ -13866,7 +13866,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB12_8 ; VI-NEXT: .LBB12_2: ; VI-NEXT: ; implicit-def: $vgpr8 -; VI-NEXT: .LBB12_3: ; %frem.compute +; VI-NEXT: .LBB12_3: ; %frem.compute77 ; VI-NEXT: v_frexp_mant_f32_e64 v9, |v4| ; VI-NEXT: v_ldexp_f32 v9, v9, 1 ; VI-NEXT: v_div_scale_f32 v15, s[2:3], v9, v9, 1.0 @@ -13891,10 +13891,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v10 ; VI-NEXT: v_div_fixup_f32 v12, v12, v9, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB12_7 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; VI-NEXT: v_sub_u32_e32 v10, vcc, v13, v14 ; VI-NEXT: v_add_u32_e32 v10, vcc, 12, v10 -; VI-NEXT: .LBB12_5: ; %frem.loop_body +; VI-NEXT: .LBB12_5: ; %frem.loop_body85 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v13, v11 ; VI-NEXT: v_mul_f32_e32 v11, v13, v12 @@ -13909,7 +13909,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB12_5 ; VI-NEXT: ; %bb.6: ; %Flow125 ; VI-NEXT: v_mov_b32_e32 v11, v13 -; VI-NEXT: .LBB12_7: ; %frem.loop_exit +; VI-NEXT: .LBB12_7: ; %frem.loop_exit86 ; VI-NEXT: v_add_u32_e32 v10, vcc, -11, v10 ; VI-NEXT: v_ldexp_f32 v10, v11, v10 ; VI-NEXT: v_mul_f32_e32 v11, v10, v12 @@ -13925,7 +13925,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v1|, |v5| ; VI-NEXT: s_and_b64 vcc, exec, s[2:3] ; VI-NEXT: s_cbranch_vccz .LBB12_10 -; VI-NEXT: ; %bb.9: ; %frem.else16 +; VI-NEXT: ; %bb.9: ; %frem.else47 ; VI-NEXT: s_brev_b32 s2, -2 ; VI-NEXT: v_bfi_b32 v9, s2, 0, v1 ; VI-NEXT: v_cmp_eq_f32_e64 vcc, |v1|, |v5| @@ -13934,7 +13934,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB12_16 ; VI-NEXT: .LBB12_10: ; VI-NEXT: ; implicit-def: $vgpr9 -; VI-NEXT: .LBB12_11: ; %frem.compute15 +; VI-NEXT: .LBB12_11: ; %frem.compute46 ; VI-NEXT: v_frexp_mant_f32_e64 v10, |v5| ; VI-NEXT: v_ldexp_f32 v10, v10, 1 ; VI-NEXT: v_div_scale_f32 v16, s[2:3], v10, v10, 1.0 @@ -13959,10 +13959,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v11 ; VI-NEXT: v_div_fixup_f32 v13, v13, v10, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB12_15 -; VI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; VI-NEXT: v_sub_u32_e32 v11, vcc, v14, v15 ; VI-NEXT: v_add_u32_e32 v11, vcc, 12, v11 -; VI-NEXT: .LBB12_13: ; %frem.loop_body23 +; VI-NEXT: .LBB12_13: ; %frem.loop_body54 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v14, v12 ; VI-NEXT: v_mul_f32_e32 v12, v14, v13 @@ -13977,7 +13977,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB12_13 ; VI-NEXT: ; %bb.14: ; %Flow121 ; VI-NEXT: v_mov_b32_e32 v12, v14 -; VI-NEXT: .LBB12_15: ; %frem.loop_exit24 +; VI-NEXT: .LBB12_15: ; %frem.loop_exit55 ; VI-NEXT: v_add_u32_e32 v11, vcc, -11, v11 ; VI-NEXT: v_ldexp_f32 v11, v12, v11 ; VI-NEXT: v_mul_f32_e32 v12, v11, v13 @@ -13993,7 +13993,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v2|, |v6| ; VI-NEXT: s_and_b64 vcc, exec, s[2:3] ; VI-NEXT: s_cbranch_vccz .LBB12_18 -; VI-NEXT: ; %bb.17: ; %frem.else47 +; VI-NEXT: ; %bb.17: ; %frem.else16 ; VI-NEXT: s_brev_b32 s2, -2 ; VI-NEXT: v_bfi_b32 v10, s2, 0, v2 ; VI-NEXT: v_cmp_eq_f32_e64 vcc, |v2|, |v6| @@ -14002,7 +14002,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB12_24 ; VI-NEXT: .LBB12_18: ; VI-NEXT: ; implicit-def: $vgpr10 -; VI-NEXT: .LBB12_19: ; %frem.compute46 +; VI-NEXT: .LBB12_19: ; %frem.compute15 ; VI-NEXT: v_frexp_mant_f32_e64 v11, |v6| ; VI-NEXT: v_ldexp_f32 v11, v11, 1 ; VI-NEXT: v_div_scale_f32 v17, s[2:3], v11, v11, 1.0 @@ -14027,10 +14027,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v12 ; VI-NEXT: v_div_fixup_f32 v14, v14, v11, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB12_23 -; VI-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; VI-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; VI-NEXT: v_sub_u32_e32 v12, vcc, v15, v16 ; VI-NEXT: v_add_u32_e32 v12, vcc, 12, v12 -; VI-NEXT: .LBB12_21: ; %frem.loop_body54 +; VI-NEXT: .LBB12_21: ; %frem.loop_body23 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v15, v13 ; VI-NEXT: v_mul_f32_e32 v13, v15, v14 @@ -14045,7 +14045,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB12_21 ; VI-NEXT: ; %bb.22: ; %Flow117 ; VI-NEXT: v_mov_b32_e32 v13, v15 -; VI-NEXT: .LBB12_23: ; %frem.loop_exit55 +; VI-NEXT: .LBB12_23: ; %frem.loop_exit24 ; VI-NEXT: v_add_u32_e32 v12, vcc, -11, v12 ; VI-NEXT: v_ldexp_f32 v12, v13, v12 ; VI-NEXT: v_mul_f32_e32 v13, v12, v14 @@ -14061,7 +14061,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v3|, |v7| ; VI-NEXT: s_and_b64 vcc, exec, s[2:3] ; VI-NEXT: s_cbranch_vccz .LBB12_26 -; VI-NEXT: ; %bb.25: ; %frem.else78 +; VI-NEXT: ; %bb.25: ; %frem.else ; VI-NEXT: s_brev_b32 s2, -2 ; VI-NEXT: v_bfi_b32 v11, s2, 0, v3 ; VI-NEXT: v_cmp_eq_f32_e64 vcc, |v3|, |v7| @@ -14070,7 +14070,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB12_32 ; VI-NEXT: .LBB12_26: ; VI-NEXT: ; implicit-def: $vgpr11 -; VI-NEXT: .LBB12_27: ; %frem.compute77 +; VI-NEXT: .LBB12_27: ; %frem.compute ; VI-NEXT: v_frexp_mant_f32_e64 v12, |v7| ; VI-NEXT: v_ldexp_f32 v12, v12, 1 ; VI-NEXT: v_div_scale_f32 v18, s[2:3], v12, v12, 1.0 @@ -14095,10 +14095,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 13, v13 ; VI-NEXT: v_div_fixup_f32 v15, v15, v12, 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB12_31 -; VI-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; VI-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; VI-NEXT: v_sub_u32_e32 v13, vcc, v16, v17 ; VI-NEXT: v_add_u32_e32 v13, vcc, 12, v13 -; VI-NEXT: .LBB12_29: ; %frem.loop_body85 +; VI-NEXT: .LBB12_29: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v16, v14 ; VI-NEXT: v_mul_f32_e32 v14, v16, v15 @@ -14113,7 +14113,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_cbranch_vccnz .LBB12_29 ; VI-NEXT: ; %bb.30: ; %Flow ; VI-NEXT: v_mov_b32_e32 v14, v16 -; VI-NEXT: .LBB12_31: ; %frem.loop_exit86 +; VI-NEXT: .LBB12_31: ; %frem.loop_exit ; VI-NEXT: v_add_u32_e32 v13, vcc, -11, v13 ; VI-NEXT: v_ldexp_f32 v13, v14, v13 ; VI-NEXT: v_mul_f32_e32 v14, v13, v15 @@ -14161,7 +14161,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v0|, |v4| ; GFX9-NEXT: s_and_b64 vcc, exec, s[2:3] ; GFX9-NEXT: s_cbranch_vccz .LBB12_2 -; GFX9-NEXT: ; %bb.1: ; %frem.else +; GFX9-NEXT: ; %bb.1: ; %frem.else78 ; GFX9-NEXT: s_brev_b32 s2, -2 ; GFX9-NEXT: v_bfi_b32 v8, s2, 0, v0 ; GFX9-NEXT: v_cmp_eq_f32_e64 vcc, |v0|, |v4| @@ -14170,7 +14170,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB12_8 ; GFX9-NEXT: .LBB12_2: ; GFX9-NEXT: ; implicit-def: $vgpr8 -; GFX9-NEXT: .LBB12_3: ; %frem.compute +; GFX9-NEXT: .LBB12_3: ; %frem.compute77 ; GFX9-NEXT: v_frexp_mant_f32_e64 v9, |v4| ; GFX9-NEXT: v_ldexp_f32 v9, v9, 1 ; GFX9-NEXT: v_div_scale_f32 v15, s[2:3], v9, v9, 1.0 @@ -14195,10 +14195,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 13, v10 ; GFX9-NEXT: v_div_fixup_f32 v12, v12, v9, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB12_7 -; GFX9-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX9-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; GFX9-NEXT: v_sub_u32_e32 v10, v13, v14 ; GFX9-NEXT: v_add_u32_e32 v10, 12, v10 -; GFX9-NEXT: .LBB12_5: ; %frem.loop_body +; GFX9-NEXT: .LBB12_5: ; %frem.loop_body85 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v13, v11 ; GFX9-NEXT: v_mul_f32_e32 v11, v13, v12 @@ -14213,7 +14213,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB12_5 ; GFX9-NEXT: ; %bb.6: ; %Flow125 ; GFX9-NEXT: v_mov_b32_e32 v11, v13 -; GFX9-NEXT: .LBB12_7: ; %frem.loop_exit +; GFX9-NEXT: .LBB12_7: ; %frem.loop_exit86 ; GFX9-NEXT: v_add_u32_e32 v10, -11, v10 ; GFX9-NEXT: v_ldexp_f32 v10, v11, v10 ; GFX9-NEXT: v_mul_f32_e32 v11, v10, v12 @@ -14229,7 +14229,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v1|, |v5| ; GFX9-NEXT: s_and_b64 vcc, exec, s[2:3] ; GFX9-NEXT: s_cbranch_vccz .LBB12_10 -; GFX9-NEXT: ; %bb.9: ; %frem.else16 +; GFX9-NEXT: ; %bb.9: ; %frem.else47 ; GFX9-NEXT: s_brev_b32 s2, -2 ; GFX9-NEXT: v_bfi_b32 v9, s2, 0, v1 ; GFX9-NEXT: v_cmp_eq_f32_e64 vcc, |v1|, |v5| @@ -14238,7 +14238,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB12_16 ; GFX9-NEXT: .LBB12_10: ; GFX9-NEXT: ; implicit-def: $vgpr9 -; GFX9-NEXT: .LBB12_11: ; %frem.compute15 +; GFX9-NEXT: .LBB12_11: ; %frem.compute46 ; GFX9-NEXT: v_frexp_mant_f32_e64 v10, |v5| ; GFX9-NEXT: v_ldexp_f32 v10, v10, 1 ; GFX9-NEXT: v_div_scale_f32 v16, s[2:3], v10, v10, 1.0 @@ -14263,10 +14263,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 13, v11 ; GFX9-NEXT: v_div_fixup_f32 v13, v13, v10, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB12_15 -; GFX9-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX9-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; GFX9-NEXT: v_sub_u32_e32 v11, v14, v15 ; GFX9-NEXT: v_add_u32_e32 v11, 12, v11 -; GFX9-NEXT: .LBB12_13: ; %frem.loop_body23 +; GFX9-NEXT: .LBB12_13: ; %frem.loop_body54 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v14, v12 ; GFX9-NEXT: v_mul_f32_e32 v12, v14, v13 @@ -14281,7 +14281,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB12_13 ; GFX9-NEXT: ; %bb.14: ; %Flow121 ; GFX9-NEXT: v_mov_b32_e32 v12, v14 -; GFX9-NEXT: .LBB12_15: ; %frem.loop_exit24 +; GFX9-NEXT: .LBB12_15: ; %frem.loop_exit55 ; GFX9-NEXT: v_add_u32_e32 v11, -11, v11 ; GFX9-NEXT: v_ldexp_f32 v11, v12, v11 ; GFX9-NEXT: v_mul_f32_e32 v12, v11, v13 @@ -14297,7 +14297,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v2|, |v6| ; GFX9-NEXT: s_and_b64 vcc, exec, s[2:3] ; GFX9-NEXT: s_cbranch_vccz .LBB12_18 -; GFX9-NEXT: ; %bb.17: ; %frem.else47 +; GFX9-NEXT: ; %bb.17: ; %frem.else16 ; GFX9-NEXT: s_brev_b32 s2, -2 ; GFX9-NEXT: v_bfi_b32 v10, s2, 0, v2 ; GFX9-NEXT: v_cmp_eq_f32_e64 vcc, |v2|, |v6| @@ -14306,7 +14306,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB12_24 ; GFX9-NEXT: .LBB12_18: ; GFX9-NEXT: ; implicit-def: $vgpr10 -; GFX9-NEXT: .LBB12_19: ; %frem.compute46 +; GFX9-NEXT: .LBB12_19: ; %frem.compute15 ; GFX9-NEXT: v_frexp_mant_f32_e64 v11, |v6| ; GFX9-NEXT: v_ldexp_f32 v11, v11, 1 ; GFX9-NEXT: v_div_scale_f32 v17, s[2:3], v11, v11, 1.0 @@ -14331,10 +14331,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 13, v12 ; GFX9-NEXT: v_div_fixup_f32 v14, v14, v11, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB12_23 -; GFX9-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; GFX9-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; GFX9-NEXT: v_sub_u32_e32 v12, v15, v16 ; GFX9-NEXT: v_add_u32_e32 v12, 12, v12 -; GFX9-NEXT: .LBB12_21: ; %frem.loop_body54 +; GFX9-NEXT: .LBB12_21: ; %frem.loop_body23 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v15, v13 ; GFX9-NEXT: v_mul_f32_e32 v13, v15, v14 @@ -14349,7 +14349,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB12_21 ; GFX9-NEXT: ; %bb.22: ; %Flow117 ; GFX9-NEXT: v_mov_b32_e32 v13, v15 -; GFX9-NEXT: .LBB12_23: ; %frem.loop_exit55 +; GFX9-NEXT: .LBB12_23: ; %frem.loop_exit24 ; GFX9-NEXT: v_add_u32_e32 v12, -11, v12 ; GFX9-NEXT: v_ldexp_f32 v12, v13, v12 ; GFX9-NEXT: v_mul_f32_e32 v13, v12, v14 @@ -14365,7 +14365,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_ngt_f32_e64 s[2:3], |v3|, |v7| ; GFX9-NEXT: s_and_b64 vcc, exec, s[2:3] ; GFX9-NEXT: s_cbranch_vccz .LBB12_26 -; GFX9-NEXT: ; %bb.25: ; %frem.else78 +; GFX9-NEXT: ; %bb.25: ; %frem.else ; GFX9-NEXT: s_brev_b32 s2, -2 ; GFX9-NEXT: v_bfi_b32 v11, s2, 0, v3 ; GFX9-NEXT: v_cmp_eq_f32_e64 vcc, |v3|, |v7| @@ -14374,7 +14374,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB12_32 ; GFX9-NEXT: .LBB12_26: ; GFX9-NEXT: ; implicit-def: $vgpr11 -; GFX9-NEXT: .LBB12_27: ; %frem.compute77 +; GFX9-NEXT: .LBB12_27: ; %frem.compute ; GFX9-NEXT: v_frexp_mant_f32_e64 v12, |v7| ; GFX9-NEXT: v_ldexp_f32 v12, v12, 1 ; GFX9-NEXT: v_div_scale_f32 v18, s[2:3], v12, v12, 1.0 @@ -14399,10 +14399,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 13, v13 ; GFX9-NEXT: v_div_fixup_f32 v15, v15, v12, 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB12_31 -; GFX9-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; GFX9-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX9-NEXT: v_sub_u32_e32 v13, v16, v17 ; GFX9-NEXT: v_add_u32_e32 v13, 12, v13 -; GFX9-NEXT: .LBB12_29: ; %frem.loop_body85 +; GFX9-NEXT: .LBB12_29: ; %frem.loop_body ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v16, v14 ; GFX9-NEXT: v_mul_f32_e32 v14, v16, v15 @@ -14417,7 +14417,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_cbranch_vccnz .LBB12_29 ; GFX9-NEXT: ; %bb.30: ; %Flow ; GFX9-NEXT: v_mov_b32_e32 v14, v16 -; GFX9-NEXT: .LBB12_31: ; %frem.loop_exit86 +; GFX9-NEXT: .LBB12_31: ; %frem.loop_exit ; GFX9-NEXT: v_add_u32_e32 v13, -11, v13 ; GFX9-NEXT: v_ldexp_f32 v13, v14, v13 ; GFX9-NEXT: v_mul_f32_e32 v14, v13, v15 @@ -14466,7 +14466,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_ngt_f32_e64 s2, |v0|, |v4| ; GFX10-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX10-NEXT: s_cbranch_vccz .LBB12_2 -; GFX10-NEXT: ; %bb.1: ; %frem.else +; GFX10-NEXT: ; %bb.1: ; %frem.else78 ; GFX10-NEXT: v_bfi_b32 v8, 0x7fffffff, 0, v0 ; GFX10-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v0|, |v4| ; GFX10-NEXT: v_cndmask_b32_e32 v8, v0, v8, vcc_lo @@ -14474,7 +14474,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB12_8 ; GFX10-NEXT: .LBB12_2: ; GFX10-NEXT: ; implicit-def: $vgpr8 -; GFX10-NEXT: .LBB12_3: ; %frem.compute +; GFX10-NEXT: .LBB12_3: ; %frem.compute77 ; GFX10-NEXT: v_frexp_mant_f32_e64 v9, |v4| ; GFX10-NEXT: v_frexp_mant_f32_e64 v8, |v0| ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v11, v0 @@ -14501,10 +14501,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v12 ; GFX10-NEXT: v_div_fixup_f32 v11, v11, v9, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB12_7 -; GFX10-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX10-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 12 -; GFX10-NEXT: .LBB12_5: ; %frem.loop_body +; GFX10-NEXT: .LBB12_5: ; %frem.loop_body85 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v13, v10 ; GFX10-NEXT: s_add_i32 s2, s2, -12 @@ -14520,7 +14520,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.6: ; %Flow125 ; GFX10-NEXT: v_mov_b32_e32 v12, s2 ; GFX10-NEXT: v_mov_b32_e32 v10, v13 -; GFX10-NEXT: .LBB12_7: ; %frem.loop_exit +; GFX10-NEXT: .LBB12_7: ; %frem.loop_exit86 ; GFX10-NEXT: v_add_nc_u32_e32 v12, -11, v12 ; GFX10-NEXT: v_ldexp_f32 v10, v10, v12 ; GFX10-NEXT: v_mul_f32_e32 v11, v10, v11 @@ -14535,7 +14535,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_ngt_f32_e64 s2, |v1|, |v5| ; GFX10-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX10-NEXT: s_cbranch_vccz .LBB12_10 -; GFX10-NEXT: ; %bb.9: ; %frem.else16 +; GFX10-NEXT: ; %bb.9: ; %frem.else47 ; GFX10-NEXT: v_bfi_b32 v9, 0x7fffffff, 0, v1 ; GFX10-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v1|, |v5| ; GFX10-NEXT: v_cndmask_b32_e32 v9, v1, v9, vcc_lo @@ -14543,7 +14543,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB12_16 ; GFX10-NEXT: .LBB12_10: ; GFX10-NEXT: ; implicit-def: $vgpr9 -; GFX10-NEXT: .LBB12_11: ; %frem.compute15 +; GFX10-NEXT: .LBB12_11: ; %frem.compute46 ; GFX10-NEXT: v_frexp_mant_f32_e64 v10, |v5| ; GFX10-NEXT: v_frexp_mant_f32_e64 v9, |v1| ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v12, v1 @@ -14570,10 +14570,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v13 ; GFX10-NEXT: v_div_fixup_f32 v12, v12, v10, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB12_15 -; GFX10-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX10-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 12 -; GFX10-NEXT: .LBB12_13: ; %frem.loop_body23 +; GFX10-NEXT: .LBB12_13: ; %frem.loop_body54 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v14, v11 ; GFX10-NEXT: s_add_i32 s2, s2, -12 @@ -14589,7 +14589,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.14: ; %Flow121 ; GFX10-NEXT: v_mov_b32_e32 v13, s2 ; GFX10-NEXT: v_mov_b32_e32 v11, v14 -; GFX10-NEXT: .LBB12_15: ; %frem.loop_exit24 +; GFX10-NEXT: .LBB12_15: ; %frem.loop_exit55 ; GFX10-NEXT: v_add_nc_u32_e32 v13, -11, v13 ; GFX10-NEXT: v_ldexp_f32 v11, v11, v13 ; GFX10-NEXT: v_mul_f32_e32 v12, v11, v12 @@ -14604,7 +14604,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_ngt_f32_e64 s2, |v2|, |v6| ; GFX10-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX10-NEXT: s_cbranch_vccz .LBB12_18 -; GFX10-NEXT: ; %bb.17: ; %frem.else47 +; GFX10-NEXT: ; %bb.17: ; %frem.else16 ; GFX10-NEXT: v_bfi_b32 v10, 0x7fffffff, 0, v2 ; GFX10-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v2|, |v6| ; GFX10-NEXT: v_cndmask_b32_e32 v10, v2, v10, vcc_lo @@ -14612,7 +14612,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB12_24 ; GFX10-NEXT: .LBB12_18: ; GFX10-NEXT: ; implicit-def: $vgpr10 -; GFX10-NEXT: .LBB12_19: ; %frem.compute46 +; GFX10-NEXT: .LBB12_19: ; %frem.compute15 ; GFX10-NEXT: v_frexp_mant_f32_e64 v11, |v6| ; GFX10-NEXT: v_frexp_mant_f32_e64 v10, |v2| ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v13, v2 @@ -14639,10 +14639,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v14 ; GFX10-NEXT: v_div_fixup_f32 v13, v13, v11, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB12_23 -; GFX10-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; GFX10-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 12 -; GFX10-NEXT: .LBB12_21: ; %frem.loop_body54 +; GFX10-NEXT: .LBB12_21: ; %frem.loop_body23 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v15, v12 ; GFX10-NEXT: s_add_i32 s2, s2, -12 @@ -14658,7 +14658,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.22: ; %Flow117 ; GFX10-NEXT: v_mov_b32_e32 v14, s2 ; GFX10-NEXT: v_mov_b32_e32 v12, v15 -; GFX10-NEXT: .LBB12_23: ; %frem.loop_exit55 +; GFX10-NEXT: .LBB12_23: ; %frem.loop_exit24 ; GFX10-NEXT: v_add_nc_u32_e32 v14, -11, v14 ; GFX10-NEXT: v_ldexp_f32 v12, v12, v14 ; GFX10-NEXT: v_mul_f32_e32 v13, v12, v13 @@ -14673,7 +14673,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_ngt_f32_e64 s2, |v3|, |v7| ; GFX10-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX10-NEXT: s_cbranch_vccz .LBB12_26 -; GFX10-NEXT: ; %bb.25: ; %frem.else78 +; GFX10-NEXT: ; %bb.25: ; %frem.else ; GFX10-NEXT: v_bfi_b32 v11, 0x7fffffff, 0, v3 ; GFX10-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v3|, |v7| ; GFX10-NEXT: v_cndmask_b32_e32 v11, v3, v11, vcc_lo @@ -14681,7 +14681,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB12_32 ; GFX10-NEXT: .LBB12_26: ; GFX10-NEXT: ; implicit-def: $vgpr11 -; GFX10-NEXT: .LBB12_27: ; %frem.compute77 +; GFX10-NEXT: .LBB12_27: ; %frem.compute ; GFX10-NEXT: v_frexp_mant_f32_e64 v12, |v7| ; GFX10-NEXT: v_frexp_mant_f32_e64 v11, |v3| ; GFX10-NEXT: v_frexp_exp_i32_f32_e32 v14, v3 @@ -14708,10 +14708,10 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v15 ; GFX10-NEXT: v_div_fixup_f32 v14, v14, v12, 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB12_31 -; GFX10-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; GFX10-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 12 -; GFX10-NEXT: .LBB12_29: ; %frem.loop_body85 +; GFX10-NEXT: .LBB12_29: ; %frem.loop_body ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v16, v13 ; GFX10-NEXT: s_add_i32 s2, s2, -12 @@ -14727,7 +14727,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: ; %bb.30: ; %Flow ; GFX10-NEXT: v_mov_b32_e32 v15, s2 ; GFX10-NEXT: v_mov_b32_e32 v13, v16 -; GFX10-NEXT: .LBB12_31: ; %frem.loop_exit86 +; GFX10-NEXT: .LBB12_31: ; %frem.loop_exit ; GFX10-NEXT: v_add_nc_u32_e32 v15, -11, v15 ; GFX10-NEXT: v_ldexp_f32 v13, v13, v15 ; GFX10-NEXT: v_mul_f32_e32 v14, v13, v14 @@ -14773,7 +14773,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: v_cmp_ngt_f32_e64 s2, |v0|, |v4| ; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX11-NEXT: s_cbranch_vccz .LBB12_2 -; GFX11-NEXT: ; %bb.1: ; %frem.else +; GFX11-NEXT: ; %bb.1: ; %frem.else78 ; GFX11-NEXT: v_bfi_b32 v8, 0x7fffffff, 0, v0 ; GFX11-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v0|, |v4| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -14782,7 +14782,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_branch .LBB12_8 ; GFX11-NEXT: .LBB12_2: ; GFX11-NEXT: ; implicit-def: $vgpr8 -; GFX11-NEXT: .LBB12_3: ; %frem.compute +; GFX11-NEXT: .LBB12_3: ; %frem.compute77 ; GFX11-NEXT: v_frexp_mant_f32_e64 v9, |v4| ; GFX11-NEXT: v_frexp_mant_f32_e64 v8, |v0| ; GFX11-NEXT: v_frexp_exp_i32_f32_e32 v11, v0 @@ -14818,11 +14818,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_div_fixup_f32 v11, v11, v9, 1.0 ; GFX11-NEXT: s_cbranch_vccnz .LBB12_7 -; GFX11-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX11-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; GFX11-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_add_i32 s2, s2, 12 -; GFX11-NEXT: .LBB12_5: ; %frem.loop_body +; GFX11-NEXT: .LBB12_5: ; %frem.loop_body85 ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-NEXT: v_mov_b32_e32 v13, v10 @@ -14842,7 +14842,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: ; %bb.6: ; %Flow125 ; GFX11-NEXT: v_mov_b32_e32 v12, s2 ; GFX11-NEXT: v_mov_b32_e32 v10, v13 -; GFX11-NEXT: .LBB12_7: ; %frem.loop_exit +; GFX11-NEXT: .LBB12_7: ; %frem.loop_exit86 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-NEXT: v_add_nc_u32_e32 v12, -11, v12 ; GFX11-NEXT: v_ldexp_f32 v10, v10, v12 @@ -14862,7 +14862,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: v_cmp_ngt_f32_e64 s2, |v1|, |v5| ; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX11-NEXT: s_cbranch_vccz .LBB12_10 -; GFX11-NEXT: ; %bb.9: ; %frem.else16 +; GFX11-NEXT: ; %bb.9: ; %frem.else47 ; GFX11-NEXT: v_bfi_b32 v9, 0x7fffffff, 0, v1 ; GFX11-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v1|, |v5| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -14871,7 +14871,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_branch .LBB12_16 ; GFX11-NEXT: .LBB12_10: ; GFX11-NEXT: ; implicit-def: $vgpr9 -; GFX11-NEXT: .LBB12_11: ; %frem.compute15 +; GFX11-NEXT: .LBB12_11: ; %frem.compute46 ; GFX11-NEXT: v_frexp_mant_f32_e64 v10, |v5| ; GFX11-NEXT: v_frexp_mant_f32_e64 v9, |v1| ; GFX11-NEXT: v_frexp_exp_i32_f32_e32 v12, v1 @@ -14907,11 +14907,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_div_fixup_f32 v12, v12, v10, 1.0 ; GFX11-NEXT: s_cbranch_vccnz .LBB12_15 -; GFX11-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX11-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; GFX11-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_add_i32 s2, s2, 12 -; GFX11-NEXT: .LBB12_13: ; %frem.loop_body23 +; GFX11-NEXT: .LBB12_13: ; %frem.loop_body54 ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-NEXT: v_mov_b32_e32 v14, v11 @@ -14931,7 +14931,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: ; %bb.14: ; %Flow121 ; GFX11-NEXT: v_mov_b32_e32 v13, s2 ; GFX11-NEXT: v_mov_b32_e32 v11, v14 -; GFX11-NEXT: .LBB12_15: ; %frem.loop_exit24 +; GFX11-NEXT: .LBB12_15: ; %frem.loop_exit55 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-NEXT: v_add_nc_u32_e32 v13, -11, v13 ; GFX11-NEXT: v_ldexp_f32 v11, v11, v13 @@ -14951,7 +14951,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: v_cmp_ngt_f32_e64 s2, |v2|, |v6| ; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX11-NEXT: s_cbranch_vccz .LBB12_18 -; GFX11-NEXT: ; %bb.17: ; %frem.else47 +; GFX11-NEXT: ; %bb.17: ; %frem.else16 ; GFX11-NEXT: v_bfi_b32 v10, 0x7fffffff, 0, v2 ; GFX11-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v2|, |v6| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -14960,7 +14960,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_branch .LBB12_24 ; GFX11-NEXT: .LBB12_18: ; GFX11-NEXT: ; implicit-def: $vgpr10 -; GFX11-NEXT: .LBB12_19: ; %frem.compute46 +; GFX11-NEXT: .LBB12_19: ; %frem.compute15 ; GFX11-NEXT: v_frexp_mant_f32_e64 v11, |v6| ; GFX11-NEXT: v_frexp_mant_f32_e64 v10, |v2| ; GFX11-NEXT: v_frexp_exp_i32_f32_e32 v13, v2 @@ -14996,11 +14996,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_div_fixup_f32 v13, v13, v11, 1.0 ; GFX11-NEXT: s_cbranch_vccnz .LBB12_23 -; GFX11-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; GFX11-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; GFX11-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_add_i32 s2, s2, 12 -; GFX11-NEXT: .LBB12_21: ; %frem.loop_body54 +; GFX11-NEXT: .LBB12_21: ; %frem.loop_body23 ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-NEXT: v_mov_b32_e32 v15, v12 @@ -15020,7 +15020,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: ; %bb.22: ; %Flow117 ; GFX11-NEXT: v_mov_b32_e32 v14, s2 ; GFX11-NEXT: v_mov_b32_e32 v12, v15 -; GFX11-NEXT: .LBB12_23: ; %frem.loop_exit55 +; GFX11-NEXT: .LBB12_23: ; %frem.loop_exit24 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-NEXT: v_add_nc_u32_e32 v14, -11, v14 ; GFX11-NEXT: v_ldexp_f32 v12, v12, v14 @@ -15040,7 +15040,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: v_cmp_ngt_f32_e64 s2, |v3|, |v7| ; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX11-NEXT: s_cbranch_vccz .LBB12_26 -; GFX11-NEXT: ; %bb.25: ; %frem.else78 +; GFX11-NEXT: ; %bb.25: ; %frem.else ; GFX11-NEXT: v_bfi_b32 v11, 0x7fffffff, 0, v3 ; GFX11-NEXT: v_cmp_eq_f32_e64 vcc_lo, |v3|, |v7| ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) @@ -15049,7 +15049,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_branch .LBB12_32 ; GFX11-NEXT: .LBB12_26: ; GFX11-NEXT: ; implicit-def: $vgpr11 -; GFX11-NEXT: .LBB12_27: ; %frem.compute77 +; GFX11-NEXT: .LBB12_27: ; %frem.compute ; GFX11-NEXT: v_frexp_mant_f32_e64 v12, |v7| ; GFX11-NEXT: v_frexp_mant_f32_e64 v11, |v3| ; GFX11-NEXT: v_frexp_exp_i32_f32_e32 v14, v3 @@ -15085,11 +15085,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_div_fixup_f32 v14, v14, v12, 1.0 ; GFX11-NEXT: s_cbranch_vccnz .LBB12_31 -; GFX11-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; GFX11-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX11-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_add_i32 s2, s2, 12 -; GFX11-NEXT: .LBB12_29: ; %frem.loop_body85 +; GFX11-NEXT: .LBB12_29: ; %frem.loop_body ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-NEXT: v_mov_b32_e32 v16, v13 @@ -15109,7 +15109,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: ; %bb.30: ; %Flow ; GFX11-NEXT: v_mov_b32_e32 v15, s2 ; GFX11-NEXT: v_mov_b32_e32 v13, v16 -; GFX11-NEXT: .LBB12_31: ; %frem.loop_exit86 +; GFX11-NEXT: .LBB12_31: ; %frem.loop_exit ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-NEXT: v_add_nc_u32_e32 v15, -11, v15 ; GFX11-NEXT: v_ldexp_f32 v13, v13, v15 @@ -15170,7 +15170,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_cmp_ngt_f32 s5, s12 ; GFX1150-NEXT: s_cbranch_scc0 .LBB12_2 -; GFX1150-NEXT: ; %bb.1: ; %frem.else +; GFX1150-NEXT: ; %bb.1: ; %frem.else78 ; GFX1150-NEXT: s_cmp_eq_f32 s5, s12 ; GFX1150-NEXT: v_bfi_b32 v0, 0x7fffffff, 0, s8 ; GFX1150-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -15180,7 +15180,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_branch .LBB12_8 ; GFX1150-NEXT: .LBB12_2: ; GFX1150-NEXT: ; implicit-def: $vgpr0 -; GFX1150-NEXT: .LBB12_3: ; %frem.compute +; GFX1150-NEXT: .LBB12_3: ; %frem.compute77 ; GFX1150-NEXT: v_frexp_mant_f32_e64 v1, |s6| ; GFX1150-NEXT: v_frexp_mant_f32_e64 v0, |s8| ; GFX1150-NEXT: v_frexp_exp_i32_f32_e32 v3, s8 @@ -15215,11 +15215,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v4 ; GFX1150-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1150-NEXT: s_cbranch_vccnz .LBB12_7 -; GFX1150-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1150-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; GFX1150-NEXT: s_sub_i32 s11, s11, s12 ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_add_i32 s11, s11, 12 -; GFX1150-NEXT: .LBB12_5: ; %frem.loop_body +; GFX1150-NEXT: .LBB12_5: ; %frem.loop_body85 ; GFX1150-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-NEXT: v_mov_b32_e32 v5, v2 @@ -15241,7 +15241,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: ; %bb.6: ; %Flow125 ; GFX1150-NEXT: v_mov_b32_e32 v4, s11 ; GFX1150-NEXT: v_mov_b32_e32 v2, v5 -; GFX1150-NEXT: .LBB12_7: ; %frem.loop_exit +; GFX1150-NEXT: .LBB12_7: ; %frem.loop_exit86 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-NEXT: v_add_nc_u32_e32 v4, -11, v4 ; GFX1150-NEXT: v_ldexp_f32 v2, v2, v4 @@ -15264,7 +15264,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_cmp_ngt_f32 s8, s12 ; GFX1150-NEXT: s_cbranch_scc0 .LBB12_10 -; GFX1150-NEXT: ; %bb.9: ; %frem.else16 +; GFX1150-NEXT: ; %bb.9: ; %frem.else47 ; GFX1150-NEXT: s_cmp_eq_f32 s8, s12 ; GFX1150-NEXT: v_bfi_b32 v1, 0x7fffffff, 0, s10 ; GFX1150-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -15274,7 +15274,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_branch .LBB12_16 ; GFX1150-NEXT: .LBB12_10: ; GFX1150-NEXT: ; implicit-def: $vgpr1 -; GFX1150-NEXT: .LBB12_11: ; %frem.compute15 +; GFX1150-NEXT: .LBB12_11: ; %frem.compute46 ; GFX1150-NEXT: v_frexp_mant_f32_e64 v2, |s4| ; GFX1150-NEXT: v_frexp_mant_f32_e64 v1, |s10| ; GFX1150-NEXT: v_frexp_exp_i32_f32_e32 v4, s10 @@ -15309,11 +15309,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v5 ; GFX1150-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1150-NEXT: s_cbranch_vccnz .LBB12_15 -; GFX1150-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX1150-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; GFX1150-NEXT: s_sub_i32 s11, s11, s12 ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_add_i32 s11, s11, 12 -; GFX1150-NEXT: .LBB12_13: ; %frem.loop_body23 +; GFX1150-NEXT: .LBB12_13: ; %frem.loop_body54 ; GFX1150-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-NEXT: v_mov_b32_e32 v6, v3 @@ -15335,7 +15335,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: ; %bb.14: ; %Flow121 ; GFX1150-NEXT: v_mov_b32_e32 v5, s11 ; GFX1150-NEXT: v_mov_b32_e32 v3, v6 -; GFX1150-NEXT: .LBB12_15: ; %frem.loop_exit24 +; GFX1150-NEXT: .LBB12_15: ; %frem.loop_exit55 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-NEXT: v_add_nc_u32_e32 v5, -11, v5 ; GFX1150-NEXT: v_ldexp_f32 v3, v3, v5 @@ -15358,7 +15358,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_cmp_ngt_f32 s10, s12 ; GFX1150-NEXT: s_cbranch_scc0 .LBB12_18 -; GFX1150-NEXT: ; %bb.17: ; %frem.else47 +; GFX1150-NEXT: ; %bb.17: ; %frem.else16 ; GFX1150-NEXT: s_cmp_eq_f32 s10, s12 ; GFX1150-NEXT: v_bfi_b32 v2, 0x7fffffff, 0, s9 ; GFX1150-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -15368,7 +15368,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_branch .LBB12_24 ; GFX1150-NEXT: .LBB12_18: ; GFX1150-NEXT: ; implicit-def: $vgpr2 -; GFX1150-NEXT: .LBB12_19: ; %frem.compute46 +; GFX1150-NEXT: .LBB12_19: ; %frem.compute15 ; GFX1150-NEXT: v_frexp_mant_f32_e64 v3, |s3| ; GFX1150-NEXT: v_frexp_mant_f32_e64 v2, |s9| ; GFX1150-NEXT: v_frexp_exp_i32_f32_e32 v5, s9 @@ -15403,11 +15403,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v6 ; GFX1150-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX1150-NEXT: s_cbranch_vccnz .LBB12_23 -; GFX1150-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; GFX1150-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; GFX1150-NEXT: s_sub_i32 s11, s11, s12 ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_add_i32 s11, s11, 12 -; GFX1150-NEXT: .LBB12_21: ; %frem.loop_body54 +; GFX1150-NEXT: .LBB12_21: ; %frem.loop_body23 ; GFX1150-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-NEXT: v_mov_b32_e32 v7, v4 @@ -15429,7 +15429,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: ; %bb.22: ; %Flow117 ; GFX1150-NEXT: v_mov_b32_e32 v6, s11 ; GFX1150-NEXT: v_mov_b32_e32 v4, v7 -; GFX1150-NEXT: .LBB12_23: ; %frem.loop_exit55 +; GFX1150-NEXT: .LBB12_23: ; %frem.loop_exit24 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-NEXT: v_add_nc_u32_e32 v6, -11, v6 ; GFX1150-NEXT: v_ldexp_f32 v4, v4, v6 @@ -15452,7 +15452,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_cmp_ngt_f32 s9, s12 ; GFX1150-NEXT: s_cbranch_scc0 .LBB12_26 -; GFX1150-NEXT: ; %bb.25: ; %frem.else78 +; GFX1150-NEXT: ; %bb.25: ; %frem.else ; GFX1150-NEXT: s_cmp_eq_f32 s9, s12 ; GFX1150-NEXT: v_bfi_b32 v3, 0x7fffffff, 0, s7 ; GFX1150-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -15462,7 +15462,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_branch .LBB12_32 ; GFX1150-NEXT: .LBB12_26: ; GFX1150-NEXT: ; implicit-def: $vgpr3 -; GFX1150-NEXT: .LBB12_27: ; %frem.compute77 +; GFX1150-NEXT: .LBB12_27: ; %frem.compute ; GFX1150-NEXT: v_frexp_mant_f32_e64 v4, |s2| ; GFX1150-NEXT: v_frexp_mant_f32_e64 v3, |s7| ; GFX1150-NEXT: v_frexp_exp_i32_f32_e32 v6, s7 @@ -15497,11 +15497,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v7 ; GFX1150-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; GFX1150-NEXT: s_cbranch_vccnz .LBB12_31 -; GFX1150-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; GFX1150-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX1150-NEXT: s_sub_i32 s11, s11, s12 ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_add_i32 s11, s11, 12 -; GFX1150-NEXT: .LBB12_29: ; %frem.loop_body85 +; GFX1150-NEXT: .LBB12_29: ; %frem.loop_body ; GFX1150-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-NEXT: v_mov_b32_e32 v8, v5 @@ -15523,7 +15523,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: ; %bb.30: ; %Flow ; GFX1150-NEXT: v_mov_b32_e32 v7, s11 ; GFX1150-NEXT: v_mov_b32_e32 v5, v8 -; GFX1150-NEXT: .LBB12_31: ; %frem.loop_exit86 +; GFX1150-NEXT: .LBB12_31: ; %frem.loop_exit ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-NEXT: v_add_nc_u32_e32 v7, -11, v7 ; GFX1150-NEXT: v_ldexp_f32 v5, v5, v7 @@ -15597,7 +15597,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1200-NEXT: s_cmp_ngt_f32 s5, s12 ; GFX1200-NEXT: s_cbranch_scc0 .LBB12_2 -; GFX1200-NEXT: ; %bb.1: ; %frem.else +; GFX1200-NEXT: ; %bb.1: ; %frem.else78 ; GFX1200-NEXT: s_cmp_eq_f32 s5, s12 ; GFX1200-NEXT: v_bfi_b32 v0, 0x7fffffff, 0, s8 ; GFX1200-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -15607,7 +15607,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_branch .LBB12_8 ; GFX1200-NEXT: .LBB12_2: ; GFX1200-NEXT: ; implicit-def: $vgpr0 -; GFX1200-NEXT: .LBB12_3: ; %frem.compute +; GFX1200-NEXT: .LBB12_3: ; %frem.compute77 ; GFX1200-NEXT: v_frexp_mant_f32_e64 v1, |s6| ; GFX1200-NEXT: v_frexp_mant_f32_e64 v0, |s8| ; GFX1200-NEXT: v_frexp_exp_i32_f32_e32 v3, s8 @@ -15643,11 +15643,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v4 ; GFX1200-NEXT: v_div_fixup_f32 v3, v3, v1, 1.0 ; GFX1200-NEXT: s_cbranch_vccnz .LBB12_7 -; GFX1200-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1200-NEXT: ; %bb.4: ; %frem.loop_body85.preheader ; GFX1200-NEXT: s_sub_co_i32 s11, s11, s12 ; GFX1200-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1200-NEXT: s_add_co_i32 s11, s11, 12 -; GFX1200-NEXT: .LBB12_5: ; %frem.loop_body +; GFX1200-NEXT: .LBB12_5: ; %frem.loop_body85 ; GFX1200-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1200-NEXT: v_mov_b32_e32 v5, v2 @@ -15670,7 +15670,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: ; %bb.6: ; %Flow125 ; GFX1200-NEXT: v_mov_b32_e32 v4, s11 ; GFX1200-NEXT: v_mov_b32_e32 v2, v5 -; GFX1200-NEXT: .LBB12_7: ; %frem.loop_exit +; GFX1200-NEXT: .LBB12_7: ; %frem.loop_exit86 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-NEXT: v_add_nc_u32_e32 v4, -11, v4 ; GFX1200-NEXT: v_ldexp_f32 v2, v2, v4 @@ -15694,7 +15694,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_cmp_ngt_f32 s8, s12 ; GFX1200-NEXT: s_cbranch_scc0 .LBB12_10 -; GFX1200-NEXT: ; %bb.9: ; %frem.else16 +; GFX1200-NEXT: ; %bb.9: ; %frem.else47 ; GFX1200-NEXT: s_cmp_eq_f32 s8, s12 ; GFX1200-NEXT: v_bfi_b32 v1, 0x7fffffff, 0, s10 ; GFX1200-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -15705,7 +15705,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_branch .LBB12_16 ; GFX1200-NEXT: .LBB12_10: ; GFX1200-NEXT: ; implicit-def: $vgpr1 -; GFX1200-NEXT: .LBB12_11: ; %frem.compute15 +; GFX1200-NEXT: .LBB12_11: ; %frem.compute46 ; GFX1200-NEXT: v_frexp_mant_f32_e64 v2, |s4| ; GFX1200-NEXT: v_frexp_mant_f32_e64 v1, |s10| ; GFX1200-NEXT: v_frexp_exp_i32_f32_e32 v4, s10 @@ -15741,11 +15741,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v5 ; GFX1200-NEXT: v_div_fixup_f32 v4, v4, v2, 1.0 ; GFX1200-NEXT: s_cbranch_vccnz .LBB12_15 -; GFX1200-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX1200-NEXT: ; %bb.12: ; %frem.loop_body54.preheader ; GFX1200-NEXT: s_sub_co_i32 s11, s11, s12 ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_add_co_i32 s11, s11, 12 -; GFX1200-NEXT: .LBB12_13: ; %frem.loop_body23 +; GFX1200-NEXT: .LBB12_13: ; %frem.loop_body54 ; GFX1200-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-NEXT: v_mov_b32_e32 v6, v3 @@ -15769,7 +15769,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: ; %bb.14: ; %Flow121 ; GFX1200-NEXT: v_mov_b32_e32 v5, s11 ; GFX1200-NEXT: v_mov_b32_e32 v3, v6 -; GFX1200-NEXT: .LBB12_15: ; %frem.loop_exit24 +; GFX1200-NEXT: .LBB12_15: ; %frem.loop_exit55 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-NEXT: v_add_nc_u32_e32 v5, -11, v5 ; GFX1200-NEXT: v_ldexp_f32 v3, v3, v5 @@ -15793,7 +15793,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_cmp_ngt_f32 s10, s12 ; GFX1200-NEXT: s_cbranch_scc0 .LBB12_18 -; GFX1200-NEXT: ; %bb.17: ; %frem.else47 +; GFX1200-NEXT: ; %bb.17: ; %frem.else16 ; GFX1200-NEXT: s_cmp_eq_f32 s10, s12 ; GFX1200-NEXT: v_bfi_b32 v2, 0x7fffffff, 0, s9 ; GFX1200-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -15804,7 +15804,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_branch .LBB12_24 ; GFX1200-NEXT: .LBB12_18: ; GFX1200-NEXT: ; implicit-def: $vgpr2 -; GFX1200-NEXT: .LBB12_19: ; %frem.compute46 +; GFX1200-NEXT: .LBB12_19: ; %frem.compute15 ; GFX1200-NEXT: v_frexp_mant_f32_e64 v3, |s3| ; GFX1200-NEXT: v_frexp_mant_f32_e64 v2, |s9| ; GFX1200-NEXT: v_frexp_exp_i32_f32_e32 v5, s9 @@ -15840,11 +15840,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v6 ; GFX1200-NEXT: v_div_fixup_f32 v5, v5, v3, 1.0 ; GFX1200-NEXT: s_cbranch_vccnz .LBB12_23 -; GFX1200-NEXT: ; %bb.20: ; %frem.loop_body54.preheader +; GFX1200-NEXT: ; %bb.20: ; %frem.loop_body23.preheader ; GFX1200-NEXT: s_sub_co_i32 s11, s11, s12 ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_add_co_i32 s11, s11, 12 -; GFX1200-NEXT: .LBB12_21: ; %frem.loop_body54 +; GFX1200-NEXT: .LBB12_21: ; %frem.loop_body23 ; GFX1200-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-NEXT: v_mov_b32_e32 v7, v4 @@ -15868,7 +15868,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: ; %bb.22: ; %Flow117 ; GFX1200-NEXT: v_mov_b32_e32 v6, s11 ; GFX1200-NEXT: v_mov_b32_e32 v4, v7 -; GFX1200-NEXT: .LBB12_23: ; %frem.loop_exit55 +; GFX1200-NEXT: .LBB12_23: ; %frem.loop_exit24 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-NEXT: v_add_nc_u32_e32 v6, -11, v6 ; GFX1200-NEXT: v_ldexp_f32 v4, v4, v6 @@ -15892,7 +15892,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_cmp_ngt_f32 s9, s12 ; GFX1200-NEXT: s_cbranch_scc0 .LBB12_26 -; GFX1200-NEXT: ; %bb.25: ; %frem.else78 +; GFX1200-NEXT: ; %bb.25: ; %frem.else ; GFX1200-NEXT: s_cmp_eq_f32 s9, s12 ; GFX1200-NEXT: v_bfi_b32 v3, 0x7fffffff, 0, s7 ; GFX1200-NEXT: s_cselect_b32 vcc_lo, -1, 0 @@ -15903,7 +15903,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_branch .LBB12_32 ; GFX1200-NEXT: .LBB12_26: ; GFX1200-NEXT: ; implicit-def: $vgpr3 -; GFX1200-NEXT: .LBB12_27: ; %frem.compute77 +; GFX1200-NEXT: .LBB12_27: ; %frem.compute ; GFX1200-NEXT: v_frexp_mant_f32_e64 v4, |s2| ; GFX1200-NEXT: v_frexp_mant_f32_e64 v3, |s7| ; GFX1200-NEXT: v_frexp_exp_i32_f32_e32 v6, s7 @@ -15939,11 +15939,11 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_gt_i32_e32 vcc_lo, 13, v7 ; GFX1200-NEXT: v_div_fixup_f32 v6, v6, v4, 1.0 ; GFX1200-NEXT: s_cbranch_vccnz .LBB12_31 -; GFX1200-NEXT: ; %bb.28: ; %frem.loop_body85.preheader +; GFX1200-NEXT: ; %bb.28: ; %frem.loop_body.preheader ; GFX1200-NEXT: s_sub_co_i32 s11, s11, s12 ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_add_co_i32 s11, s11, 12 -; GFX1200-NEXT: .LBB12_29: ; %frem.loop_body85 +; GFX1200-NEXT: .LBB12_29: ; %frem.loop_body ; GFX1200-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX1200-NEXT: v_mov_b32_e32 v8, v5 @@ -15967,7 +15967,7 @@ define amdgpu_kernel void @frem_v4f32(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: ; %bb.30: ; %Flow ; GFX1200-NEXT: v_mov_b32_e32 v7, s11 ; GFX1200-NEXT: v_mov_b32_e32 v5, v8 -; GFX1200-NEXT: .LBB12_31: ; %frem.loop_exit86 +; GFX1200-NEXT: .LBB12_31: ; %frem.loop_exit ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-NEXT: v_add_nc_u32_e32 v7, -11, v7 ; GFX1200-NEXT: v_ldexp_f32 v5, v5, v7 @@ -16048,7 +16048,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cmp_ngt_f64_e64 s[0:1], |v[0:1]|, |v[4:5]| ; SI-NEXT: s_and_b64 vcc, exec, s[0:1] ; SI-NEXT: s_cbranch_vccz .LBB13_2 -; SI-NEXT: ; %bb.1: ; %frem.else +; SI-NEXT: ; %bb.1: ; %frem.else16 ; SI-NEXT: v_and_b32_e32 v8, 0x80000000, v1 ; SI-NEXT: v_cmp_eq_f64_e64 vcc, |v[0:1]|, |v[4:5]| ; SI-NEXT: v_cndmask_b32_e32 v9, v1, v8, vcc @@ -16059,7 +16059,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB13_2: ; SI-NEXT: ; implicit-def: $vgpr8_vgpr9 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB13_3: ; %frem.compute +; SI-NEXT: .LBB13_3: ; %frem.compute15 ; SI-NEXT: s_brev_b32 s5, -2 ; SI-NEXT: v_and_b32_e32 v10, 0x7fffffff, v1 ; SI-NEXT: s_mov_b32 s0, 0 @@ -16105,13 +16105,13 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f64 v[12:13], v[12:13], v[8:9], 1.0 ; SI-NEXT: s_cmp_lt_i32 s6, 27 ; SI-NEXT: s_cbranch_scc1 .LBB13_7 -; SI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; SI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; SI-NEXT: s_sub_i32 s0, s3, s7 ; SI-NEXT: s_add_i32 s6, s0, 26 ; SI-NEXT: s_mov_b32 s3, 0x432fffff ; SI-NEXT: v_mov_b32_e32 v18, 0x43300000 ; SI-NEXT: v_mov_b32_e32 v14, 0 -; SI-NEXT: .LBB13_5: ; %frem.loop_body +; SI-NEXT: .LBB13_5: ; %frem.loop_body23 ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v17, v11 ; SI-NEXT: v_mov_b32_e32 v16, v10 @@ -16134,7 +16134,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: ; %bb.6: ; %Flow51 ; SI-NEXT: v_mov_b32_e32 v10, v16 ; SI-NEXT: v_mov_b32_e32 v11, v17 -; SI-NEXT: .LBB13_7: ; %frem.loop_exit +; SI-NEXT: .LBB13_7: ; %frem.loop_exit24 ; SI-NEXT: s_sub_i32 s0, s6, 25 ; SI-NEXT: v_ldexp_f64 v[10:11], v[10:11], s0 ; SI-NEXT: v_mul_f64 v[12:13], v[10:11], v[12:13] @@ -16160,7 +16160,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_cmp_ngt_f64_e64 s[0:1], |v[2:3]|, |v[6:7]| ; SI-NEXT: s_and_b64 vcc, exec, s[0:1] ; SI-NEXT: s_cbranch_vccz .LBB13_10 -; SI-NEXT: ; %bb.9: ; %frem.else16 +; SI-NEXT: ; %bb.9: ; %frem.else ; SI-NEXT: v_and_b32_e32 v10, 0x80000000, v3 ; SI-NEXT: v_cmp_eq_f64_e64 vcc, |v[2:3]|, |v[6:7]| ; SI-NEXT: v_cndmask_b32_e32 v11, v3, v10, vcc @@ -16171,7 +16171,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: .LBB13_10: ; SI-NEXT: ; implicit-def: $vgpr10_vgpr11 ; SI-NEXT: s_mov_b64 vcc, 0 -; SI-NEXT: .LBB13_11: ; %frem.compute15 +; SI-NEXT: .LBB13_11: ; %frem.compute ; SI-NEXT: s_brev_b32 s5, -2 ; SI-NEXT: v_and_b32_e32 v12, 0x7fffffff, v3 ; SI-NEXT: s_mov_b32 s0, 0 @@ -16217,13 +16217,13 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: v_div_fixup_f64 v[14:15], v[14:15], v[10:11], 1.0 ; SI-NEXT: s_cmp_lt_i32 s6, 27 ; SI-NEXT: s_cbranch_scc1 .LBB13_15 -; SI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; SI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; SI-NEXT: s_sub_i32 s0, s3, s7 ; SI-NEXT: s_add_i32 s6, s0, 26 ; SI-NEXT: s_mov_b32 s3, 0x432fffff ; SI-NEXT: v_mov_b32_e32 v20, 0x43300000 ; SI-NEXT: v_mov_b32_e32 v16, 0 -; SI-NEXT: .LBB13_13: ; %frem.loop_body23 +; SI-NEXT: .LBB13_13: ; %frem.loop_body ; SI-NEXT: ; =>This Inner Loop Header: Depth=1 ; SI-NEXT: v_mov_b32_e32 v19, v13 ; SI-NEXT: v_mov_b32_e32 v18, v12 @@ -16246,7 +16246,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; SI-NEXT: ; %bb.14: ; %Flow ; SI-NEXT: v_mov_b32_e32 v12, v18 ; SI-NEXT: v_mov_b32_e32 v13, v19 -; SI-NEXT: .LBB13_15: ; %frem.loop_exit24 +; SI-NEXT: .LBB13_15: ; %frem.loop_exit ; SI-NEXT: s_sub_i32 s0, s6, 25 ; SI-NEXT: v_ldexp_f64 v[12:13], v[12:13], s0 ; SI-NEXT: v_mul_f64 v[14:15], v[12:13], v[14:15] @@ -16304,7 +16304,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f64_e64 s[2:3], |v[0:1]|, |v[4:5]| ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB13_2 -; CI-NEXT: ; %bb.1: ; %frem.else +; CI-NEXT: ; %bb.1: ; %frem.else16 ; CI-NEXT: v_cmp_eq_f64_e64 vcc, |v[0:1]|, |v[4:5]| ; CI-NEXT: v_and_b32_e32 v8, 0x80000000, v1 ; CI-NEXT: v_cndmask_b32_e32 v9, v1, v8, vcc @@ -16313,7 +16313,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB13_8 ; CI-NEXT: .LBB13_2: ; CI-NEXT: ; implicit-def: $vgpr8_vgpr9 -; CI-NEXT: .LBB13_3: ; %frem.compute +; CI-NEXT: .LBB13_3: ; %frem.compute15 ; CI-NEXT: v_frexp_mant_f64_e64 v[8:9], |v[0:1]| ; CI-NEXT: v_frexp_exp_i32_f64_e32 v15, v[4:5] ; CI-NEXT: v_frexp_exp_i32_f64_e32 v14, v[0:1] @@ -16337,10 +16337,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 27, v17 ; CI-NEXT: v_div_fixup_f64 v[12:13], v[12:13], v[8:9], 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB13_7 -; CI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; CI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; CI-NEXT: v_sub_i32_e32 v14, vcc, v14, v15 ; CI-NEXT: v_add_i32_e32 v17, vcc, 26, v14 -; CI-NEXT: .LBB13_5: ; %frem.loop_body +; CI-NEXT: .LBB13_5: ; %frem.loop_body23 ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v15, v11 ; CI-NEXT: v_mov_b32_e32 v14, v10 @@ -16358,7 +16358,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: ; %bb.6: ; %Flow51 ; CI-NEXT: v_mov_b32_e32 v10, v14 ; CI-NEXT: v_mov_b32_e32 v11, v15 -; CI-NEXT: .LBB13_7: ; %frem.loop_exit +; CI-NEXT: .LBB13_7: ; %frem.loop_exit24 ; CI-NEXT: v_subrev_i32_e32 v14, vcc, 25, v17 ; CI-NEXT: v_ldexp_f64 v[10:11], v[10:11], v14 ; CI-NEXT: s_brev_b32 s2, -2 @@ -16375,7 +16375,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_ngt_f64_e64 s[2:3], |v[2:3]|, |v[6:7]| ; CI-NEXT: s_and_b64 vcc, exec, s[2:3] ; CI-NEXT: s_cbranch_vccz .LBB13_10 -; CI-NEXT: ; %bb.9: ; %frem.else16 +; CI-NEXT: ; %bb.9: ; %frem.else ; CI-NEXT: v_cmp_eq_f64_e64 vcc, |v[2:3]|, |v[6:7]| ; CI-NEXT: v_and_b32_e32 v10, 0x80000000, v3 ; CI-NEXT: v_cndmask_b32_e32 v11, v3, v10, vcc @@ -16384,7 +16384,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: s_branch .LBB13_16 ; CI-NEXT: .LBB13_10: ; CI-NEXT: ; implicit-def: $vgpr10_vgpr11 -; CI-NEXT: .LBB13_11: ; %frem.compute15 +; CI-NEXT: .LBB13_11: ; %frem.compute ; CI-NEXT: v_frexp_mant_f64_e64 v[10:11], |v[2:3]| ; CI-NEXT: v_frexp_exp_i32_f64_e32 v17, v[6:7] ; CI-NEXT: v_frexp_exp_i32_f64_e32 v16, v[2:3] @@ -16408,10 +16408,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: v_cmp_gt_i32_e32 vcc, 27, v19 ; CI-NEXT: v_div_fixup_f64 v[14:15], v[14:15], v[10:11], 1.0 ; CI-NEXT: s_cbranch_vccnz .LBB13_15 -; CI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; CI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; CI-NEXT: v_sub_i32_e32 v16, vcc, v16, v17 ; CI-NEXT: v_add_i32_e32 v19, vcc, 26, v16 -; CI-NEXT: .LBB13_13: ; %frem.loop_body23 +; CI-NEXT: .LBB13_13: ; %frem.loop_body ; CI-NEXT: ; =>This Inner Loop Header: Depth=1 ; CI-NEXT: v_mov_b32_e32 v17, v13 ; CI-NEXT: v_mov_b32_e32 v16, v12 @@ -16429,7 +16429,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; CI-NEXT: ; %bb.14: ; %Flow ; CI-NEXT: v_mov_b32_e32 v12, v16 ; CI-NEXT: v_mov_b32_e32 v13, v17 -; CI-NEXT: .LBB13_15: ; %frem.loop_exit24 +; CI-NEXT: .LBB13_15: ; %frem.loop_exit ; CI-NEXT: v_subrev_i32_e32 v16, vcc, 25, v19 ; CI-NEXT: v_ldexp_f64 v[12:13], v[12:13], v16 ; CI-NEXT: s_brev_b32 s2, -2 @@ -16478,7 +16478,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f64_e64 s[2:3], |v[0:1]|, |v[4:5]| ; VI-NEXT: s_and_b64 vcc, exec, s[2:3] ; VI-NEXT: s_cbranch_vccz .LBB13_2 -; VI-NEXT: ; %bb.1: ; %frem.else +; VI-NEXT: ; %bb.1: ; %frem.else16 ; VI-NEXT: v_cmp_eq_f64_e64 vcc, |v[0:1]|, |v[4:5]| ; VI-NEXT: v_and_b32_e32 v8, 0x80000000, v1 ; VI-NEXT: v_cndmask_b32_e32 v9, v1, v8, vcc @@ -16487,7 +16487,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB13_8 ; VI-NEXT: .LBB13_2: ; VI-NEXT: ; implicit-def: $vgpr8_vgpr9 -; VI-NEXT: .LBB13_3: ; %frem.compute +; VI-NEXT: .LBB13_3: ; %frem.compute15 ; VI-NEXT: v_frexp_mant_f64_e64 v[8:9], |v[0:1]| ; VI-NEXT: v_frexp_exp_i32_f64_e32 v15, v[4:5] ; VI-NEXT: v_frexp_exp_i32_f64_e32 v14, v[0:1] @@ -16511,10 +16511,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 27, v17 ; VI-NEXT: v_div_fixup_f64 v[12:13], v[12:13], v[8:9], 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB13_7 -; VI-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; VI-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; VI-NEXT: v_sub_u32_e32 v14, vcc, v14, v15 ; VI-NEXT: v_add_u32_e32 v17, vcc, 26, v14 -; VI-NEXT: .LBB13_5: ; %frem.loop_body +; VI-NEXT: .LBB13_5: ; %frem.loop_body23 ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v15, v11 ; VI-NEXT: v_mov_b32_e32 v14, v10 @@ -16532,7 +16532,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: ; %bb.6: ; %Flow51 ; VI-NEXT: v_mov_b32_e32 v10, v14 ; VI-NEXT: v_mov_b32_e32 v11, v15 -; VI-NEXT: .LBB13_7: ; %frem.loop_exit +; VI-NEXT: .LBB13_7: ; %frem.loop_exit24 ; VI-NEXT: v_subrev_u32_e32 v14, vcc, 25, v17 ; VI-NEXT: v_ldexp_f64 v[10:11], v[10:11], v14 ; VI-NEXT: s_brev_b32 s2, -2 @@ -16549,7 +16549,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_ngt_f64_e64 s[2:3], |v[2:3]|, |v[6:7]| ; VI-NEXT: s_and_b64 vcc, exec, s[2:3] ; VI-NEXT: s_cbranch_vccz .LBB13_10 -; VI-NEXT: ; %bb.9: ; %frem.else16 +; VI-NEXT: ; %bb.9: ; %frem.else ; VI-NEXT: v_cmp_eq_f64_e64 vcc, |v[2:3]|, |v[6:7]| ; VI-NEXT: v_and_b32_e32 v10, 0x80000000, v3 ; VI-NEXT: v_cndmask_b32_e32 v11, v3, v10, vcc @@ -16558,7 +16558,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: s_branch .LBB13_16 ; VI-NEXT: .LBB13_10: ; VI-NEXT: ; implicit-def: $vgpr10_vgpr11 -; VI-NEXT: .LBB13_11: ; %frem.compute15 +; VI-NEXT: .LBB13_11: ; %frem.compute ; VI-NEXT: v_frexp_mant_f64_e64 v[10:11], |v[2:3]| ; VI-NEXT: v_frexp_exp_i32_f64_e32 v17, v[6:7] ; VI-NEXT: v_frexp_exp_i32_f64_e32 v16, v[2:3] @@ -16582,10 +16582,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: v_cmp_gt_i32_e32 vcc, 27, v19 ; VI-NEXT: v_div_fixup_f64 v[14:15], v[14:15], v[10:11], 1.0 ; VI-NEXT: s_cbranch_vccnz .LBB13_15 -; VI-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; VI-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; VI-NEXT: v_sub_u32_e32 v16, vcc, v16, v17 ; VI-NEXT: v_add_u32_e32 v19, vcc, 26, v16 -; VI-NEXT: .LBB13_13: ; %frem.loop_body23 +; VI-NEXT: .LBB13_13: ; %frem.loop_body ; VI-NEXT: ; =>This Inner Loop Header: Depth=1 ; VI-NEXT: v_mov_b32_e32 v17, v13 ; VI-NEXT: v_mov_b32_e32 v16, v12 @@ -16603,7 +16603,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; VI-NEXT: ; %bb.14: ; %Flow ; VI-NEXT: v_mov_b32_e32 v12, v16 ; VI-NEXT: v_mov_b32_e32 v13, v17 -; VI-NEXT: .LBB13_15: ; %frem.loop_exit24 +; VI-NEXT: .LBB13_15: ; %frem.loop_exit ; VI-NEXT: v_subrev_u32_e32 v16, vcc, 25, v19 ; VI-NEXT: v_ldexp_f64 v[12:13], v[12:13], v16 ; VI-NEXT: s_brev_b32 s2, -2 @@ -16647,7 +16647,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_ngt_f64_e64 s[2:3], |v[0:1]|, |v[4:5]| ; GFX9-NEXT: s_and_b64 vcc, exec, s[2:3] ; GFX9-NEXT: s_cbranch_vccz .LBB13_2 -; GFX9-NEXT: ; %bb.1: ; %frem.else +; GFX9-NEXT: ; %bb.1: ; %frem.else16 ; GFX9-NEXT: v_cmp_eq_f64_e64 vcc, |v[0:1]|, |v[4:5]| ; GFX9-NEXT: v_and_b32_e32 v8, 0x80000000, v1 ; GFX9-NEXT: v_cndmask_b32_e32 v9, v1, v8, vcc @@ -16656,7 +16656,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB13_8 ; GFX9-NEXT: .LBB13_2: ; GFX9-NEXT: ; implicit-def: $vgpr8_vgpr9 -; GFX9-NEXT: .LBB13_3: ; %frem.compute +; GFX9-NEXT: .LBB13_3: ; %frem.compute15 ; GFX9-NEXT: v_frexp_mant_f64_e64 v[8:9], |v[0:1]| ; GFX9-NEXT: v_frexp_exp_i32_f64_e32 v15, v[4:5] ; GFX9-NEXT: v_frexp_exp_i32_f64_e32 v14, v[0:1] @@ -16680,10 +16680,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 27, v17 ; GFX9-NEXT: v_div_fixup_f64 v[12:13], v[12:13], v[8:9], 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB13_7 -; GFX9-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX9-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX9-NEXT: v_sub_u32_e32 v14, v14, v15 ; GFX9-NEXT: v_add_u32_e32 v17, 26, v14 -; GFX9-NEXT: .LBB13_5: ; %frem.loop_body +; GFX9-NEXT: .LBB13_5: ; %frem.loop_body23 ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v15, v11 ; GFX9-NEXT: v_mov_b32_e32 v14, v10 @@ -16701,7 +16701,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: ; %bb.6: ; %Flow51 ; GFX9-NEXT: v_mov_b32_e32 v10, v14 ; GFX9-NEXT: v_mov_b32_e32 v11, v15 -; GFX9-NEXT: .LBB13_7: ; %frem.loop_exit +; GFX9-NEXT: .LBB13_7: ; %frem.loop_exit24 ; GFX9-NEXT: v_subrev_u32_e32 v14, 25, v17 ; GFX9-NEXT: v_ldexp_f64 v[10:11], v[10:11], v14 ; GFX9-NEXT: s_brev_b32 s2, -2 @@ -16718,7 +16718,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_ngt_f64_e64 s[2:3], |v[2:3]|, |v[6:7]| ; GFX9-NEXT: s_and_b64 vcc, exec, s[2:3] ; GFX9-NEXT: s_cbranch_vccz .LBB13_10 -; GFX9-NEXT: ; %bb.9: ; %frem.else16 +; GFX9-NEXT: ; %bb.9: ; %frem.else ; GFX9-NEXT: v_cmp_eq_f64_e64 vcc, |v[2:3]|, |v[6:7]| ; GFX9-NEXT: v_and_b32_e32 v10, 0x80000000, v3 ; GFX9-NEXT: v_cndmask_b32_e32 v11, v3, v10, vcc @@ -16727,7 +16727,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: s_branch .LBB13_16 ; GFX9-NEXT: .LBB13_10: ; GFX9-NEXT: ; implicit-def: $vgpr10_vgpr11 -; GFX9-NEXT: .LBB13_11: ; %frem.compute15 +; GFX9-NEXT: .LBB13_11: ; %frem.compute ; GFX9-NEXT: v_frexp_mant_f64_e64 v[10:11], |v[2:3]| ; GFX9-NEXT: v_frexp_exp_i32_f64_e32 v17, v[6:7] ; GFX9-NEXT: v_frexp_exp_i32_f64_e32 v16, v[2:3] @@ -16751,10 +16751,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: v_cmp_gt_i32_e32 vcc, 27, v19 ; GFX9-NEXT: v_div_fixup_f64 v[14:15], v[14:15], v[10:11], 1.0 ; GFX9-NEXT: s_cbranch_vccnz .LBB13_15 -; GFX9-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX9-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX9-NEXT: v_sub_u32_e32 v16, v16, v17 ; GFX9-NEXT: v_add_u32_e32 v19, 26, v16 -; GFX9-NEXT: .LBB13_13: ; %frem.loop_body23 +; GFX9-NEXT: .LBB13_13: ; %frem.loop_body ; GFX9-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX9-NEXT: v_mov_b32_e32 v17, v13 ; GFX9-NEXT: v_mov_b32_e32 v16, v12 @@ -16772,7 +16772,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX9-NEXT: ; %bb.14: ; %Flow ; GFX9-NEXT: v_mov_b32_e32 v12, v16 ; GFX9-NEXT: v_mov_b32_e32 v13, v17 -; GFX9-NEXT: .LBB13_15: ; %frem.loop_exit24 +; GFX9-NEXT: .LBB13_15: ; %frem.loop_exit ; GFX9-NEXT: v_subrev_u32_e32 v16, 25, v19 ; GFX9-NEXT: v_ldexp_f64 v[12:13], v[12:13], v16 ; GFX9-NEXT: s_brev_b32 s2, -2 @@ -16817,7 +16817,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_ngt_f64_e64 s2, |v[0:1]|, |v[4:5]| ; GFX10-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX10-NEXT: s_cbranch_vccz .LBB13_2 -; GFX10-NEXT: ; %bb.1: ; %frem.else +; GFX10-NEXT: ; %bb.1: ; %frem.else16 ; GFX10-NEXT: v_cmp_eq_f64_e64 vcc_lo, |v[0:1]|, |v[4:5]| ; GFX10-NEXT: v_and_b32_e32 v8, 0x80000000, v1 ; GFX10-NEXT: v_cndmask_b32_e32 v9, v1, v8, vcc_lo @@ -16826,7 +16826,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB13_8 ; GFX10-NEXT: .LBB13_2: ; GFX10-NEXT: ; implicit-def: $vgpr8_vgpr9 -; GFX10-NEXT: .LBB13_3: ; %frem.compute +; GFX10-NEXT: .LBB13_3: ; %frem.compute15 ; GFX10-NEXT: v_frexp_mant_f64_e64 v[8:9], |v[0:1]| ; GFX10-NEXT: v_frexp_exp_i32_f64_e32 v13, v[4:5] ; GFX10-NEXT: v_frexp_exp_i32_f64_e32 v12, v[0:1] @@ -16851,10 +16851,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 27, v17 ; GFX10-NEXT: v_div_fixup_f64 v[12:13], v[12:13], v[8:9], 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB13_7 -; GFX10-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX10-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 26 -; GFX10-NEXT: .LBB13_5: ; %frem.loop_body +; GFX10-NEXT: .LBB13_5: ; %frem.loop_body23 ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v15, v11 ; GFX10-NEXT: v_mov_b32_e32 v14, v10 @@ -16873,7 +16873,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_mov_b32_e32 v10, v14 ; GFX10-NEXT: v_mov_b32_e32 v17, s2 ; GFX10-NEXT: v_mov_b32_e32 v11, v15 -; GFX10-NEXT: .LBB13_7: ; %frem.loop_exit +; GFX10-NEXT: .LBB13_7: ; %frem.loop_exit24 ; GFX10-NEXT: v_subrev_nc_u32_e32 v14, 25, v17 ; GFX10-NEXT: v_ldexp_f64 v[10:11], v[10:11], v14 ; GFX10-NEXT: v_mul_f64 v[12:13], v[10:11], v[12:13] @@ -16889,7 +16889,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_ngt_f64_e64 s2, |v[2:3]|, |v[6:7]| ; GFX10-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX10-NEXT: s_cbranch_vccz .LBB13_10 -; GFX10-NEXT: ; %bb.9: ; %frem.else16 +; GFX10-NEXT: ; %bb.9: ; %frem.else ; GFX10-NEXT: v_cmp_eq_f64_e64 vcc_lo, |v[2:3]|, |v[6:7]| ; GFX10-NEXT: v_and_b32_e32 v10, 0x80000000, v3 ; GFX10-NEXT: v_cndmask_b32_e32 v11, v3, v10, vcc_lo @@ -16898,7 +16898,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: s_branch .LBB13_16 ; GFX10-NEXT: .LBB13_10: ; GFX10-NEXT: ; implicit-def: $vgpr10_vgpr11 -; GFX10-NEXT: .LBB13_11: ; %frem.compute15 +; GFX10-NEXT: .LBB13_11: ; %frem.compute ; GFX10-NEXT: v_frexp_mant_f64_e64 v[10:11], |v[2:3]| ; GFX10-NEXT: v_frexp_exp_i32_f64_e32 v15, v[6:7] ; GFX10-NEXT: v_frexp_exp_i32_f64_e32 v14, v[2:3] @@ -16923,10 +16923,10 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_cmp_gt_i32_e32 vcc_lo, 27, v19 ; GFX10-NEXT: v_div_fixup_f64 v[14:15], v[14:15], v[10:11], 1.0 ; GFX10-NEXT: s_cbranch_vccnz .LBB13_15 -; GFX10-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX10-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX10-NEXT: s_sub_i32 s2, s2, s3 ; GFX10-NEXT: s_add_i32 s2, s2, 26 -; GFX10-NEXT: .LBB13_13: ; %frem.loop_body23 +; GFX10-NEXT: .LBB13_13: ; %frem.loop_body ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: v_mov_b32_e32 v17, v13 ; GFX10-NEXT: v_mov_b32_e32 v16, v12 @@ -16945,7 +16945,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX10-NEXT: v_mov_b32_e32 v12, v16 ; GFX10-NEXT: v_mov_b32_e32 v19, s2 ; GFX10-NEXT: v_mov_b32_e32 v13, v17 -; GFX10-NEXT: .LBB13_15: ; %frem.loop_exit24 +; GFX10-NEXT: .LBB13_15: ; %frem.loop_exit ; GFX10-NEXT: v_subrev_nc_u32_e32 v16, 25, v19 ; GFX10-NEXT: v_ldexp_f64 v[12:13], v[12:13], v16 ; GFX10-NEXT: v_mul_f64 v[14:15], v[12:13], v[14:15] @@ -16986,7 +16986,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: v_cmp_ngt_f64_e64 s2, |v[0:1]|, |v[4:5]| ; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX11-NEXT: s_cbranch_vccz .LBB13_2 -; GFX11-NEXT: ; %bb.1: ; %frem.else +; GFX11-NEXT: ; %bb.1: ; %frem.else16 ; GFX11-NEXT: v_cmp_eq_f64_e64 vcc_lo, |v[0:1]|, |v[4:5]| ; GFX11-NEXT: v_and_b32_e32 v8, 0x80000000, v1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) @@ -16996,7 +16996,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_branch .LBB13_8 ; GFX11-NEXT: .LBB13_2: ; GFX11-NEXT: ; implicit-def: $vgpr8_vgpr9 -; GFX11-NEXT: .LBB13_3: ; %frem.compute +; GFX11-NEXT: .LBB13_3: ; %frem.compute15 ; GFX11-NEXT: v_frexp_mant_f64_e64 v[8:9], |v[0:1]| ; GFX11-NEXT: v_frexp_exp_i32_f64_e32 v13, v[4:5] ; GFX11-NEXT: v_frexp_exp_i32_f64_e32 v12, v[0:1] @@ -17029,12 +17029,12 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_div_fixup_f64 v[12:13], v[12:13], v[8:9], 1.0 ; GFX11-NEXT: s_cbranch_vccnz .LBB13_7 -; GFX11-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX11-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX11-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_add_i32 s2, s2, 26 ; GFX11-NEXT: .p2align 6 -; GFX11-NEXT: .LBB13_5: ; %frem.loop_body +; GFX11-NEXT: .LBB13_5: ; %frem.loop_body23 ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-NEXT: v_dual_mov_b32 v15, v11 :: v_dual_mov_b32 v14, v10 @@ -17054,7 +17054,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: ; %bb.6: ; %Flow51 ; GFX11-NEXT: v_dual_mov_b32 v17, s2 :: v_dual_mov_b32 v10, v14 ; GFX11-NEXT: v_mov_b32_e32 v11, v15 -; GFX11-NEXT: .LBB13_7: ; %frem.loop_exit +; GFX11-NEXT: .LBB13_7: ; %frem.loop_exit24 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-NEXT: v_subrev_nc_u32_e32 v14, 25, v17 ; GFX11-NEXT: v_ldexp_f64 v[10:11], v[10:11], v14 @@ -17074,7 +17074,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: v_cmp_ngt_f64_e64 s2, |v[2:3]|, |v[6:7]| ; GFX11-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX11-NEXT: s_cbranch_vccz .LBB13_10 -; GFX11-NEXT: ; %bb.9: ; %frem.else16 +; GFX11-NEXT: ; %bb.9: ; %frem.else ; GFX11-NEXT: v_cmp_eq_f64_e64 vcc_lo, |v[2:3]|, |v[6:7]| ; GFX11-NEXT: v_and_b32_e32 v10, 0x80000000, v3 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) @@ -17084,7 +17084,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_branch .LBB13_16 ; GFX11-NEXT: .LBB13_10: ; GFX11-NEXT: ; implicit-def: $vgpr10_vgpr11 -; GFX11-NEXT: .LBB13_11: ; %frem.compute15 +; GFX11-NEXT: .LBB13_11: ; %frem.compute ; GFX11-NEXT: v_frexp_mant_f64_e64 v[10:11], |v[2:3]| ; GFX11-NEXT: v_frexp_exp_i32_f64_e32 v15, v[6:7] ; GFX11-NEXT: v_frexp_exp_i32_f64_e32 v14, v[2:3] @@ -17117,12 +17117,12 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) ; GFX11-NEXT: v_div_fixup_f64 v[14:15], v[14:15], v[10:11], 1.0 ; GFX11-NEXT: s_cbranch_vccnz .LBB13_15 -; GFX11-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX11-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX11-NEXT: s_sub_i32 s2, s2, s3 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_add_i32 s2, s2, 26 ; GFX11-NEXT: .p2align 6 -; GFX11-NEXT: .LBB13_13: ; %frem.loop_body23 +; GFX11-NEXT: .LBB13_13: ; %frem.loop_body ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX11-NEXT: v_dual_mov_b32 v17, v13 :: v_dual_mov_b32 v16, v12 @@ -17142,7 +17142,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX11-NEXT: ; %bb.14: ; %Flow ; GFX11-NEXT: v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v12, v16 ; GFX11-NEXT: v_mov_b32_e32 v13, v17 -; GFX11-NEXT: .LBB13_15: ; %frem.loop_exit24 +; GFX11-NEXT: .LBB13_15: ; %frem.loop_exit ; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-NEXT: v_subrev_nc_u32_e32 v16, 25, v19 ; GFX11-NEXT: v_ldexp_f64 v[12:13], v[12:13], v16 @@ -17187,7 +17187,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_ngt_f64_e64 s2, |v[0:1]|, |v[4:5]| ; GFX1150-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX1150-NEXT: s_cbranch_vccz .LBB13_2 -; GFX1150-NEXT: ; %bb.1: ; %frem.else +; GFX1150-NEXT: ; %bb.1: ; %frem.else16 ; GFX1150-NEXT: v_cmp_eq_f64_e64 vcc_lo, |v[0:1]|, |v[4:5]| ; GFX1150-NEXT: v_and_b32_e32 v8, 0x80000000, v1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) @@ -17197,7 +17197,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_branch .LBB13_8 ; GFX1150-NEXT: .LBB13_2: ; GFX1150-NEXT: ; implicit-def: $vgpr8_vgpr9 -; GFX1150-NEXT: .LBB13_3: ; %frem.compute +; GFX1150-NEXT: .LBB13_3: ; %frem.compute15 ; GFX1150-NEXT: v_frexp_mant_f64_e64 v[8:9], |v[0:1]| ; GFX1150-NEXT: v_frexp_exp_i32_f64_e32 v13, v[4:5] ; GFX1150-NEXT: v_frexp_exp_i32_f64_e32 v12, v[0:1] @@ -17229,12 +17229,12 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_gt_i32_e32 vcc_lo, 27, v17 ; GFX1150-NEXT: v_div_fixup_f64 v[12:13], v[12:13], v[8:9], 1.0 ; GFX1150-NEXT: s_cbranch_vccnz .LBB13_7 -; GFX1150-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1150-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX1150-NEXT: s_sub_i32 s2, s2, s3 ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_add_i32 s2, s2, 26 ; GFX1150-NEXT: .p2align 6 -; GFX1150-NEXT: .LBB13_5: ; %frem.loop_body +; GFX1150-NEXT: .LBB13_5: ; %frem.loop_body23 ; GFX1150-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-NEXT: v_dual_mov_b32 v15, v11 :: v_dual_mov_b32 v14, v10 @@ -17254,7 +17254,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: ; %bb.6: ; %Flow51 ; GFX1150-NEXT: v_dual_mov_b32 v17, s2 :: v_dual_mov_b32 v10, v14 ; GFX1150-NEXT: v_mov_b32_e32 v11, v15 -; GFX1150-NEXT: .LBB13_7: ; %frem.loop_exit +; GFX1150-NEXT: .LBB13_7: ; %frem.loop_exit24 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-NEXT: v_subrev_nc_u32_e32 v14, 25, v17 ; GFX1150-NEXT: v_ldexp_f64 v[10:11], v[10:11], v14 @@ -17274,7 +17274,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_ngt_f64_e64 s2, |v[2:3]|, |v[6:7]| ; GFX1150-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX1150-NEXT: s_cbranch_vccz .LBB13_10 -; GFX1150-NEXT: ; %bb.9: ; %frem.else16 +; GFX1150-NEXT: ; %bb.9: ; %frem.else ; GFX1150-NEXT: v_cmp_eq_f64_e64 vcc_lo, |v[2:3]|, |v[6:7]| ; GFX1150-NEXT: v_and_b32_e32 v10, 0x80000000, v3 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) @@ -17284,7 +17284,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: s_branch .LBB13_16 ; GFX1150-NEXT: .LBB13_10: ; GFX1150-NEXT: ; implicit-def: $vgpr10_vgpr11 -; GFX1150-NEXT: .LBB13_11: ; %frem.compute15 +; GFX1150-NEXT: .LBB13_11: ; %frem.compute ; GFX1150-NEXT: v_frexp_mant_f64_e64 v[10:11], |v[2:3]| ; GFX1150-NEXT: v_frexp_exp_i32_f64_e32 v15, v[6:7] ; GFX1150-NEXT: v_frexp_exp_i32_f64_e32 v14, v[2:3] @@ -17316,12 +17316,12 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: v_cmp_gt_i32_e32 vcc_lo, 27, v19 ; GFX1150-NEXT: v_div_fixup_f64 v[14:15], v[14:15], v[10:11], 1.0 ; GFX1150-NEXT: s_cbranch_vccnz .LBB13_15 -; GFX1150-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX1150-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX1150-NEXT: s_sub_i32 s2, s2, s3 ; GFX1150-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1150-NEXT: s_add_i32 s2, s2, 26 ; GFX1150-NEXT: .p2align 6 -; GFX1150-NEXT: .LBB13_13: ; %frem.loop_body23 +; GFX1150-NEXT: .LBB13_13: ; %frem.loop_body ; GFX1150-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1150-NEXT: v_dual_mov_b32 v17, v13 :: v_dual_mov_b32 v16, v12 @@ -17341,7 +17341,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1150-NEXT: ; %bb.14: ; %Flow ; GFX1150-NEXT: v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v12, v16 ; GFX1150-NEXT: v_mov_b32_e32 v13, v17 -; GFX1150-NEXT: .LBB13_15: ; %frem.loop_exit24 +; GFX1150-NEXT: .LBB13_15: ; %frem.loop_exit ; GFX1150-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1150-NEXT: v_subrev_nc_u32_e32 v16, 25, v19 ; GFX1150-NEXT: v_ldexp_f64 v[12:13], v[12:13], v16 @@ -17386,7 +17386,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_ngt_f64_e64 s2, |v[0:1]|, |v[4:5]| ; GFX1200-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX1200-NEXT: s_cbranch_vccz .LBB13_2 -; GFX1200-NEXT: ; %bb.1: ; %frem.else +; GFX1200-NEXT: ; %bb.1: ; %frem.else16 ; GFX1200-NEXT: v_cmp_eq_f64_e64 vcc_lo, |v[0:1]|, |v[4:5]| ; GFX1200-NEXT: v_and_b32_e32 v8, 0x80000000, v1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_3) @@ -17396,7 +17396,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_branch .LBB13_8 ; GFX1200-NEXT: .LBB13_2: ; GFX1200-NEXT: ; implicit-def: $vgpr8_vgpr9 -; GFX1200-NEXT: .LBB13_3: ; %frem.compute +; GFX1200-NEXT: .LBB13_3: ; %frem.compute15 ; GFX1200-NEXT: v_frexp_mant_f64_e64 v[8:9], |v[0:1]| ; GFX1200-NEXT: v_frexp_exp_i32_f64_e32 v13, v[4:5] ; GFX1200-NEXT: v_frexp_exp_i32_f64_e32 v12, v[0:1] @@ -17429,11 +17429,11 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_gt_i32_e32 vcc_lo, 27, v17 ; GFX1200-NEXT: v_div_fixup_f64 v[12:13], v[12:13], v[8:9], 1.0 ; GFX1200-NEXT: s_cbranch_vccnz .LBB13_7 -; GFX1200-NEXT: ; %bb.4: ; %frem.loop_body.preheader +; GFX1200-NEXT: ; %bb.4: ; %frem.loop_body23.preheader ; GFX1200-NEXT: s_sub_co_i32 s2, s2, s3 ; GFX1200-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1200-NEXT: s_add_co_i32 s2, s2, 26 -; GFX1200-NEXT: .LBB13_5: ; %frem.loop_body +; GFX1200-NEXT: .LBB13_5: ; %frem.loop_body23 ; GFX1200-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1200-NEXT: v_dual_mov_b32 v15, v11 :: v_dual_mov_b32 v14, v10 @@ -17454,7 +17454,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: ; %bb.6: ; %Flow51 ; GFX1200-NEXT: v_dual_mov_b32 v17, s2 :: v_dual_mov_b32 v10, v14 ; GFX1200-NEXT: v_mov_b32_e32 v11, v15 -; GFX1200-NEXT: .LBB13_7: ; %frem.loop_exit +; GFX1200-NEXT: .LBB13_7: ; %frem.loop_exit24 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-NEXT: v_subrev_nc_u32_e32 v14, 25, v17 ; GFX1200-NEXT: v_ldexp_f64 v[10:11], v[10:11], v14 @@ -17476,7 +17476,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_and_b32 vcc_lo, exec_lo, s2 ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_cbranch_vccz .LBB13_10 -; GFX1200-NEXT: ; %bb.9: ; %frem.else16 +; GFX1200-NEXT: ; %bb.9: ; %frem.else ; GFX1200-NEXT: v_cmp_eq_f64_e64 vcc_lo, |v[2:3]|, |v[6:7]| ; GFX1200-NEXT: v_and_b32_e32 v10, 0x80000000, v3 ; GFX1200-NEXT: s_wait_alu 0xfffd @@ -17487,7 +17487,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: s_branch .LBB13_16 ; GFX1200-NEXT: .LBB13_10: ; GFX1200-NEXT: ; implicit-def: $vgpr10_vgpr11 -; GFX1200-NEXT: .LBB13_11: ; %frem.compute15 +; GFX1200-NEXT: .LBB13_11: ; %frem.compute ; GFX1200-NEXT: v_frexp_mant_f64_e64 v[10:11], |v[2:3]| ; GFX1200-NEXT: v_frexp_exp_i32_f64_e32 v15, v[6:7] ; GFX1200-NEXT: v_frexp_exp_i32_f64_e32 v14, v[2:3] @@ -17520,11 +17520,11 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: v_cmp_gt_i32_e32 vcc_lo, 27, v19 ; GFX1200-NEXT: v_div_fixup_f64 v[14:15], v[14:15], v[10:11], 1.0 ; GFX1200-NEXT: s_cbranch_vccnz .LBB13_15 -; GFX1200-NEXT: ; %bb.12: ; %frem.loop_body23.preheader +; GFX1200-NEXT: ; %bb.12: ; %frem.loop_body.preheader ; GFX1200-NEXT: s_sub_co_i32 s2, s2, s3 ; GFX1200-NEXT: s_wait_alu 0xfffe ; GFX1200-NEXT: s_add_co_i32 s2, s2, 26 -; GFX1200-NEXT: .LBB13_13: ; %frem.loop_body23 +; GFX1200-NEXT: .LBB13_13: ; %frem.loop_body ; GFX1200-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_2) ; GFX1200-NEXT: v_dual_mov_b32 v17, v13 :: v_dual_mov_b32 v16, v12 @@ -17547,7 +17547,7 @@ define amdgpu_kernel void @frem_v2f64(ptr addrspace(1) %out, ptr addrspace(1) %i ; GFX1200-NEXT: ; %bb.14: ; %Flow ; GFX1200-NEXT: v_dual_mov_b32 v19, s2 :: v_dual_mov_b32 v12, v16 ; GFX1200-NEXT: v_mov_b32_e32 v13, v17 -; GFX1200-NEXT: .LBB13_15: ; %frem.loop_exit24 +; GFX1200-NEXT: .LBB13_15: ; %frem.loop_exit ; GFX1200-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX1200-NEXT: v_subrev_nc_u32_e32 v16, 25, v19 ; GFX1200-NEXT: v_ldexp_f64 v[12:13], v[12:13], v16 diff --git a/llvm/test/CodeGen/AMDGPU/function-args.ll b/llvm/test/CodeGen/AMDGPU/function-args.ll index 3c41cc43a089..5babe9fb3d85 100644 --- a/llvm/test/CodeGen/AMDGPU/function-args.ll +++ b/llvm/test/CodeGen/AMDGPU/function-args.ll @@ -1111,15 +1111,11 @@ define void @void_func_v4i8(<4 x i8> %arg0) #0 { ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: buffer_store_b32 v0, off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -1190,18 +1186,15 @@ define void @void_func_v5i8(<5 x i8> %arg0) #0 { ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 4 -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: buffer_store_b8 v4, off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v3, v2 ; GFX11-TRUE16-NEXT: buffer_store_b32 v0, off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; @@ -1281,28 +1274,22 @@ define void @void_func_v8i8(<8 x i8> %arg0) #0 { ; GFX11-TRUE16-LABEL: void_func_v8i8: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v5.h, v4.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v6.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v6.l, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v2.l, v1.h ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v4, v6 -; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v6.l ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v0, v6 -; GFX11-TRUE16-NEXT: buffer_store_b64 v[1:2], off, s[0:3], 0 +; GFX11-TRUE16-NEXT: buffer_store_b64 v[2:3], off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: void_func_v8i8: @@ -1416,44 +1403,34 @@ define void @void_func_v16i8(<16 x i8> %arg0) #0 { ; GFX11-TRUE16-LABEL: void_func_v16i8: ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v15.l -; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v13.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.l, 8, v15.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v13.l ; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, 0 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v13.h, v12.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.l, 8, v11.l ; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v12.l, v13.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v14.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.l, 8, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v10.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v12, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v10.l, v9.h -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v10.h, v9.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v14.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v8.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v5.h, v4.h -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v4.l, v5.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v14.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v10.h, v15.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v12.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v10.l, v13.l +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v6.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v6.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v0.h, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v2.l, v1.h +; GFX11-TRUE16-NEXT: v_or_b16 v8.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v4, v14 -; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v14.l ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v0, v14 -; GFX11-TRUE16-NEXT: buffer_store_b128 v[5:8], off, s[0:3], 0 +; GFX11-TRUE16-NEXT: buffer_store_b128 v[8:11], off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-FAKE16-LABEL: void_func_v16i8: @@ -1649,78 +1626,59 @@ define void @void_func_v32i8(<32 x i8> %arg0) #0 { ; GFX11-TRUE16: ; %bb.0: ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-TRUE16-NEXT: scratch_load_d16_u8 v31, off, s32 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, 0 ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v15.l ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v13.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.h, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v10.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v9.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v32.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v12.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v11.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v10.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.h, 8, v9.l +; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v8.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v7.l ; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v7.h, v6.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v12, v32 -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v5.h, v4.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.h, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v11.h, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v13.h, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v15.h, 8, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.h, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v9.h, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v5.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v9, v32 -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v6.l, v7.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.h, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v10, v32 -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v0.h, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v6.h, v5.h -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v11, v32 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.h, v32.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v23.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.h, 8, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v20.l -; GFX11-TRUE16-NEXT: v_or_b16 v14.l, v7.h, v6.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.h, v32.l -; GFX11-TRUE16-NEXT: v_and_b16 v8.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v4.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v6.h, v5.h +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v8.l, v7.h +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v6.l, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v4.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v7.l, v11.l, v10.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v12.l, v11.h +; GFX11-TRUE16-NEXT: v_or_b16 v6.l, v13.l, v12.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v14.l, v13.h +; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v15.l, v14.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v16.h, v15.h +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v16.l, v17.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v9.h ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 16 -; GFX11-TRUE16-NEXT: v_or_b16 v15.l, v6.h, v5.h ; GFX11-TRUE16-NEXT: s_mov_b32 s3, 0x31016000 ; GFX11-TRUE16-NEXT: s_mov_b32 s2, -1 ; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v31.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.h, v7.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v13, v32 -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v5.l, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v19.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v17.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v14, v32 -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.h, v8.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v8.h, v5.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v15, v32 -; GFX11-TRUE16-NEXT: v_or_b16 v32.h, v4.h, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v31.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v9, v32 +; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v10.l, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v9.l, v8.h ; GFX11-TRUE16-NEXT: buffer_store_b128 v[4:7], off, s[0:3], 0 ; GFX11-TRUE16-NEXT: s_mov_b64 s[0:1], 0 ; GFX11-TRUE16-NEXT: buffer_store_b128 v[0:3], off, s[0:3], 0 diff --git a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll index f67ab18dd8ef..234eaa8af7ed 100644 --- a/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll +++ b/llvm/test/CodeGen/AMDGPU/gfx-callable-argument-types.ll @@ -4985,21 +4985,17 @@ define amdgpu_gfx void @test_call_external_void_func_v4i8_ret() #0 { ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h ; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v2.l ; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1 ; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v2 ; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33 -; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2 ; GFX11-TRUE16-NEXT: global_store_b32 v[40:41], v0, off ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33 ; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s33 offset:4 +; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2 ; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s1, -1 ; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s33 offset:8 ; 4-byte Folded Reload ; GFX11-TRUE16-NEXT: s_mov_b32 exec_lo, s1 @@ -5243,18 +5239,14 @@ define amdgpu_gfx void @test_call_external_void_func_v5i8_ret() #0 { ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, v2.l -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, 4 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, 0 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v3, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, 4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, 0 +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l ; GFX11-TRUE16-NEXT: s_clause 0x1 -; GFX11-TRUE16-NEXT: global_store_b8 v[0:1], v4, off -; GFX11-TRUE16-NEXT: global_store_b32 v[40:41], v2, off +; GFX11-TRUE16-NEXT: global_store_b8 v[2:3], v4, off +; GFX11-TRUE16-NEXT: global_store_b32 v[40:41], v0, off ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33 ; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s33 offset:4 @@ -5528,27 +5520,21 @@ define amdgpu_gfx void @test_call_external_void_func_v8i8_ret() #0 { ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v5.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, 0 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v5.l, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, v4.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l ; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v5, v4 ; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v4.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v0, v4 -; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1 -; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0 -; GFX11-TRUE16-NEXT: global_store_b64 v[40:41], v[1:2], off +; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v2.l, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v0.l, v1.l +; GFX11-TRUE16-NEXT: global_store_b64 v[40:41], v[3:4], off ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: scratch_load_b32 v41, off, s33 ; GFX11-TRUE16-NEXT: scratch_load_b32 v40, off, s33 offset:4 +; GFX11-TRUE16-NEXT: v_readlane_b32 s31, v42, 1 +; GFX11-TRUE16-NEXT: v_readlane_b32 s30, v42, 0 ; GFX11-TRUE16-NEXT: s_mov_b32 s32, s33 ; GFX11-TRUE16-NEXT: v_readlane_b32 s0, v42, 2 ; GFX11-TRUE16-NEXT: s_or_saveexec_b32 s1, -1 @@ -5994,73 +5980,53 @@ define amdgpu_gfx void @test_call_external_void_func_v32i8_ret() #0 { ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v14.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v13.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v12.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, 0 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v7.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v6.l ; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v11.l -; GFX11-TRUE16-NEXT: v_or_b16 v13.l, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.h, v12.l +; GFX11-TRUE16-NEXT: v_or_b16 v12.l, v3.h, v2.h ; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v10.l ; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v9.l ; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v8.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v13, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v7.l -; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v6.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v5.l -; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v9, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v3.l -; GFX11-TRUE16-NEXT: v_or_b16 v4.l, v3.h, v2.h -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v2.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.l, 0xff, v0.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, v12.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v4, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v1.h, v0.h -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.l, v1.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v31.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v30.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v29.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v28.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v2, v12 -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v17.l -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v0.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v27.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v1.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v26.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v25.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v24.l -; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v16.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v2, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v0.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v23.l -; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v1.h, v1.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v22.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v21.l -; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v20.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) -; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v2, v12 -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v0.h, v0.l -; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v19.l -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v1.h, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, v12.l -; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v18.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v1, v12 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v0.h, v0.l -; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v4.h, v4.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.h, v12.l -; GFX11-TRUE16-NEXT: v_or_b32_e32 v0, v0, v12 +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v5.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v3.l, 8, v3.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v11.l, v3.h, v2.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v5.h, v4.h +; GFX11-TRUE16-NEXT: v_or_b16 v10.l, v4.l, v5.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v2.l, v3.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.h, 8, v31.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v30.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v29.l +; GFX11-TRUE16-NEXT: v_and_b16 v2.h, 0xff, v28.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v27.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v26.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v5.l, 8, v1.l +; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v1.h, v0.h +; GFX11-TRUE16-NEXT: v_or_b16 v3.l, v2.h, v2.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v4.h, v4.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v0.l, 8, v25.l +; GFX11-TRUE16-NEXT: v_and_b16 v0.h, 0xff, v24.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v1.l, 8, v23.l +; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v22.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v21.l +; GFX11-TRUE16-NEXT: v_and_b16 v4.h, 0xff, v20.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v19.l +; GFX11-TRUE16-NEXT: v_and_b16 v6.h, 0xff, v18.l +; GFX11-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v17.l +; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v16.l +; GFX11-TRUE16-NEXT: v_or_b16 v2.l, v0.h, v0.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v1.l +; GFX11-TRUE16-NEXT: v_or_b16 v1.l, v4.h, v4.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.h, v6.h, v6.l +; GFX11-TRUE16-NEXT: v_or_b16 v0.l, v7.h, v7.l +; GFX11-TRUE16-NEXT: v_or_b16 v9.l, v5.h, v5.l ; GFX11-TRUE16-NEXT: s_clause 0x1 ; GFX11-TRUE16-NEXT: global_store_b128 v[42:43], v[0:3], off -; GFX11-TRUE16-NEXT: global_store_b128 v[40:41], v[5:8], off +; GFX11-TRUE16-NEXT: global_store_b128 v[40:41], v[9:12], off ; GFX11-TRUE16-NEXT: s_clause 0x3 ; GFX11-TRUE16-NEXT: scratch_load_b32 v43, off, s33 ; GFX11-TRUE16-NEXT: scratch_load_b32 v42, off, s33 offset:4 diff --git a/llvm/test/CodeGen/AMDGPU/global-constant.ll b/llvm/test/CodeGen/AMDGPU/global-constant.ll index 866d3a1e3089..b04602aff8e6 100644 --- a/llvm/test/CodeGen/AMDGPU/global-constant.ll +++ b/llvm/test/CodeGen/AMDGPU/global-constant.ll @@ -12,21 +12,21 @@ ; Non-R600 OSes use relocations. ; GCN-DEFAULT: s_getpc_b64 s[[[PC0_LO:[0-9]+]]:[[PC0_HI:[0-9]+]]] -; GCN-DEFAULT: s_add_u32 s{{[0-9]+}}, s[[PC0_LO]], private1@rel32@lo+4 -; GCN-DEFAULT: s_addc_u32 s{{[0-9]+}}, s[[PC0_HI]], private1@rel32@hi+12 +; GCN-DEFAULT: s_add_u32 s{{[0-9]+}}, s[[PC0_LO]], .Lprivate1@rel32@lo+4 +; GCN-DEFAULT: s_addc_u32 s{{[0-9]+}}, s[[PC0_HI]], .Lprivate1@rel32@hi+12 ; GCN-DEFAULT: s_getpc_b64 s[[[PC1_LO:[0-9]+]]:[[PC1_HI:[0-9]+]]] -; GCN-DEFAULT: s_add_u32 s{{[0-9]+}}, s[[PC1_LO]], private2@rel32@lo+4 -; GCN-DEFAULT: s_addc_u32 s{{[0-9]+}}, s[[PC1_HI]], private2@rel32@hi+12 +; GCN-DEFAULT: s_add_u32 s{{[0-9]+}}, s[[PC1_LO]], .Lprivate2@rel32@lo+4 +; GCN-DEFAULT: s_addc_u32 s{{[0-9]+}}, s[[PC1_HI]], .Lprivate2@rel32@hi+12 ; MESA uses absolute relocations. -; GCN-MESA: s_add_u32 s2, private1@abs32@lo, s4 -; GCN-MESA: s_addc_u32 s3, private1@abs32@hi, s5 +; GCN-MESA: s_add_u32 s2, .Lprivate1@abs32@lo, s4 +; GCN-MESA: s_addc_u32 s3, .Lprivate1@abs32@hi, s5 ; PAL uses absolute relocations. -; GCN-PAL: s_add_u32 s2, private1@abs32@lo, s4 -; GCN-PAL: s_addc_u32 s3, private1@abs32@hi, s5 -; GCN-PAL: s_add_u32 s4, private2@abs32@lo, s4 -; GCN-PAL: s_addc_u32 s5, private2@abs32@hi, s5 +; GCN-PAL: s_add_u32 s2, .Lprivate1@abs32@lo, s4 +; GCN-PAL: s_addc_u32 s3, .Lprivate1@abs32@hi, s5 +; GCN-PAL: s_add_u32 s4, .Lprivate2@abs32@lo, s4 +; GCN-PAL: s_addc_u32 s5, .Lprivate2@abs32@hi, s5 ; R600-LABEL: private_test define amdgpu_kernel void @private_test(i32 %index, ptr addrspace(1) %out) { diff --git a/llvm/test/CodeGen/AMDGPU/global-variable-relocs.ll b/llvm/test/CodeGen/AMDGPU/global-variable-relocs.ll index b8cfcbf2d266..6d55e79edbef 100644 --- a/llvm/test/CodeGen/AMDGPU/global-variable-relocs.ll +++ b/llvm/test/CodeGen/AMDGPU/global-variable-relocs.ll @@ -14,8 +14,8 @@ ; CHECK-LABEL: private_test: ; CHECK: s_getpc_b64 s[[[PC_LO:[0-9]+]]:[[PC_HI:[0-9]+]]] -; CHECK: s_add_u32 s[[ADDR_LO:[0-9]+]], s[[PC_LO]], private@rel32@lo+8 -; CHECK: s_addc_u32 s[[ADDR_HI:[0-9]+]], s[[PC_HI]], private@rel32@hi+16 +; CHECK: s_add_u32 s[[ADDR_LO:[0-9]+]], s[[PC_LO]], .Lprivate@rel32@lo+8 +; CHECK: s_addc_u32 s[[ADDR_HI:[0-9]+]], s[[PC_HI]], .Lprivate@rel32@hi+16 ; CHECK: s_load_dword s{{[0-9]+}}, s[[[ADDR_LO]]:[[ADDR_HI]]] define amdgpu_kernel void @private_test(ptr addrspace(1) %out) { %ptr = getelementptr [256 x i32], ptr addrspace(1) @private, i32 0, i32 1 @@ -153,7 +153,7 @@ define amdgpu_kernel void @external_w_init_test(ptr addrspace(1) %out) { ret void } -; CHECK: .local private +; CHECK: .local .Lprivate ; CHECK: .local internal ; CHECK: .weak linkonce ; CHECK: .weak weak diff --git a/llvm/test/CodeGen/AMDGPU/idot4u.ll b/llvm/test/CodeGen/AMDGPU/idot4u.ll index 049663a1e1bb..f80d50b56f55 100644 --- a/llvm/test/CodeGen/AMDGPU/idot4u.ll +++ b/llvm/test/CodeGen/AMDGPU/idot4u.ll @@ -2730,18 +2730,15 @@ define amdgpu_kernel void @udot4_acc8_vecMul(ptr addrspace(1) %src1, ; GFX11-DL-TRUE16-NEXT: v_lshrrev_b16 v1.h, 8, v4.l ; GFX11-DL-TRUE16-NEXT: v_mad_u16 v0.l, v3.l, v4.l, v0.l ; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v2.l, v2.l, v6.l -; GFX11-DL-TRUE16-NEXT: v_mov_b16_e32 v6.l, 0 +; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-DL-TRUE16-NEXT: v_and_b16 v1.l, 0xff, v1.l ; GFX11-DL-TRUE16-NEXT: v_mul_lo_u16 v0.h, v0.h, v1.h -; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) +; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v2.l -; GFX11-DL-TRUE16-NEXT: v_mov_b16_e32 v7.h, v6.l -; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) -; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v7.l, 8, v0.h +; GFX11-DL-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v0.h +; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-DL-TRUE16-NEXT: v_or_b16 v6.h, v1.l, v2.l -; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-DL-TRUE16-NEXT: v_or_b32_e32 v1, v7, v6 -; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v1 +; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v6 ; GFX11-DL-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX11-DL-TRUE16-NEXT: v_add_nc_u16 v0.l, v0.l, v1.l ; GFX11-DL-TRUE16-NEXT: v_lshrrev_b32_e32 v1, 8, v2 diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.is.private.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.is.private.ll index 67d0410434a9..3324018ca723 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.is.private.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.is.private.ll @@ -3,11 +3,11 @@ ; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s | FileCheck -check-prefixes=CI,CI-SDAG %s ; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX9-SDAG %s ; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250,GFX1250-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s | FileCheck -check-prefixes=CI,CI-GISEL %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250,GFX1250-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s | FileCheck -check-prefixes=CI,CI-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250,GFX1250-GISEL %s define amdgpu_kernel void @is_private_vgpr(ptr addrspace(1) %ptr.ptr) { ; SI-LABEL: is_private_vgpr: diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.is.shared.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.is.shared.ll index 63333ed165a3..355d0026091d 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.is.shared.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.is.shared.ll @@ -3,11 +3,11 @@ ; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s | FileCheck -check-prefixes=CI,CI-SDAG %s ; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX9-SDAG %s ; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250,GFX1250-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s | FileCheck -check-prefixes=CI,CI-GISEL %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL %s -; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250,GFX1250-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii < %s | FileCheck -check-prefixes=CI,CI-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1250 < %s | FileCheck -check-prefixes=GFX1250,GFX1250-GISEL %s define amdgpu_kernel void @is_local_vgpr(ptr addrspace(1) %ptr.ptr) { ; CIT-LABEL: is_local_vgpr: diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.smfmac.gfx950.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.smfmac.gfx950.ll index ee11b9295a24..0c1448a0b8fb 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.smfmac.gfx950.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.smfmac.gfx950.ll @@ -44,23 +44,23 @@ define amdgpu_kernel void @test_smfmac_f32_16x16x64_f16__vgpr(ptr addrspace(1) % ; GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0 ; GISEL-NEXT: v_lshlrev_b32_e32 v0, 4, v0 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: global_load_dwordx4 v[14:17], v0, s[6:7] +; GISEL-NEXT: global_load_dwordx4 v[8:11], v0, s[6:7] ; GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x44 ; GISEL-NEXT: s_load_dword s16, s[4:5], 0x64 -; GISEL-NEXT: v_mov_b64_e32 v[10:11], s[2:3] -; GISEL-NEXT: v_mov_b64_e32 v[8:9], s[0:1] +; GISEL-NEXT: v_mov_b64_e32 v[14:15], s[2:3] +; GISEL-NEXT: v_mov_b64_e32 v[12:13], s[0:1] ; GISEL-NEXT: s_waitcnt lgkmcnt(0) ; GISEL-NEXT: v_mov_b64_e32 v[0:1], s[8:9] ; GISEL-NEXT: v_mov_b64_e32 v[2:3], s[10:11] ; GISEL-NEXT: v_mov_b64_e32 v[4:5], s[12:13] ; GISEL-NEXT: v_mov_b64_e32 v[6:7], s[14:15] -; GISEL-NEXT: v_mov_b32_e32 v12, s16 +; GISEL-NEXT: v_mov_b32_e32 v16, s16 ; GISEL-NEXT: s_waitcnt vmcnt(0) ; GISEL-NEXT: s_nop 0 -; GISEL-NEXT: v_smfmac_f32_16x16x64_f16 v[14:17], v[8:11], v[0:7], v12 cbsz:1 abid:2 +; GISEL-NEXT: v_smfmac_f32_16x16x64_f16 v[8:11], v[12:15], v[0:7], v16 cbsz:1 abid:2 ; GISEL-NEXT: v_mov_b32_e32 v0, 0 ; GISEL-NEXT: s_nop 6 -; GISEL-NEXT: global_store_dwordx4 v0, v[14:17], s[6:7] +; GISEL-NEXT: global_store_dwordx4 v0, v[8:11], s[6:7] ; GISEL-NEXT: s_endpgm bb: %id = call i32 @llvm.amdgcn.workitem.id.x() @@ -834,24 +834,24 @@ define amdgpu_kernel void @test_smfmac_i32_16x16x128_i8__vgpr(ptr addrspace(1) % ; GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0 ; GISEL-NEXT: v_lshlrev_b32_e32 v0, 4, v0 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: global_load_dwordx4 v[14:17], v0, s[0:1] +; GISEL-NEXT: global_load_dwordx4 v[8:11], v0, s[0:1] ; GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x34 ; GISEL-NEXT: s_load_dwordx4 s[16:19], s[4:5], 0x54 ; GISEL-NEXT: s_load_dword s2, s[4:5], 0x64 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: v_mov_b64_e32 v[8:9], s[8:9] -; GISEL-NEXT: v_mov_b64_e32 v[10:11], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[14:15], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[12:13], s[8:9] ; GISEL-NEXT: v_mov_b64_e32 v[0:1], s[12:13] ; GISEL-NEXT: v_mov_b64_e32 v[2:3], s[14:15] ; GISEL-NEXT: v_mov_b64_e32 v[4:5], s[16:17] ; GISEL-NEXT: v_mov_b64_e32 v[6:7], s[18:19] -; GISEL-NEXT: v_mov_b32_e32 v12, s2 +; GISEL-NEXT: v_mov_b32_e32 v16, s2 ; GISEL-NEXT: s_waitcnt vmcnt(0) ; GISEL-NEXT: s_nop 0 -; GISEL-NEXT: v_smfmac_i32_16x16x128_i8 v[14:17], v[8:11], v[0:7], v12 cbsz:1 abid:2 +; GISEL-NEXT: v_smfmac_i32_16x16x128_i8 v[8:11], v[12:15], v[0:7], v16 cbsz:1 abid:2 ; GISEL-NEXT: v_mov_b32_e32 v0, 0 ; GISEL-NEXT: s_nop 6 -; GISEL-NEXT: global_store_dwordx4 v0, v[14:17], s[0:1] +; GISEL-NEXT: global_store_dwordx4 v0, v[8:11], s[0:1] ; GISEL-NEXT: s_endpgm bb: %id = call i32 @llvm.amdgcn.workitem.id.x() @@ -1349,24 +1349,24 @@ define amdgpu_kernel void @test_smfmac_f32_16x16x128_bf8_bf8__vgpr(ptr addrspace ; GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0 ; GISEL-NEXT: v_lshlrev_b32_e32 v0, 4, v0 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: global_load_dwordx4 v[14:17], v0, s[0:1] +; GISEL-NEXT: global_load_dwordx4 v[8:11], v0, s[0:1] ; GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x34 ; GISEL-NEXT: s_load_dwordx4 s[16:19], s[4:5], 0x54 ; GISEL-NEXT: s_load_dword s2, s[4:5], 0x64 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: v_mov_b64_e32 v[8:9], s[8:9] -; GISEL-NEXT: v_mov_b64_e32 v[10:11], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[14:15], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[12:13], s[8:9] ; GISEL-NEXT: v_mov_b64_e32 v[0:1], s[12:13] ; GISEL-NEXT: v_mov_b64_e32 v[2:3], s[14:15] ; GISEL-NEXT: v_mov_b64_e32 v[4:5], s[16:17] ; GISEL-NEXT: v_mov_b64_e32 v[6:7], s[18:19] -; GISEL-NEXT: v_mov_b32_e32 v12, s2 +; GISEL-NEXT: v_mov_b32_e32 v16, s2 ; GISEL-NEXT: s_waitcnt vmcnt(0) ; GISEL-NEXT: s_nop 0 -; GISEL-NEXT: v_smfmac_f32_16x16x128_bf8_bf8 v[14:17], v[8:11], v[0:7], v12 cbsz:1 abid:2 +; GISEL-NEXT: v_smfmac_f32_16x16x128_bf8_bf8 v[8:11], v[12:15], v[0:7], v16 cbsz:1 abid:2 ; GISEL-NEXT: v_mov_b32_e32 v0, 0 ; GISEL-NEXT: s_nop 6 -; GISEL-NEXT: global_store_dwordx4 v0, v[14:17], s[0:1] +; GISEL-NEXT: global_store_dwordx4 v0, v[8:11], s[0:1] ; GISEL-NEXT: s_endpgm bb: %id = call i32 @llvm.amdgcn.workitem.id.x() @@ -1513,24 +1513,24 @@ define amdgpu_kernel void @test_smfmac_f32_16x16x128_bf8_fp8__vgpr(ptr addrspace ; GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0 ; GISEL-NEXT: v_lshlrev_b32_e32 v0, 4, v0 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: global_load_dwordx4 v[14:17], v0, s[0:1] +; GISEL-NEXT: global_load_dwordx4 v[8:11], v0, s[0:1] ; GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x34 ; GISEL-NEXT: s_load_dwordx4 s[16:19], s[4:5], 0x54 ; GISEL-NEXT: s_load_dword s2, s[4:5], 0x64 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: v_mov_b64_e32 v[8:9], s[8:9] -; GISEL-NEXT: v_mov_b64_e32 v[10:11], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[14:15], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[12:13], s[8:9] ; GISEL-NEXT: v_mov_b64_e32 v[0:1], s[12:13] ; GISEL-NEXT: v_mov_b64_e32 v[2:3], s[14:15] ; GISEL-NEXT: v_mov_b64_e32 v[4:5], s[16:17] ; GISEL-NEXT: v_mov_b64_e32 v[6:7], s[18:19] -; GISEL-NEXT: v_mov_b32_e32 v12, s2 +; GISEL-NEXT: v_mov_b32_e32 v16, s2 ; GISEL-NEXT: s_waitcnt vmcnt(0) ; GISEL-NEXT: s_nop 0 -; GISEL-NEXT: v_smfmac_f32_16x16x128_bf8_fp8 v[14:17], v[8:11], v[0:7], v12 cbsz:1 abid:2 +; GISEL-NEXT: v_smfmac_f32_16x16x128_bf8_fp8 v[8:11], v[12:15], v[0:7], v16 cbsz:1 abid:2 ; GISEL-NEXT: v_mov_b32_e32 v0, 0 ; GISEL-NEXT: s_nop 6 -; GISEL-NEXT: global_store_dwordx4 v0, v[14:17], s[0:1] +; GISEL-NEXT: global_store_dwordx4 v0, v[8:11], s[0:1] ; GISEL-NEXT: s_endpgm bb: %id = call i32 @llvm.amdgcn.workitem.id.x() @@ -1677,24 +1677,24 @@ define amdgpu_kernel void @test_smfmac_f32_16x16x128_fp8_bf8__vgpr(ptr addrspace ; GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0 ; GISEL-NEXT: v_lshlrev_b32_e32 v0, 4, v0 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: global_load_dwordx4 v[14:17], v0, s[0:1] +; GISEL-NEXT: global_load_dwordx4 v[8:11], v0, s[0:1] ; GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x34 ; GISEL-NEXT: s_load_dwordx4 s[16:19], s[4:5], 0x54 ; GISEL-NEXT: s_load_dword s2, s[4:5], 0x64 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: v_mov_b64_e32 v[8:9], s[8:9] -; GISEL-NEXT: v_mov_b64_e32 v[10:11], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[14:15], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[12:13], s[8:9] ; GISEL-NEXT: v_mov_b64_e32 v[0:1], s[12:13] ; GISEL-NEXT: v_mov_b64_e32 v[2:3], s[14:15] ; GISEL-NEXT: v_mov_b64_e32 v[4:5], s[16:17] ; GISEL-NEXT: v_mov_b64_e32 v[6:7], s[18:19] -; GISEL-NEXT: v_mov_b32_e32 v12, s2 +; GISEL-NEXT: v_mov_b32_e32 v16, s2 ; GISEL-NEXT: s_waitcnt vmcnt(0) ; GISEL-NEXT: s_nop 0 -; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 v[14:17], v[8:11], v[0:7], v12 cbsz:1 abid:2 +; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_bf8 v[8:11], v[12:15], v[0:7], v16 cbsz:1 abid:2 ; GISEL-NEXT: v_mov_b32_e32 v0, 0 ; GISEL-NEXT: s_nop 6 -; GISEL-NEXT: global_store_dwordx4 v0, v[14:17], s[0:1] +; GISEL-NEXT: global_store_dwordx4 v0, v[8:11], s[0:1] ; GISEL-NEXT: s_endpgm bb: %id = call i32 @llvm.amdgcn.workitem.id.x() @@ -1841,24 +1841,24 @@ define amdgpu_kernel void @test_smfmac_f32_16x16x128_fp8_fp8__vgpr(ptr addrspace ; GISEL-NEXT: v_and_b32_e32 v0, 0x3ff, v0 ; GISEL-NEXT: v_lshlrev_b32_e32 v0, 4, v0 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: global_load_dwordx4 v[14:17], v0, s[0:1] +; GISEL-NEXT: global_load_dwordx4 v[8:11], v0, s[0:1] ; GISEL-NEXT: s_load_dwordx8 s[8:15], s[4:5], 0x34 ; GISEL-NEXT: s_load_dwordx4 s[16:19], s[4:5], 0x54 ; GISEL-NEXT: s_load_dword s2, s[4:5], 0x64 ; GISEL-NEXT: s_waitcnt lgkmcnt(0) -; GISEL-NEXT: v_mov_b64_e32 v[8:9], s[8:9] -; GISEL-NEXT: v_mov_b64_e32 v[10:11], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[14:15], s[10:11] +; GISEL-NEXT: v_mov_b64_e32 v[12:13], s[8:9] ; GISEL-NEXT: v_mov_b64_e32 v[0:1], s[12:13] ; GISEL-NEXT: v_mov_b64_e32 v[2:3], s[14:15] ; GISEL-NEXT: v_mov_b64_e32 v[4:5], s[16:17] ; GISEL-NEXT: v_mov_b64_e32 v[6:7], s[18:19] -; GISEL-NEXT: v_mov_b32_e32 v12, s2 +; GISEL-NEXT: v_mov_b32_e32 v16, s2 ; GISEL-NEXT: s_waitcnt vmcnt(0) ; GISEL-NEXT: s_nop 0 -; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_fp8 v[14:17], v[8:11], v[0:7], v12 cbsz:1 abid:2 +; GISEL-NEXT: v_smfmac_f32_16x16x128_fp8_fp8 v[8:11], v[12:15], v[0:7], v16 cbsz:1 abid:2 ; GISEL-NEXT: v_mov_b32_e32 v0, 0 ; GISEL-NEXT: s_nop 6 -; GISEL-NEXT: global_store_dwordx4 v0, v[14:17], s[0:1] +; GISEL-NEXT: global_store_dwordx4 v0, v[8:11], s[0:1] ; GISEL-NEXT: s_endpgm bb: %id = call i32 @llvm.amdgcn.workitem.id.x() diff --git a/llvm/test/CodeGen/AMDGPU/llvm.exp.ll b/llvm/test/CodeGen/AMDGPU/llvm.exp.ll index af79c911f29f..ac356fad5b2d 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.exp.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.exp.ll @@ -6011,8 +6011,7 @@ define half @v_exp_f16_fast(half %in) { ; SI-GISEL: ; %bb.0: ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, 0x3dc5 -; SI-GISEL-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-GISEL-NEXT: v_mul_f32_e32 v0, 0x3fb8a000, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 ; SI-GISEL-NEXT: v_exp_f32_e32 v0, v0 @@ -6512,10 +6511,9 @@ define <2 x half> @v_exp_v2f16_fast(<2 x half> %in) { ; SI-GISEL: ; %bb.0: ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, 0x3dc5 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, v1 -; SI-GISEL-NEXT: v_mul_f32_e32 v0, v0, v2 -; SI-GISEL-NEXT: v_mul_f32_e32 v1, v1, v2 +; SI-GISEL-NEXT: v_mul_f32_e32 v0, 0x3fb8a000, v0 +; SI-GISEL-NEXT: v_mul_f32_e32 v1, 0x3fb8a000, v1 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v1, v1 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 @@ -6709,12 +6707,11 @@ define <3 x half> @v_exp_v3f16_afn(<3 x half> %in) { ; SI-GISEL: ; %bb.0: ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v3, 0x3dc5 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, v1 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v2 -; SI-GISEL-NEXT: v_mul_f32_e32 v0, v0, v3 -; SI-GISEL-NEXT: v_mul_f32_e32 v1, v1, v3 -; SI-GISEL-NEXT: v_mul_f32_e32 v2, v2, v3 +; SI-GISEL-NEXT: v_mul_f32_e32 v0, 0x3fb8a000, v0 +; SI-GISEL-NEXT: v_mul_f32_e32 v1, 0x3fb8a000, v1 +; SI-GISEL-NEXT: v_mul_f32_e32 v2, 0x3fb8a000, v2 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v1, v1 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v2, v2 diff --git a/llvm/test/CodeGen/AMDGPU/llvm.exp10.ll b/llvm/test/CodeGen/AMDGPU/llvm.exp10.ll index a99c1991a790..d12ebe49814d 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.exp10.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.exp10.ll @@ -6092,8 +6092,7 @@ define half @v_exp10_f16_fast(half %in) { ; SI-GISEL: ; %bb.0: ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, 0x3dc5 -; SI-GISEL-NEXT: v_mul_f32_e32 v0, v0, v1 +; SI-GISEL-NEXT: v_mul_f32_e32 v0, 0x3fb8a000, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 ; SI-GISEL-NEXT: v_exp_f32_e32 v0, v0 @@ -6594,10 +6593,9 @@ define <2 x half> @v_exp10_v2f16_fast(<2 x half> %in) { ; SI-GISEL: ; %bb.0: ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, 0x3dc5 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, v1 -; SI-GISEL-NEXT: v_mul_f32_e32 v0, v0, v2 -; SI-GISEL-NEXT: v_mul_f32_e32 v1, v1, v2 +; SI-GISEL-NEXT: v_mul_f32_e32 v0, 0x3fb8a000, v0 +; SI-GISEL-NEXT: v_mul_f32_e32 v1, 0x3fb8a000, v1 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v1, v1 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 @@ -6791,12 +6789,11 @@ define <3 x half> @v_exp10_v3f16_afn(<3 x half> %in) { ; SI-GISEL: ; %bb.0: ; SI-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v3, 0x3dc5 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v1, v1 ; SI-GISEL-NEXT: v_cvt_f32_f16_e32 v2, v2 -; SI-GISEL-NEXT: v_mul_f32_e32 v0, v0, v3 -; SI-GISEL-NEXT: v_mul_f32_e32 v1, v1, v3 -; SI-GISEL-NEXT: v_mul_f32_e32 v2, v2, v3 +; SI-GISEL-NEXT: v_mul_f32_e32 v0, 0x3fb8a000, v0 +; SI-GISEL-NEXT: v_mul_f32_e32 v1, 0x3fb8a000, v1 +; SI-GISEL-NEXT: v_mul_f32_e32 v2, 0x3fb8a000, v2 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v1, v1 ; SI-GISEL-NEXT: v_cvt_f16_f32_e32 v2, v2 diff --git a/llvm/test/CodeGen/AMDGPU/llvm.memcpy.ll b/llvm/test/CodeGen/AMDGPU/llvm.memcpy.ll index 63e9eef3297a..66b795876d70 100644 --- a/llvm/test/CodeGen/AMDGPU/llvm.memcpy.ll +++ b/llvm/test/CodeGen/AMDGPU/llvm.memcpy.ll @@ -315,7 +315,7 @@ define amdgpu_kernel void @test_small_memcpy_i64_global_to_global_align16(ptr ad ; FUNC-LABEL: {{^}}test_memcpy_const_string_align4: ; SI: s_getpc_b64 -; SI: s_add_u32 s{{[0-9]+}}, s{{[0-9]+}}, hello.align4@rel32@lo+4 +; SI: s_add_u32 s{{[0-9]+}}, s{{[0-9]+}}, .Lhello.align4@rel32@lo+4 ; SI: s_addc_u32 ; SI-DAG: s_load_dwordx8 ; SI-DAG: s_load_dwordx2 diff --git a/llvm/test/CodeGen/AMDGPU/mad-mix-hi.ll b/llvm/test/CodeGen/AMDGPU/mad-mix-hi.ll index 3f66c23e1a73..259ee0b26d2d 100644 --- a/llvm/test/CodeGen/AMDGPU/mad-mix-hi.ll +++ b/llvm/test/CodeGen/AMDGPU/mad-mix-hi.ll @@ -488,13 +488,11 @@ define <2 x half> @v_mad_mixhi_f16_f16lo_f16lo_f16lo_undeflo_clamp_postcvt(half ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, v2 ; GISEL-CI-NEXT: v_mac_f32_e32 v2, v0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v2 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, 0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GISEL-CI-NEXT: v_max_f32_e32 v0, v0, v1 +; GISEL-CI-NEXT: v_max_f32_e32 v0, 0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GISEL-CI-NEXT: v_min_f32_e32 v0, v0, v1 +; GISEL-CI-NEXT: v_min_f32_e32 v0, 1.0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v0 ; GISEL-CI-NEXT: s_setpc_b64 s[30:31] %src0.ext = fpext half %src0 to float @@ -582,15 +580,13 @@ define <2 x half> @v_mad_mixhi_f16_f16lo_f16lo_f16lo_undeflo_clamp_postcvt_multi ; GISEL-CI-NEXT: s_mov_b32 s7, 0xf000 ; GISEL-CI-NEXT: v_mac_f32_e32 v2, v0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v2 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, 0 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, v0 +; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GISEL-CI-NEXT: buffer_store_short v0, off, s[4:7], 0 ; GISEL-CI-NEXT: s_waitcnt vmcnt(0) -; GISEL-CI-NEXT: v_max_f32_e32 v1, v2, v1 +; GISEL-CI-NEXT: v_max_f32_e32 v1, 0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GISEL-CI-NEXT: v_min_f32_e32 v1, v1, v2 +; GISEL-CI-NEXT: v_min_f32_e32 v1, 1.0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GISEL-CI-NEXT: s_setpc_b64 s[30:31] %src0.ext = fpext half %src0 to float diff --git a/llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll b/llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll index 21e6faf46f58..ba77552e5809 100644 --- a/llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll +++ b/llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll @@ -313,13 +313,11 @@ define half @v_mad_mixlo_f16_f16lo_f16lo_f32_clamp_post_cvt(half %src0, half %sr ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 ; GISEL-CI-NEXT: v_mac_f32_e32 v2, v0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v2 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, 0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GISEL-CI-NEXT: v_max_f32_e32 v0, v0, v1 +; GISEL-CI-NEXT: v_max_f32_e32 v0, 0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, 1.0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GISEL-CI-NEXT: v_min_f32_e32 v0, v0, v1 +; GISEL-CI-NEXT: v_min_f32_e32 v0, 1.0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GISEL-CI-NEXT: s_setpc_b64 s[30:31] %src0.ext = fpext half %src0 to float @@ -1009,28 +1007,26 @@ define <2 x half> @v_mad_mix_v2f32_clamp_postcvt(<2 x half> %src0, <2 x half> %s ; GISEL-CI: ; %bb.0: ; GISEL-CI-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 +; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, v2 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v4, v4 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v3, v3 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v5, v5 ; GISEL-CI-NEXT: v_mac_f32_e32 v4, v0, v2 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v4 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 0 ; GISEL-CI-NEXT: v_mac_f32_e32 v5, v1, v3 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v5 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GISEL-CI-NEXT: v_max_f32_e32 v0, v0, v2 +; GISEL-CI-NEXT: v_max_f32_e32 v0, 0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GISEL-CI-NEXT: v_max_f32_e32 v1, v1, v2 +; GISEL-CI-NEXT: v_max_f32_e32 v1, 0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GISEL-CI-NEXT: v_min_f32_e32 v0, v0, v2 +; GISEL-CI-NEXT: v_min_f32_e32 v0, 1.0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 -; GISEL-CI-NEXT: v_min_f32_e32 v1, v1, v2 +; GISEL-CI-NEXT: v_min_f32_e32 v1, 1.0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GISEL-CI-NEXT: s_setpc_b64 s[30:31] %src0.ext = fpext <2 x half> %src0 to <2 x float> @@ -1225,25 +1221,23 @@ define <3 x half> @v_mad_mix_v3f32_clamp_postcvt(<3 x half> %src0, <3 x half> %s ; GISEL-CI-NEXT: v_mac_f32_e32 v7, v1, v4 ; GISEL-CI-NEXT: v_mac_f32_e32 v8, v2, v5 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v7 -; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v3, v8 +; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v2, v8 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v3, v3 -; GISEL-CI-NEXT: v_max_f32_e32 v0, v0, v2 -; GISEL-CI-NEXT: v_max_f32_e32 v1, v1, v2 -; GISEL-CI-NEXT: v_max_f32_e32 v2, v3, v2 +; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, v2 +; GISEL-CI-NEXT: v_max_f32_e32 v0, 0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GISEL-CI-NEXT: v_max_f32_e32 v1, 0, v1 +; GISEL-CI-NEXT: v_max_f32_e32 v2, 0, v2 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v2, v2 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v3, 1.0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, v2 -; GISEL-CI-NEXT: v_min_f32_e32 v0, v0, v3 -; GISEL-CI-NEXT: v_min_f32_e32 v1, v1, v3 -; GISEL-CI-NEXT: v_min_f32_e32 v2, v2, v3 +; GISEL-CI-NEXT: v_min_f32_e32 v0, 1.0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 +; GISEL-CI-NEXT: v_min_f32_e32 v1, 1.0, v1 +; GISEL-CI-NEXT: v_min_f32_e32 v2, 1.0, v2 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v2, v2 ; GISEL-CI-NEXT: s_setpc_b64 s[30:31] @@ -1441,30 +1435,28 @@ define <4 x half> @v_mad_mix_v4f32_clamp_postcvt(<4 x half> %src0, <4 x half> %s ; GISEL-CI-NEXT: v_mac_f32_e32 v11, v3, v7 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v8 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v9 -; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v3, v10 -; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v4, v11 +; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v2, v10 +; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v3, v11 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 +; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, v2 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v3, v3 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v4, v4 -; GISEL-CI-NEXT: v_max_f32_e32 v0, v0, v2 -; GISEL-CI-NEXT: v_max_f32_e32 v1, v1, v2 -; GISEL-CI-NEXT: v_max_f32_e32 v3, v3, v2 -; GISEL-CI-NEXT: v_max_f32_e32 v2, v4, v2 +; GISEL-CI-NEXT: v_max_f32_e32 v0, 0, v0 +; GISEL-CI-NEXT: v_max_f32_e32 v1, 0, v1 +; GISEL-CI-NEXT: v_max_f32_e32 v2, 0, v2 +; GISEL-CI-NEXT: v_max_f32_e32 v3, 0, v3 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 -; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v2, v2 +; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v3, v3 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v5, 1.0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 +; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, v2 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v3, v3 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v4, v2 -; GISEL-CI-NEXT: v_min_f32_e32 v0, v0, v5 -; GISEL-CI-NEXT: v_min_f32_e32 v1, v1, v5 -; GISEL-CI-NEXT: v_min_f32_e32 v2, v3, v5 -; GISEL-CI-NEXT: v_min_f32_e32 v3, v4, v5 +; GISEL-CI-NEXT: v_min_f32_e32 v0, 1.0, v0 +; GISEL-CI-NEXT: v_min_f32_e32 v1, 1.0, v1 +; GISEL-CI-NEXT: v_min_f32_e32 v2, 1.0, v2 +; GISEL-CI-NEXT: v_min_f32_e32 v3, 1.0, v3 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v2, v2 @@ -1622,16 +1614,14 @@ define <2 x half> @v_mad_mix_v2f32_clamp_postcvt_lo(<2 x half> %src0, <2 x half> ; GISEL-CI-NEXT: v_mac_f32_e32 v4, v0, v2 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v5 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v4 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 0 ; GISEL-CI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 ; GISEL-CI-NEXT: v_or_b32_e32 v0, v1, v0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v0 ; GISEL-CI-NEXT: v_and_b32_e32 v0, 0xffff0000, v0 -; GISEL-CI-NEXT: v_max_f32_e32 v1, v1, v2 +; GISEL-CI-NEXT: v_max_f32_e32 v1, 0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GISEL-CI-NEXT: v_min_f32_e32 v1, v1, v2 +; GISEL-CI-NEXT: v_min_f32_e32 v1, 1.0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GISEL-CI-NEXT: v_or_b32_e32 v0, v0, v1 ; GISEL-CI-NEXT: v_lshrrev_b32_e32 v1, 16, v0 @@ -1790,17 +1780,15 @@ define <2 x half> @v_mad_mix_v2f32_clamp_postcvt_hi(<2 x half> %src0, <2 x half> ; GISEL-CI-NEXT: v_mac_f32_e32 v4, v0, v2 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v0, v5 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v4 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 0 ; GISEL-CI-NEXT: v_lshlrev_b32_e32 v0, 16, v0 ; GISEL-CI-NEXT: v_or_b32_e32 v0, v1, v0 ; GISEL-CI-NEXT: v_lshrrev_b32_e32 v1, 16, v0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 ; GISEL-CI-NEXT: v_and_b32_e32 v0, 0xffff, v0 -; GISEL-CI-NEXT: v_max_f32_e32 v1, v1, v2 +; GISEL-CI-NEXT: v_max_f32_e32 v1, 0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 -; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v2, 1.0 ; GISEL-CI-NEXT: v_cvt_f32_f16_e32 v1, v1 -; GISEL-CI-NEXT: v_min_f32_e32 v1, v1, v2 +; GISEL-CI-NEXT: v_min_f32_e32 v1, 1.0, v1 ; GISEL-CI-NEXT: v_cvt_f16_f32_e32 v1, v1 ; GISEL-CI-NEXT: v_lshlrev_b32_e32 v1, 16, v1 ; GISEL-CI-NEXT: v_or_b32_e32 v0, v0, v1 diff --git a/llvm/test/CodeGen/AMDGPU/maximumnum.ll b/llvm/test/CodeGen/AMDGPU/maximumnum.ll index 4f73e8e9c188..c90b2c917041 100644 --- a/llvm/test/CodeGen/AMDGPU/maximumnum.ll +++ b/llvm/test/CodeGen/AMDGPU/maximumnum.ll @@ -271,8 +271,7 @@ define half @v_maximumnum_f16_1.0(half %x) { ; GFX7-GISEL: ; %bb.0: ; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX7-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX7-GISEL-NEXT: v_cvt_f32_f16_e32 v1, 1.0 -; GFX7-GISEL-NEXT: v_max_f32_e32 v0, v0, v1 +; GFX7-GISEL-NEXT: v_max_f32_e32 v0, 1.0, v0 ; GFX7-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/mfma-loop.ll b/llvm/test/CodeGen/AMDGPU/mfma-loop.ll index 0af655dfbbee..4bb653848cbf 100644 --- a/llvm/test/CodeGen/AMDGPU/mfma-loop.ll +++ b/llvm/test/CodeGen/AMDGPU/mfma-loop.ll @@ -2399,8 +2399,9 @@ define amdgpu_kernel void @test_mfma_nested_loop_zeroinit(ptr addrspace(1) %arg) ; GFX90A-NEXT: v_accvgpr_mov_b32 a29, a0 ; GFX90A-NEXT: v_accvgpr_mov_b32 a30, a0 ; GFX90A-NEXT: v_accvgpr_mov_b32 a31, a0 -; GFX90A-NEXT: v_mov_b32_e32 v0, 2.0 -; GFX90A-NEXT: v_mov_b32_e32 v1, 1.0 +; GFX90A-NEXT: v_mov_b32_e32 v0, 1.0 +; GFX90A-NEXT: v_mov_b32_e32 v1, 2.0 +; GFX90A-NEXT: ; kill: def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 killed $exec ; GFX90A-NEXT: .LBB9_1: ; %for.cond.preheader ; GFX90A-NEXT: ; =>This Loop Header: Depth=1 ; GFX90A-NEXT: ; Child Loop BB9_2 Depth 2 @@ -2409,7 +2410,7 @@ define amdgpu_kernel void @test_mfma_nested_loop_zeroinit(ptr addrspace(1) %arg) ; GFX90A-NEXT: ; Parent Loop BB9_1 Depth=1 ; GFX90A-NEXT: ; => This Inner Loop Header: Depth=2 ; GFX90A-NEXT: s_nop 0 -; GFX90A-NEXT: v_mfma_f32_32x32x1f32 a[0:31], v1, v0, a[0:31] +; GFX90A-NEXT: v_mfma_f32_32x32x1f32 a[0:31], v0, v1, a[0:31] ; GFX90A-NEXT: s_add_i32 s1, s1, -1 ; GFX90A-NEXT: s_cmp_lg_u32 s1, 0 ; GFX90A-NEXT: s_cbranch_scc1 .LBB9_2 @@ -2468,8 +2469,9 @@ define amdgpu_kernel void @test_mfma_nested_loop_zeroinit(ptr addrspace(1) %arg) ; GFX942-NEXT: v_accvgpr_mov_b32 a29, a0 ; GFX942-NEXT: v_accvgpr_mov_b32 a30, a0 ; GFX942-NEXT: v_accvgpr_mov_b32 a31, a0 -; GFX942-NEXT: v_mov_b32_e32 v0, 2.0 -; GFX942-NEXT: v_mov_b32_e32 v1, 1.0 +; GFX942-NEXT: v_mov_b32_e32 v0, 1.0 +; GFX942-NEXT: v_mov_b32_e32 v1, 2.0 +; GFX942-NEXT: ; kill: def $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 killed $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15_agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31 killed $exec ; GFX942-NEXT: .LBB9_1: ; %for.cond.preheader ; GFX942-NEXT: ; =>This Loop Header: Depth=1 ; GFX942-NEXT: ; Child Loop BB9_2 Depth 2 @@ -2478,7 +2480,7 @@ define amdgpu_kernel void @test_mfma_nested_loop_zeroinit(ptr addrspace(1) %arg) ; GFX942-NEXT: ; Parent Loop BB9_1 Depth=1 ; GFX942-NEXT: ; => This Inner Loop Header: Depth=2 ; GFX942-NEXT: s_nop 0 -; GFX942-NEXT: v_mfma_f32_32x32x1_2b_f32 a[0:31], v1, v0, a[0:31] +; GFX942-NEXT: v_mfma_f32_32x32x1_2b_f32 a[0:31], v0, v1, a[0:31] ; GFX942-NEXT: s_add_i32 s1, s1, -1 ; GFX942-NEXT: s_cmp_lg_u32 s1, 0 ; GFX942-NEXT: s_cbranch_scc1 .LBB9_2 diff --git a/llvm/test/CodeGen/AMDGPU/minimumnum.ll b/llvm/test/CodeGen/AMDGPU/minimumnum.ll index 558006d2b695..64e8b7b50de0 100644 --- a/llvm/test/CodeGen/AMDGPU/minimumnum.ll +++ b/llvm/test/CodeGen/AMDGPU/minimumnum.ll @@ -271,8 +271,7 @@ define half @v_minimumnum_f16_1.0(half %x) { ; GFX7-GISEL: ; %bb.0: ; GFX7-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX7-GISEL-NEXT: v_cvt_f32_f16_e32 v0, v0 -; GFX7-GISEL-NEXT: v_cvt_f32_f16_e32 v1, 1.0 -; GFX7-GISEL-NEXT: v_min_f32_e32 v0, v0, v1 +; GFX7-GISEL-NEXT: v_min_f32_e32 v0, 1.0, v0 ; GFX7-GISEL-NEXT: v_cvt_f16_f32_e32 v0, v0 ; GFX7-GISEL-NEXT: s_setpc_b64 s[30:31] ; diff --git a/llvm/test/CodeGen/AMDGPU/naked-fn-with-frame-pointer.ll b/llvm/test/CodeGen/AMDGPU/naked-fn-with-frame-pointer.ll index 5ff2d82c1464..2509497bbcde 100644 --- a/llvm/test/CodeGen/AMDGPU/naked-fn-with-frame-pointer.ll +++ b/llvm/test/CodeGen/AMDGPU/naked-fn-with-frame-pointer.ll @@ -5,8 +5,8 @@ declare dso_local void @main() define dso_local void @naked() naked "frame-pointer"="all" { ; CHECK-LABEL: naked: -; CHECK: naked$local: -; CHECK-NEXT: .type naked$local,@function +; CHECK: .Lnaked$local: +; CHECK-NEXT: .type .Lnaked$local,@function ; CHECK-NEXT: ; %bb.0: ; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; CHECK-NEXT: s_getpc_b64 s[16:17] @@ -19,8 +19,8 @@ define dso_local void @naked() naked "frame-pointer"="all" { define dso_local void @normal() "frame-pointer"="all" { ; CHECK-LABEL: normal: -; CHECK: normal$local: -; CHECK-NEXT: .type normal$local,@function +; CHECK: .Lnormal$local: +; CHECK-NEXT: .type .Lnormal$local,@function ; CHECK-NEXT: ; %bb.0: ; CHECK-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; CHECK-NEXT: s_mov_b32 s16, s33 diff --git a/llvm/test/CodeGen/AMDGPU/regpressure_printer.mir b/llvm/test/CodeGen/AMDGPU/regpressure_printer.mir index 8d5b5e4e8cae..b41aa088bfdd 100644 --- a/llvm/test/CodeGen/AMDGPU/regpressure_printer.mir +++ b/llvm/test/CodeGen/AMDGPU/regpressure_printer.mir @@ -510,14 +510,14 @@ body: | ; RPU-NEXT: 0 0 $sgpr0 = S_BUFFER_LOAD_DWORD_IMM $sgpr0_sgpr1_sgpr2_sgpr3, 0, 0 ; RPU-NEXT: 0 0 ; RPU-NEXT: 0 1 undef %0.sub5:vreg_512 = V_MOV_B32_e32 5, implicit $exec - ; RPU-NEXT: 0 0 - ; RPU-NEXT: 0 0 S_CMP_GT_U32 $sgpr0, 15, implicit-def $scc - ; RPU-NEXT: 0 0 - ; RPU-NEXT: 0 0 S_CBRANCH_SCC1 %bb.2, implicit $scc - ; RPU-NEXT: 0 0 - ; RPU-NEXT: 0 0 S_BRANCH %bb.1 - ; RPU-NEXT: 0 0 - ; RPU-NEXT: Live-out: + ; RPU-NEXT: 0 1 + ; RPU-NEXT: 0 1 S_CMP_GT_U32 $sgpr0, 15, implicit-def $scc + ; RPU-NEXT: 0 1 + ; RPU-NEXT: 0 1 S_CBRANCH_SCC1 %bb.2, implicit $scc + ; RPU-NEXT: 0 1 + ; RPU-NEXT: 0 1 S_BRANCH %bb.1 + ; RPU-NEXT: 0 1 + ; RPU-NEXT: Live-out: %0:0000000000000C00 ; RPU-NEXT: Live-thr: ; RPU-NEXT: 0 0 ; RPU-NEXT: bb.1: @@ -571,8 +571,6 @@ body: | ; RPD-NEXT: 0 1 S_BRANCH %bb.1 ; RPD-NEXT: 0 1 ; RPD-NEXT: Live-out: %0:0000000000000C00 - ; RPD-NEXT: mis LIS: - ; RPD-NEXT: %0:L0000000000000C00 isn't found in LIS reported set ; RPD-NEXT: Live-thr: ; RPD-NEXT: 0 0 ; RPD-NEXT: bb.1: diff --git a/llvm/test/CodeGen/AMDGPU/s_cmp_0.ll b/llvm/test/CodeGen/AMDGPU/s_cmp_0.ll new file mode 100644 index 000000000000..f53aaaad87e1 --- /dev/null +++ b/llvm/test/CodeGen/AMDGPU/s_cmp_0.ll @@ -0,0 +1,625 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 +; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s | FileCheck %s + +declare i32 @llvm.ctpop.i32(i32) +declare i64 @llvm.ctpop.i64(i64) +declare i32 @llvm.amdgcn.s.quadmask.i32(i32) +declare i64 @llvm.amdgcn.s.quadmask.i64(i64) + +define amdgpu_ps i32 @shl32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: shl32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_lshl_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = shl i32 %val0, %val1 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @shl64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: shl64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_lshl_b64 s[0:1], s[0:1], s2 +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = shl i64 %val0, %val1 + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @lshr32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: lshr32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_lshr_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = lshr i32 %val0, %val1 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @lshr64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: lshr64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_lshr_b64 s[0:1], s[0:1], s2 +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = lshr i64 %val0, %val1 + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @ashr32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: ashr32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_ashr_i32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = ashr i32 %val0, %val1 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @ashr64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: ashr64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_ashr_i64 s[0:1], s[0:1], s2 +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = ashr i64 %val0, %val1 + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @abs32(i32 inreg %val0) { +; CHECK-LABEL: abs32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_abs_i32 s0, s0 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s0 +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %neg = sub i32 0, %val0 + %cond = icmp sgt i32 %val0, %neg + %result = select i1 %cond, i32 %val0, i32 %neg + call void asm "; use $0", "s"(i32 %result) + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @and32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: and32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_and_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = and i32 %val0, %val1 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @and64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: and64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_and_b64 s[0:1], s[0:1], s[2:3] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = and i64 %val0, %val1 + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @or32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: or32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_or_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = or i32 %val0, %val1 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @or64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: or64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_or_b64 s[0:1], s[0:1], s[2:3] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = or i64 %val0, %val1 + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @xor32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: xor32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_xor_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = xor i32 %val0, %val1 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @xor64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: xor64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_xor_b64 s[0:1], s[0:1], s[2:3] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = xor i64 %val0, %val1 + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @nand32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: nand32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_nand_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s0 +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = and i32 %val0, %val1 + %result2 = xor i32 %result, -1 + call void asm "; use $0", "s"(i32 %result2) + %cmp = icmp ne i32 %result2, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @nand64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: nand64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_nand_b64 s[0:1], s[0:1], s[2:3] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[0:1] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = and i64 %val0, %val1 + %result2 = xor i64 %result, -1 + call void asm "; use $0", "s"(i64 %result2) + %cmp = icmp ne i64 %result2, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @nor32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: nor32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_nor_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s0 +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = or i32 %val0, %val1 + %result2 = xor i32 %result, -1 + call void asm "; use $0", "s"(i32 %result2) + %cmp = icmp ne i32 %result2, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @nor64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: nor64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_nor_b64 s[0:1], s[0:1], s[2:3] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[0:1] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = or i64 %val0, %val1 + %result2 = xor i64 %result, -1 + call void asm "; use $0", "s"(i64 %result2) + %cmp = icmp ne i64 %result2, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @xnor32(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: xnor32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_xnor_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s0 +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = xor i32 %val0, %val1 + %result2 = xor i32 %result, -1 + call void asm "; use $0", "s"(i32 %result2) + %cmp = icmp ne i32 %result2, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @xnor64(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: xnor64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_xnor_b64 s[0:1], s[0:1], s[2:3] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[0:1] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = xor i64 %val0, %val1 + %result2 = xor i64 %result, -1 + call void asm "; use $0", "s"(i64 %result2) + %cmp = icmp ne i64 %result2, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @andn232(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: andn232: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_andn2_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %nval1 = xor i32 %val1, -1 + %result = and i32 %val0, %nval1 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @nandn264(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: nandn264: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_andn2_b64 s[0:1], s[0:1], s[2:3] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %nval1 = xor i64 %val1, -1 + %result = and i64 %val0, %nval1 + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @orn232(i32 inreg %val0, i32 inreg %val1) { +; CHECK-LABEL: orn232: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_orn2_b32 s0, s0, s1 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %nval1 = xor i32 %val1, -1 + %result = or i32 %val0, %nval1 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @orn264(i64 inreg %val0, i64 inreg %val1) { +; CHECK-LABEL: orn264: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_orn2_b64 s[0:1], s[0:1], s[2:3] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %nval1 = xor i64 %val1, -1 + %result = or i64 %val0, %nval1 + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @bfe_i32(i32 inreg %val0) { +; CHECK-LABEL: bfe_i32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_bfe_i32 s0, s0, 0x80010 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %shl = shl i32 %val0, 8 + %result = ashr i32 %shl, 24 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @bfe_i64(i64 inreg %val0) { +; CHECK-LABEL: bfe_i64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_bfe_i64 s[2:3], s[0:1], 0x80000 +; CHECK-NEXT: s_and_b32 s0, s0, 0xff +; CHECK-NEXT: s_mov_b32 s1, 0 +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[2:3] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: ; return to shader part epilog + %shl = shl i64 %val0, 56 + %result = ashr i64 %shl, 56 + call void asm "; use $0", "s"(i64 %result) + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @bfe_u32(i32 inreg %val0) { +; CHECK-LABEL: bfe_u32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_bfe_u32 s0, s0, 0x80010 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %shl = shl i32 %val0, 8 + %result = lshr i32 %shl, 24 + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @bfe_u64(i64 inreg %val0) { +; CHECK-LABEL: bfe_u64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_and_b32 s0, s0, 0xff +; CHECK-NEXT: s_mov_b32 s1, 0 +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[0:1] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %shl = shl i64 %val0, 56 + %result = lshr i64 %shl, 56 + call void asm "; use $0", "s"(i64 %result) + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @bcnt032(i32 inreg %val0) { +; CHECK-LABEL: bcnt032: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_bcnt1_i32_b32 s0, s0 +; CHECK-NEXT: s_sub_i32 s0, 32, s0 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s0 +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = call i32 @llvm.ctpop.i32(i32 %val0) nounwind readnone + %result2 = sub i32 32, %result + call void asm "; use $0", "s"(i32 %result2) + %cmp = icmp ne i32 %result2, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @bcnt064(i64 inreg %val0) { +; CHECK-LABEL: bcnt064: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_bcnt1_i32_b64 s0, s[0:1] +; CHECK-NEXT: s_sub_u32 s0, 64, s0 +; CHECK-NEXT: s_subb_u32 s1, 0, 0 +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[0:1] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = call i64 @llvm.ctpop.i64(i64 %val0) nounwind readnone + %result2 = sub i64 64, %result + call void asm "; use $0", "s"(i64 %result2) + %cmp = icmp ne i64 %result2, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @bcnt132(i32 inreg %val0) { +; CHECK-LABEL: bcnt132: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_bcnt1_i32_b32 s0, s0 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s0 +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = call i32 @llvm.ctpop.i32(i32 %val0) nounwind readnone + call void asm "; use $0", "s"(i32 %result) + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @bcnt164(i64 inreg %val0) { +; CHECK-LABEL: bcnt164: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_bcnt1_i32_b64 s0, s[0:1] +; CHECK-NEXT: s_mov_b32 s1, 0 +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[0:1] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = call i64 @llvm.ctpop.i64(i64 %val0) nounwind readnone + call void asm "; use $0", "s"(i64 %result) + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @quadmask32(i32 inreg %val0) { +; CHECK-LABEL: quadmask32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_quadmask_b32 s0, s0 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s0 +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = call i32 @llvm.amdgcn.s.quadmask.i32(i32 %val0) nounwind readnone + call void asm "; use $0", "s"(i32 %result) + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @quadmask64(i64 inreg %val0) { +; CHECK-LABEL: quadmask64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_quadmask_b64 s[0:1], s[0:1] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[0:1] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = call i64 @llvm.amdgcn.s.quadmask.i64(i64 %val0) nounwind readnone + call void asm "; use $0", "s"(i64 %result) + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @not32(i32 inreg %val0) { +; CHECK-LABEL: not32: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_not_b32 s0, s0 +; CHECK-NEXT: s_cmp_lg_u32 s0, 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s0 +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = xor i32 %val0, -1 + call void asm "; use $0", "s"(i32 %result) + %cmp = icmp ne i32 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} + +define amdgpu_ps i32 @not64(i64 inreg %val0) { +; CHECK-LABEL: not64: +; CHECK: ; %bb.0: +; CHECK-NEXT: s_not_b64 s[0:1], s[0:1] +; CHECK-NEXT: s_cmp_lg_u64 s[0:1], 0 +; CHECK-NEXT: ;;#ASMSTART +; CHECK-NEXT: ; use s[0:1] +; CHECK-NEXT: ;;#ASMEND +; CHECK-NEXT: s_cselect_b64 s[0:1], -1, 0 +; CHECK-NEXT: v_cndmask_b32_e64 v0, 0, 1, s[0:1] +; CHECK-NEXT: v_readfirstlane_b32 s0, v0 +; CHECK-NEXT: ; return to shader part epilog + %result = xor i64 %val0, -1 + call void asm "; use $0", "s"(i64 %result) + %cmp = icmp ne i64 %result, 0 + %zext = zext i1 %cmp to i32 + ret i32 %zext +} diff --git a/llvm/test/CodeGen/AMDGPU/uniform-select.ll b/llvm/test/CodeGen/AMDGPU/uniform-select.ll index f001bf0d5e49..b52913fbb969 100644 --- a/llvm/test/CodeGen/AMDGPU/uniform-select.ll +++ b/llvm/test/CodeGen/AMDGPU/uniform-select.ll @@ -20,34 +20,34 @@ define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) { ; GFX90A-NEXT: s_cmp_eq_u32 s1, 1 ; GFX90A-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX90A-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s7, s4, s3 +; GFX90A-NEXT: s_cselect_b32 s7, s3, s2 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 2 ; GFX90A-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX90A-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s7, s5, s7 +; GFX90A-NEXT: s_cselect_b32 s7, s4, s7 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 3 ; GFX90A-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX90A-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s7, s6, s7 +; GFX90A-NEXT: s_cselect_b32 s7, s5, s7 ; GFX90A-NEXT: s_or_b32 s7, s7, s0 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 1 ; GFX90A-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX90A-NEXT: s_and_b64 s[10:11], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s4, s7, s4 +; GFX90A-NEXT: s_cselect_b32 s3, s7, s3 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 3 ; GFX90A-NEXT: s_cselect_b64 s[10:11], -1, 0 ; GFX90A-NEXT: s_and_b64 s[12:13], s[10:11], exec -; GFX90A-NEXT: s_cselect_b32 s6, s7, s6 +; GFX90A-NEXT: s_cselect_b32 s5, s7, s5 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 2 ; GFX90A-NEXT: s_cselect_b64 s[12:13], -1, 0 ; GFX90A-NEXT: s_and_b64 s[14:15], s[12:13], exec -; GFX90A-NEXT: s_cselect_b32 s5, s7, s5 +; GFX90A-NEXT: s_cselect_b32 s4, s7, s4 ; GFX90A-NEXT: s_cmp_eq_u32 s1, 0 -; GFX90A-NEXT: s_cselect_b32 s3, s7, s3 +; GFX90A-NEXT: s_cselect_b32 s2, s7, s2 ; GFX90A-NEXT: s_or_b64 s[8:9], s[12:13], s[8:9] ; GFX90A-NEXT: s_or_b64 s[8:9], s[10:11], s[8:9] ; GFX90A-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX90A-NEXT: s_cselect_b32 s2, 0, s2 +; GFX90A-NEXT: s_cselect_b32 s6, 0, s6 ; GFX90A-NEXT: s_mov_b64 vcc, vcc ; GFX90A-NEXT: s_cbranch_vccnz .LBB0_1 ; GFX90A-NEXT: ; %bb.2: ; %DummyReturnBlock @@ -68,34 +68,34 @@ define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) { ; GFX942-NEXT: s_cmp_eq_u32 s1, 1 ; GFX942-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX942-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s7, s4, s3 +; GFX942-NEXT: s_cselect_b32 s7, s3, s2 ; GFX942-NEXT: s_cmp_eq_u32 s1, 2 ; GFX942-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX942-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s7, s5, s7 +; GFX942-NEXT: s_cselect_b32 s7, s4, s7 ; GFX942-NEXT: s_cmp_eq_u32 s1, 3 ; GFX942-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX942-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s7, s6, s7 +; GFX942-NEXT: s_cselect_b32 s7, s5, s7 ; GFX942-NEXT: s_or_b32 s7, s7, s0 ; GFX942-NEXT: s_cmp_eq_u32 s1, 1 ; GFX942-NEXT: s_cselect_b64 s[8:9], -1, 0 ; GFX942-NEXT: s_and_b64 s[10:11], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s4, s7, s4 +; GFX942-NEXT: s_cselect_b32 s3, s7, s3 ; GFX942-NEXT: s_cmp_eq_u32 s1, 3 ; GFX942-NEXT: s_cselect_b64 s[10:11], -1, 0 ; GFX942-NEXT: s_and_b64 s[12:13], s[10:11], exec -; GFX942-NEXT: s_cselect_b32 s6, s7, s6 +; GFX942-NEXT: s_cselect_b32 s5, s7, s5 ; GFX942-NEXT: s_cmp_eq_u32 s1, 2 ; GFX942-NEXT: s_cselect_b64 s[12:13], -1, 0 ; GFX942-NEXT: s_and_b64 s[14:15], s[12:13], exec -; GFX942-NEXT: s_cselect_b32 s5, s7, s5 +; GFX942-NEXT: s_cselect_b32 s4, s7, s4 ; GFX942-NEXT: s_cmp_eq_u32 s1, 0 -; GFX942-NEXT: s_cselect_b32 s3, s7, s3 +; GFX942-NEXT: s_cselect_b32 s2, s7, s2 ; GFX942-NEXT: s_or_b64 s[8:9], s[12:13], s[8:9] ; GFX942-NEXT: s_or_b64 s[8:9], s[10:11], s[8:9] ; GFX942-NEXT: s_and_b64 s[8:9], s[8:9], exec -; GFX942-NEXT: s_cselect_b32 s2, 0, s2 +; GFX942-NEXT: s_cselect_b32 s6, 0, s6 ; GFX942-NEXT: s_mov_b64 vcc, vcc ; GFX942-NEXT: s_cbranch_vccnz .LBB0_1 ; GFX942-NEXT: ; %bb.2: ; %DummyReturnBlock @@ -117,34 +117,34 @@ define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) { ; GFX1030-NEXT: s_cmp_eq_u32 s1, 1 ; GFX1030-NEXT: s_cselect_b32 s7, -1, 0 ; GFX1030-NEXT: s_and_b32 s7, s7, exec_lo -; GFX1030-NEXT: s_cselect_b32 s7, s4, s3 +; GFX1030-NEXT: s_cselect_b32 s7, s3, s2 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 2 ; GFX1030-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1030-NEXT: s_and_b32 s8, s8, exec_lo -; GFX1030-NEXT: s_cselect_b32 s7, s5, s7 +; GFX1030-NEXT: s_cselect_b32 s7, s4, s7 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 3 ; GFX1030-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1030-NEXT: s_and_b32 s8, s8, exec_lo -; GFX1030-NEXT: s_cselect_b32 s7, s6, s7 +; GFX1030-NEXT: s_cselect_b32 s7, s5, s7 ; GFX1030-NEXT: s_or_b32 s7, s7, s0 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 1 ; GFX1030-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1030-NEXT: s_and_b32 s9, s8, exec_lo -; GFX1030-NEXT: s_cselect_b32 s4, s7, s4 +; GFX1030-NEXT: s_cselect_b32 s3, s7, s3 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 3 ; GFX1030-NEXT: s_cselect_b32 s9, -1, 0 ; GFX1030-NEXT: s_and_b32 s10, s9, exec_lo -; GFX1030-NEXT: s_cselect_b32 s6, s7, s6 +; GFX1030-NEXT: s_cselect_b32 s5, s7, s5 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 2 ; GFX1030-NEXT: s_cselect_b32 s10, -1, 0 ; GFX1030-NEXT: s_and_b32 s11, s10, exec_lo -; GFX1030-NEXT: s_cselect_b32 s5, s7, s5 +; GFX1030-NEXT: s_cselect_b32 s4, s7, s4 ; GFX1030-NEXT: s_cmp_eq_u32 s1, 0 -; GFX1030-NEXT: s_cselect_b32 s3, s7, s3 +; GFX1030-NEXT: s_cselect_b32 s2, s7, s2 ; GFX1030-NEXT: s_or_b32 s7, s10, s8 ; GFX1030-NEXT: s_or_b32 s7, s9, s7 ; GFX1030-NEXT: s_and_b32 s7, s7, exec_lo -; GFX1030-NEXT: s_cselect_b32 s2, 0, s2 +; GFX1030-NEXT: s_cselect_b32 s6, 0, s6 ; GFX1030-NEXT: s_cbranch_vccnz .LBB0_1 ; GFX1030-NEXT: ; %bb.2: ; %DummyReturnBlock ; GFX1030-NEXT: s_endpgm @@ -166,38 +166,38 @@ define amdgpu_kernel void @test_insert_extract(i32 %p, i32 %q) { ; GFX1100-NEXT: s_cselect_b32 s7, -1, 0 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s7, s7, exec_lo -; GFX1100-NEXT: s_cselect_b32 s7, s4, s3 +; GFX1100-NEXT: s_cselect_b32 s7, s3, s2 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 2 ; GFX1100-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1100-NEXT: s_and_b32 s8, s8, exec_lo -; GFX1100-NEXT: s_cselect_b32 s7, s5, s7 +; GFX1100-NEXT: s_cselect_b32 s7, s4, s7 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 3 ; GFX1100-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s8, s8, exec_lo -; GFX1100-NEXT: s_cselect_b32 s7, s6, s7 +; GFX1100-NEXT: s_cselect_b32 s7, s5, s7 ; GFX1100-NEXT: s_or_b32 s7, s7, s0 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 1 ; GFX1100-NEXT: s_cselect_b32 s8, -1, 0 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s9, s8, exec_lo -; GFX1100-NEXT: s_cselect_b32 s4, s7, s4 +; GFX1100-NEXT: s_cselect_b32 s3, s7, s3 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 3 ; GFX1100-NEXT: s_cselect_b32 s9, -1, 0 ; GFX1100-NEXT: s_and_b32 s10, s9, exec_lo -; GFX1100-NEXT: s_cselect_b32 s6, s7, s6 +; GFX1100-NEXT: s_cselect_b32 s5, s7, s5 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 2 ; GFX1100-NEXT: s_cselect_b32 s10, -1, 0 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s11, s10, exec_lo -; GFX1100-NEXT: s_cselect_b32 s5, s7, s5 +; GFX1100-NEXT: s_cselect_b32 s4, s7, s4 ; GFX1100-NEXT: s_cmp_eq_u32 s1, 0 -; GFX1100-NEXT: s_cselect_b32 s3, s7, s3 +; GFX1100-NEXT: s_cselect_b32 s2, s7, s2 ; GFX1100-NEXT: s_or_b32 s7, s10, s8 ; GFX1100-NEXT: s_or_b32 s7, s9, s7 ; GFX1100-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX1100-NEXT: s_and_b32 s7, s7, exec_lo -; GFX1100-NEXT: s_cselect_b32 s2, 0, s2 +; GFX1100-NEXT: s_cselect_b32 s6, 0, s6 ; GFX1100-NEXT: s_cbranch_vccnz .LBB0_1 ; GFX1100-NEXT: ; %bb.2: ; %DummyReturnBlock ; GFX1100-NEXT: s_endpgm diff --git a/llvm/test/CodeGen/AMDGPU/unpack-non-coissue-insts-post-ra-scheduler.mir b/llvm/test/CodeGen/AMDGPU/unpack-non-coissue-insts-post-ra-scheduler.mir index 8b467eb0b054..75ae76fdee19 100644 --- a/llvm/test/CodeGen/AMDGPU/unpack-non-coissue-insts-post-ra-scheduler.mir +++ b/llvm/test/CodeGen/AMDGPU/unpack-non-coissue-insts-post-ra-scheduler.mir @@ -1,7 +1,7 @@ -# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5 +# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 6 # RUN: llc -mtriple amdgcn-amd-amdhsa -mcpu=gfx950 -run-pass=si-pre-emit-peephole -o - %s | FileCheck -check-prefix=GFX950 %s # RUN: llc -mtriple amdgcn-amd-amdhsa -mcpu=gfx942 -run-pass=si-pre-emit-peephole -o - %s | FileCheck -check-prefix=GFX942 %s -# RUN: llc -mtriple amdgcn-amd-amdhsa -mcpu=gfx90a -run-pass=si-pre-emit-peephole -o - %s | FileCheck -check-prefix=GFX90a %s +# RUN: llc -mtriple amdgcn-amd-amdhsa -mcpu=gfx90a -run-pass=si-pre-emit-peephole -o - %s | FileCheck -check-prefix=GFX90A %s --- name: test_pk_mul_unpacking_f32 @@ -57,26 +57,26 @@ body: | ; GFX942-NEXT: $vgpr17 = nofpexcept V_MUL_F32_e64 0, killed $sgpr31, 0, killed $vgpr5, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_pk_mul_unpacking_f32 - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 8, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_pk_mul_unpacking_f32 + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 8, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec S_WAITCNT 49279 @@ -150,26 +150,26 @@ body: | ; GFX942-NEXT: $vgpr17 = nofpexcept V_MUL_F32_e64 0, killed $sgpr31, 0, killed $vgpr5, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_op_sel_selection_unpacking_f32 - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 8, killed $sgpr30_sgpr31, 12, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_op_sel_selection_unpacking_f32 + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 8, killed $sgpr30_sgpr31, 12, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec S_WAITCNT 49279 @@ -243,26 +243,26 @@ body: | ; GFX942-NEXT: $vgpr17 = nofpexcept V_MUL_F32_e64 0, killed $sgpr30, 0, killed $vgpr5, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_op_sel_hi_selection_unpacking_f32 - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 0, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_op_sel_hi_selection_unpacking_f32 + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 0, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec S_WAITCNT 49279 @@ -370,41 +370,41 @@ body: | ; GFX942-NEXT: renamable $vgpr10_vgpr11 = nofpexcept V_PK_ADD_F32 8, killed $sgpr10_sgpr11, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_pk_add_unpacking_f32 - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr4 = V_MOV_B32_e32 2, implicit $exec - ; GFX90a-NEXT: renamable $vgpr5 = V_MOV_B32_e32 1, implicit $exec - ; GFX90a-NEXT: renamable $vgpr2 = V_MOV_B32_e32 4, implicit $exec - ; GFX90a-NEXT: renamable $vgpr3 = V_MOV_B32_e32 3, implicit $exec - ; GFX90a-NEXT: renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: renamable $vgpr16 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: $agpr31 = V_ACCVGPR_WRITE_B32_e64 $sgpr15, implicit $exec, implicit-def $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 $sgpr14, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 $sgpr13, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 $sgpr12, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 $sgpr11, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 $sgpr10, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 $sgpr9, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 $sgpr8, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 $sgpr7, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 $sgpr6, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 $sgpr5, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 $sgpr4, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 $sgpr3, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 $sgpr2, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 $sgpr1, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 - ; GFX90a-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 $sgpr0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15, implicit $exec - ; GFX90a-NEXT: $vgpr0 = V_MOV_B32_e32 killed $sgpr18, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr1 = V_MOV_B32_e32 killed $sgpr19, implicit $exec, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_BF8_BF8_e64 killed $vgpr4_vgpr5, killed $vgpr2_vgpr3, killed $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, 1, 2, 3, implicit $mode, implicit $exec, implicit $mode, implicit $exec - ; GFX90a-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_ADD_F32 8, killed $sgpr2_sgpr3, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: renamable $vgpr6_vgpr7 = nofpexcept V_PK_ADD_F32 8, killed $sgpr6_sgpr7, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: renamable $vgpr4_vgpr5 = nofpexcept V_PK_ADD_F32 8, killed $sgpr4_sgpr5, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: renamable $vgpr10_vgpr11 = nofpexcept V_PK_ADD_F32 8, killed $sgpr10_sgpr11, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_pk_add_unpacking_f32 + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr4 = V_MOV_B32_e32 2, implicit $exec + ; GFX90A-NEXT: renamable $vgpr5 = V_MOV_B32_e32 1, implicit $exec + ; GFX90A-NEXT: renamable $vgpr2 = V_MOV_B32_e32 4, implicit $exec + ; GFX90A-NEXT: renamable $vgpr3 = V_MOV_B32_e32 3, implicit $exec + ; GFX90A-NEXT: renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: renamable $vgpr16 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: $agpr31 = V_ACCVGPR_WRITE_B32_e64 $sgpr15, implicit $exec, implicit-def $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr30 = V_ACCVGPR_WRITE_B32_e64 $sgpr14, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr29 = V_ACCVGPR_WRITE_B32_e64 $sgpr13, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr28 = V_ACCVGPR_WRITE_B32_e64 $sgpr12, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr27 = V_ACCVGPR_WRITE_B32_e64 $sgpr11, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr26 = V_ACCVGPR_WRITE_B32_e64 $sgpr10, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr25 = V_ACCVGPR_WRITE_B32_e64 $sgpr9, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr24 = V_ACCVGPR_WRITE_B32_e64 $sgpr8, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr23 = V_ACCVGPR_WRITE_B32_e64 $sgpr7, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr22 = V_ACCVGPR_WRITE_B32_e64 $sgpr6, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr21 = V_ACCVGPR_WRITE_B32_e64 $sgpr5, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr20 = V_ACCVGPR_WRITE_B32_e64 $sgpr4, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr19 = V_ACCVGPR_WRITE_B32_e64 $sgpr3, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr18 = V_ACCVGPR_WRITE_B32_e64 $sgpr2, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr17 = V_ACCVGPR_WRITE_B32_e64 $sgpr1, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 + ; GFX90A-NEXT: $agpr16 = V_ACCVGPR_WRITE_B32_e64 $sgpr0, implicit $exec, implicit $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15, implicit $exec + ; GFX90A-NEXT: $vgpr0 = V_MOV_B32_e32 killed $sgpr18, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr1 = V_MOV_B32_e32 killed $sgpr19, implicit $exec, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_BF8_BF8_e64 killed $vgpr4_vgpr5, killed $vgpr2_vgpr3, killed $agpr16_agpr17_agpr18_agpr19_agpr20_agpr21_agpr22_agpr23_agpr24_agpr25_agpr26_agpr27_agpr28_agpr29_agpr30_agpr31, 1, 2, 3, implicit $mode, implicit $exec, implicit $mode, implicit $exec + ; GFX90A-NEXT: renamable $vgpr2_vgpr3 = nofpexcept V_PK_ADD_F32 8, killed $sgpr2_sgpr3, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: renamable $vgpr6_vgpr7 = nofpexcept V_PK_ADD_F32 8, killed $sgpr6_sgpr7, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: renamable $vgpr4_vgpr5 = nofpexcept V_PK_ADD_F32 8, killed $sgpr4_sgpr5, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: renamable $vgpr10_vgpr11 = nofpexcept V_PK_ADD_F32 8, killed $sgpr10_sgpr11, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr4 = V_MOV_B32_e32 2, implicit $exec renamable $vgpr5 = V_MOV_B32_e32 1, implicit $exec @@ -438,7 +438,6 @@ body: | renamable $vgpr10_vgpr11 = nofpexcept V_PK_ADD_F32 8, killed $sgpr10_sgpr11, 8, $vgpr0_vgpr1, 0, 0, 0, 0, 0, implicit $mode, implicit $exec S_ENDPGM 0 - ... --- name: test_pk_fma_unpacking_f32 @@ -490,24 +489,24 @@ body: | ; GFX942-NEXT: $vgpr17 = nofpexcept V_FMA_F32_e64 0, killed $sgpr30, 0, killed $vgpr5, 0, $vgpr5, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_pk_fma_unpacking_f32 - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 0, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 8, $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_pk_fma_unpacking_f32 + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 0, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 8, $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 @@ -577,25 +576,25 @@ body: | ; GFX942-NEXT: renamable $vgpr4_vgpr5 = nofpexcept V_PK_FMA_F32 8, killed $sgpr30_sgpr31, 8, $vgpr4_vgpr5, 0, $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_unpacking_does_not_introduce_rw_dependency - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: renamable $vgpr4_vgpr5 = nofpexcept V_PK_MUL_F32 8, $sgpr30_sgpr31, 12, $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: renamable $vgpr4_vgpr5 = nofpexcept V_PK_FMA_F32 8, killed $sgpr30_sgpr31, 8, $vgpr4_vgpr5, 0, $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_unpacking_does_not_introduce_rw_dependency + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: renamable $vgpr4_vgpr5 = nofpexcept V_PK_MUL_F32 8, $sgpr30_sgpr31, 12, $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: renamable $vgpr4_vgpr5 = nofpexcept V_PK_FMA_F32 8, killed $sgpr30_sgpr31, 8, $vgpr4_vgpr5, 0, $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 @@ -670,28 +669,28 @@ body: | ; GFX942-NEXT: $vgpr10_vgpr11 = V_PK_MOV_B32 12, killed $vgpr6_vgpr7, 8, killed $vgpr8_vgpr9, 0, 0, 0, 0, 0, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_opcodes_not_supported_for_unpacking_are_skipped - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_dpp $vgpr4, $vgpr4, 228, 15, 15, -1, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_CVT_PK_BF8_F32_e64 0, killed $vgpr4, 0, $vgpr4, $vgpr5, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: $vgpr8_vgpr9 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: $vgpr10_vgpr11 = V_PK_MOV_B32 12, killed $vgpr6_vgpr7, 8, killed $vgpr8_vgpr9, 0, 0, 0, 0, 0, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_opcodes_not_supported_for_unpacking_are_skipped + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_dpp $vgpr4, $vgpr4, 228, 15, 15, -1, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_CVT_PK_BF8_F32_e64 0, killed $vgpr4, 0, $vgpr4, $vgpr5, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: $vgpr8_vgpr9 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: $vgpr10_vgpr11 = V_PK_MOV_B32 12, killed $vgpr6_vgpr7, 8, killed $vgpr8_vgpr9, 0, 0, 0, 0, 0, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 @@ -705,14 +704,11 @@ body: | early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - $vgpr4 = V_MOV_B32_dpp $vgpr4, $vgpr4, 228, 15, 15, -1, implicit $exec $vgpr5 = V_CVT_PK_BF8_F32_e64 0, killed $vgpr4, 0, $vgpr4, $vgpr5, 0, implicit $mode, implicit $exec - $vgpr6_vgpr7 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec $vgpr8_vgpr9 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec $vgpr10_vgpr11 = V_PK_MOV_B32 12, killed $vgpr6_vgpr7, 8, killed $vgpr8_vgpr9, 0, 0, 0, 0, 0, implicit $exec - S_ENDPGM 0 ... @@ -770,26 +766,26 @@ body: | ; GFX942-NEXT: $vgpr17 = nofpexcept V_FMA_F32_e64 0, killed $sgpr30, 0, killed $vgpr5, 0, killed $vgpr7, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_opsel_register_is_correctly_marked_as_killed - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr6 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr7 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 0, killed $sgpr30_sgpr31, 12, killed $vgpr4_vgpr5, 8, killed $vgpr6_vgpr7, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_opsel_register_is_correctly_marked_as_killed + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr6 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr7 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 0, killed $sgpr30_sgpr31, 12, killed $vgpr4_vgpr5, 8, killed $vgpr6_vgpr7, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 @@ -861,26 +857,26 @@ body: | ; GFX942-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 0, killed $sgpr30_sgpr31, 12, killed $vgpr4_vgpr5, 8, killed $vgpr6_vgpr7, 0, 0, 0, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_inst_dependent_on_mfma_are_not_unpacked - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_ACCVGPR_READ_B32_e64 $agpr0, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec - ; GFX90a-NEXT: $vgpr6 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec - ; GFX90a-NEXT: $vgpr7 = V_ACCVGPR_READ_B32_e64 $agpr3, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 0, killed $sgpr30_sgpr31, 12, killed $vgpr4_vgpr5, 8, killed $vgpr6_vgpr7, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_inst_dependent_on_mfma_are_not_unpacked + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_ACCVGPR_READ_B32_e64 $agpr0, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec + ; GFX90A-NEXT: $vgpr6 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec + ; GFX90A-NEXT: $vgpr7 = V_ACCVGPR_READ_B32_e64 $agpr3, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 0, killed $sgpr30_sgpr31, 12, killed $vgpr4_vgpr5, 8, killed $vgpr6_vgpr7, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 @@ -950,25 +946,25 @@ body: | ; GFX942-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 8, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_mfma_def_using_instr_blocks_unpacking - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr6 = V_ACCVGPR_READ_B32_e64 $agpr3, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 8, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_mfma_def_using_instr_blocks_unpacking + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr6 = V_ACCVGPR_READ_B32_e64 $agpr3, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_FMA_F32 8, killed $sgpr30_sgpr31, 8, killed $vgpr4_vgpr5, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 @@ -1041,26 +1037,26 @@ body: | ; GFX942-NEXT: $vgpr17 = nofpexcept V_MUL_F32_e64 0, 1065353216, 0, killed $vgpr5, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_unpacking_with_imm_input - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 8, 1065353216, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_unpacking_with_imm_input + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 8, 1065353216, 8, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec S_WAITCNT 49279 @@ -1134,26 +1130,26 @@ body: | ; GFX942-NEXT: $vgpr17 = nofpexcept V_MUL_F32_e64 0, killed $sgpr31, 1, killed $vgpr5, 0, 0, implicit $mode, implicit $exec ; GFX942-NEXT: S_ENDPGM 0 ; - ; GFX90a-LABEL: name: test_neg_lo_hi_post_unpacking - ; GFX90a: liveins: $sgpr4_sgpr5 - ; GFX90a-NEXT: {{ $}} - ; GFX90a-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 - ; GFX90a-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 - ; GFX90a-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 - ; GFX90a-NEXT: S_WAITCNT 49279 - ; GFX90a-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 - ; GFX90a-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec - ; GFX90a-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 - ; GFX90a-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec - ; GFX90a-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec - ; GFX90a-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec - ; GFX90a-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 8, killed $sgpr30_sgpr31, 11, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec - ; GFX90a-NEXT: S_ENDPGM 0 + ; GFX90A-LABEL: name: test_neg_lo_hi_post_unpacking + ; GFX90A: liveins: $sgpr4_sgpr5 + ; GFX90A-NEXT: {{ $}} + ; GFX90A-NEXT: early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 + ; GFX90A-NEXT: renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: renamable $sgpr44_sgpr45_sgpr46_sgpr47 = S_LOAD_DWORDX4_IMM renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: renamable $sgpr48_sgpr49_sgpr50_sgpr51 = S_LOAD_DWORDX4_IMM renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7_sgpr8_sgpr9_sgpr10_sgpr11_sgpr12_sgpr13_sgpr14_sgpr15 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr42_sgpr43, 0, 0 + ; GFX90A-NEXT: early-clobber renamable $sgpr16_sgpr17_sgpr18_sgpr19_sgpr20_sgpr21_sgpr22_sgpr23_sgpr24_sgpr25_sgpr26_sgpr27_sgpr28_sgpr29_sgpr30_sgpr31 = S_LOAD_DWORDX16_IMM_ec killed renamable $sgpr40_sgpr41, 0, 0 + ; GFX90A-NEXT: S_WAITCNT 49279 + ; GFX90A-NEXT: $vgpr0_vgpr1 = V_MOV_B64_e32 $sgpr44_sgpr45, implicit $exec, implicit-def $vgpr0_vgpr1_vgpr2_vgpr3, implicit $sgpr44_sgpr45_sgpr46_sgpr47 + ; GFX90A-NEXT: $vgpr2_vgpr3 = V_MOV_B64_e32 killed $sgpr46_sgpr47, implicit $exec, implicit $sgpr44_sgpr45_sgpr46_sgpr47, implicit $exec + ; GFX90A-NEXT: $vgpr4_vgpr5 = V_MOV_B64_e32 $sgpr48_sgpr49, implicit $exec, implicit-def $vgpr4_vgpr5_vgpr6_vgpr7, implicit $sgpr48_sgpr49_sgpr50_sgpr51 + ; GFX90A-NEXT: $vgpr6_vgpr7 = V_MOV_B64_e32 killed $sgpr50_sgpr51, implicit $exec, implicit $sgpr48_sgpr49_sgpr50_sgpr51, implicit $exec + ; GFX90A-NEXT: early-clobber renamable $agpr0_agpr1_agpr2_agpr3_agpr4_agpr5_agpr6_agpr7_agpr8_agpr9_agpr10_agpr11_agpr12_agpr13_agpr14_agpr15 = V_MFMA_F32_32X32X16_F16_e64 killed $vgpr0_vgpr1_vgpr2_vgpr3, killed $vgpr4_vgpr5_vgpr6_vgpr7, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: $vgpr4 = V_MOV_B32_e32 killed $sgpr14, implicit $exec, implicit $exec + ; GFX90A-NEXT: $vgpr5 = V_MOV_B32_e32 killed $sgpr15, implicit $exec, implicit $exec + ; GFX90A-NEXT: renamable $vgpr16_vgpr17 = nofpexcept V_PK_MUL_F32 8, killed $sgpr30_sgpr31, 11, killed $vgpr4_vgpr5, 0, 0, 0, 0, 0, implicit $mode, implicit $exec + ; GFX90A-NEXT: S_ENDPGM 0 early-clobber renamable $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43 = S_LOAD_DWORDX8_IMM_ec killed renamable $sgpr4_sgpr5, 0, 0 renamable $vgpr18 = V_MOV_B32_e32 0, implicit $exec S_WAITCNT 49279 diff --git a/llvm/test/CodeGen/AMDGPU/v_sat_pk_u8_i16.ll b/llvm/test/CodeGen/AMDGPU/v_sat_pk_u8_i16.ll index 6b5bae06e5ef..c9b94e096fde 100644 --- a/llvm/test/CodeGen/AMDGPU/v_sat_pk_u8_i16.ll +++ b/llvm/test/CodeGen/AMDGPU/v_sat_pk_u8_i16.ll @@ -6,12 +6,12 @@ ; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=SDAG-GFX12,SDAG-GFX12-TRUE16 %s ; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=SDAG-GFX12,SDAG-GFX12-FAKE16 %s -; RUN: llc -mtriple=amdgcn -mcpu=fiji -global-isel < %s | FileCheck -check-prefixes=GISEL-VI %s -; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -global-isel < %s | FileCheck -check-prefixes=GISEL-GFX9 %s -; RUN: llc -mtriple=amdgcn -mcpu=gfx1101 -mattr=+real-true16 -global-isel < %s | FileCheck -check-prefixes=GFX11,GISEL-GFX11,GISEL-GFX11-TRUE16 %s -; RUN: llc -mtriple=amdgcn -mcpu=gfx1101 -mattr=-real-true16 -global-isel < %s | FileCheck -check-prefixes=GFX11,GISEL-GFX11,GISEL-GFX11-FAKE16 %s -; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 -global-isel < %s | FileCheck -check-prefixes=GISEL-GFX12,GISEL-GFX12-TRUE16 %s -; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -global-isel < %s | FileCheck -check-prefixes=GISEL-GFX12,GISEL-GFX12-FAKE16 %s +; RUN: llc -mtriple=amdgcn -mcpu=fiji -global-isel -new-reg-bank-select < %s | FileCheck -check-prefixes=GISEL-VI %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -global-isel -new-reg-bank-select < %s | FileCheck -check-prefixes=GISEL-GFX9 %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx1101 -mattr=+real-true16 -global-isel -new-reg-bank-select < %s | FileCheck -check-prefixes=GFX11,GISEL-GFX11,GISEL-GFX11-TRUE16 %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx1101 -mattr=-real-true16 -global-isel -new-reg-bank-select < %s | FileCheck -check-prefixes=GFX11,GISEL-GFX11,GISEL-GFX11-FAKE16 %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 -global-isel -new-reg-bank-select < %s | FileCheck -check-prefixes=GISEL-GFX12,GISEL-GFX12-TRUE16 %s +; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 -global-isel -new-reg-bank-select < %s | FileCheck -check-prefixes=GISEL-GFX12,GISEL-GFX12-FAKE16 %s ; <GFX9 has no V_SAT_PK, GFX9+ has V_SAT_PK, GFX11 has V_SAT_PK with t16 diff --git a/llvm/test/CodeGen/AMDGPU/vector-reduce-smax.ll b/llvm/test/CodeGen/AMDGPU/vector-reduce-smax.ll index b5d9d00c4804..8d0e00383d69 100644 --- a/llvm/test/CodeGen/AMDGPU/vector-reduce-smax.ll +++ b/llvm/test/CodeGen/AMDGPU/vector-reduce-smax.ll @@ -1,21 +1,21 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-SDAG,GFX11-SDAG-TRUE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-SDAG,GFX11-SDAG-FAKE16 %s ; FIXME-TRUE16. enable gisel -; XUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s +; XUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG,GFX12-SDAG-TRUE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG,GFX12-SDAG-FAKE16 %s -; XUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s +; XUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s define i8 @test_vector_reduce_smax_v2i8(<2 x i8> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smax_v2i8: @@ -1632,6 +1632,7 @@ entry: ret i8 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_smax_v2i16(<2 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smax_v2i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1678,7 +1679,7 @@ define i16 @test_vector_reduce_smax_v2i16(<2 x i16> %v) { ; GFX9-GISEL-LABEL: test_vector_reduce_smax_v2i16: ; GFX9-GISEL: ; %bb.0: ; %entry ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1692,7 +1693,7 @@ define i16 @test_vector_reduce_smax_v2i16(<2 x i16> %v) { ; GFX10-GISEL-LABEL: test_vector_reduce_smax_v2i16: ; GFX10-GISEL: ; %bb.0: ; %entry ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1713,7 +1714,7 @@ define i16 @test_vector_reduce_smax_v2i16(<2 x i16> %v) { ; GFX11-GISEL-LABEL: test_vector_reduce_smax_v2i16: ; GFX11-GISEL: ; %bb.0: ; %entry ; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1747,7 +1748,7 @@ define i16 @test_vector_reduce_smax_v2i16(<2 x i16> %v) { ; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0 ; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0 ; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1900,6 +1901,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_smax_v4i16(<4 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smax_v4i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1961,7 +1963,7 @@ define i16 @test_vector_reduce_smax_v4i16(<4 x i16> %v) { ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1977,7 +1979,7 @@ define i16 @test_vector_reduce_smax_v4i16(<4 x i16> %v) { ; GFX10-GISEL: ; %bb.0: ; %entry ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2003,7 +2005,7 @@ define i16 @test_vector_reduce_smax_v4i16(<4 x i16> %v) { ; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2041,7 +2043,7 @@ define i16 @test_vector_reduce_smax_v4i16(<4 x i16> %v) { ; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0 ; GFX12-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] entry: @@ -2049,6 +2051,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_smax_v8i16(<8 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smax_v8i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -2139,7 +2142,7 @@ define i16 @test_vector_reduce_smax_v8i16(<8 x i16> %v) { ; GFX9-GISEL-NEXT: s_nop 0 ; GFX9-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2159,7 +2162,7 @@ define i16 @test_vector_reduce_smax_v8i16(<8 x i16> %v) { ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v2 ; GFX10-GISEL-NEXT: v_pk_max_i16 v1, v1, v3 ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2192,7 +2195,7 @@ define i16 @test_vector_reduce_smax_v8i16(<8 x i16> %v) { ; GFX11-GISEL-NEXT: v_pk_max_i16 v1, v1, v3 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2238,7 +2241,7 @@ define i16 @test_vector_reduce_smax_v8i16(<8 x i16> %v) { ; GFX12-GISEL-NEXT: v_pk_max_i16 v1, v1, v3 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2247,6 +2250,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_smax_v16i16(<16 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smax_v16i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -2391,7 +2395,7 @@ define i16 @test_vector_reduce_smax_v16i16(<16 x i16> %v) { ; GFX9-GISEL-NEXT: s_nop 0 ; GFX9-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2419,7 +2423,7 @@ define i16 @test_vector_reduce_smax_v16i16(<16 x i16> %v) { ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v2 ; GFX10-GISEL-NEXT: v_pk_max_i16 v1, v1, v3 ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2467,7 +2471,7 @@ define i16 @test_vector_reduce_smax_v16i16(<16 x i16> %v) { ; GFX11-GISEL-NEXT: v_pk_max_i16 v1, v1, v3 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2528,7 +2532,7 @@ define i16 @test_vector_reduce_smax_v16i16(<16 x i16> %v) { ; GFX12-GISEL-NEXT: v_pk_max_i16 v1, v1, v3 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] diff --git a/llvm/test/CodeGen/AMDGPU/vector-reduce-smin.ll b/llvm/test/CodeGen/AMDGPU/vector-reduce-smin.ll index 2a989ecd2eba..f15ecf014ab0 100644 --- a/llvm/test/CodeGen/AMDGPU/vector-reduce-smin.ll +++ b/llvm/test/CodeGen/AMDGPU/vector-reduce-smin.ll @@ -1,21 +1,21 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-SDAG,GFX11-SDAG-TRUE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-SDAG,GFX11-SDAG-FAKE16 %s ; FIXME-TRUE16. enable gisel -; XUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s +; XUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG,GFX12-SDAG-TRUE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG,GFX12-SDAG-FAKE16 %s -; XUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s +; XUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s define i8 @test_vector_reduce_smin_v2i8(<2 x i8> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smin_v2i8: @@ -1632,6 +1632,7 @@ entry: ret i8 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_smin_v2i16(<2 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smin_v2i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1678,7 +1679,7 @@ define i16 @test_vector_reduce_smin_v2i16(<2 x i16> %v) { ; GFX9-GISEL-LABEL: test_vector_reduce_smin_v2i16: ; GFX9-GISEL: ; %bb.0: ; %entry ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1692,7 +1693,7 @@ define i16 @test_vector_reduce_smin_v2i16(<2 x i16> %v) { ; GFX10-GISEL-LABEL: test_vector_reduce_smin_v2i16: ; GFX10-GISEL: ; %bb.0: ; %entry ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1713,7 +1714,7 @@ define i16 @test_vector_reduce_smin_v2i16(<2 x i16> %v) { ; GFX11-GISEL-LABEL: test_vector_reduce_smin_v2i16: ; GFX11-GISEL: ; %bb.0: ; %entry ; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1747,7 +1748,7 @@ define i16 @test_vector_reduce_smin_v2i16(<2 x i16> %v) { ; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0 ; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0 ; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1900,6 +1901,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_smin_v4i16(<4 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smin_v4i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1961,7 +1963,7 @@ define i16 @test_vector_reduce_smin_v4i16(<4 x i16> %v) { ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1977,7 +1979,7 @@ define i16 @test_vector_reduce_smin_v4i16(<4 x i16> %v) { ; GFX10-GISEL: ; %bb.0: ; %entry ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2003,7 +2005,7 @@ define i16 @test_vector_reduce_smin_v4i16(<4 x i16> %v) { ; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2041,7 +2043,7 @@ define i16 @test_vector_reduce_smin_v4i16(<4 x i16> %v) { ; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0 ; GFX12-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] entry: @@ -2049,6 +2051,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_smin_v8i16(<8 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_smin_v8i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -2139,7 +2142,7 @@ define i16 @test_vector_reduce_smin_v8i16(<8 x i16> %v) { ; GFX9-GISEL-NEXT: s_nop 0 ; GFX9-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2159,7 +2162,7 @@ define i16 @test_vector_reduce_smin_v8i16(<8 x i16> %v) { ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v2 ; GFX10-GISEL-NEXT: v_pk_min_i16 v1, v1, v3 ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2192,7 +2195,7 @@ define i16 @test_vector_reduce_smin_v8i16(<8 x i16> %v) { ; GFX11-GISEL-NEXT: v_pk_min_i16 v1, v1, v3 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2238,7 +2241,7 @@ define i16 @test_vector_reduce_smin_v8i16(<8 x i16> %v) { ; GFX12-GISEL-NEXT: v_pk_min_i16 v1, v1, v3 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2391,7 +2394,7 @@ define i16 @test_vector_reduce_smin_v16i16(<16 x i16> %v) { ; GFX9-GISEL-NEXT: s_nop 0 ; GFX9-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2419,7 +2422,7 @@ define i16 @test_vector_reduce_smin_v16i16(<16 x i16> %v) { ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v2 ; GFX10-GISEL-NEXT: v_pk_min_i16 v1, v1, v3 ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2467,7 +2470,7 @@ define i16 @test_vector_reduce_smin_v16i16(<16 x i16> %v) { ; GFX11-GISEL-NEXT: v_pk_min_i16 v1, v1, v3 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2528,7 +2531,7 @@ define i16 @test_vector_reduce_smin_v16i16(<16 x i16> %v) { ; GFX12-GISEL-NEXT: v_pk_min_i16 v1, v1, v3 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_i16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] diff --git a/llvm/test/CodeGen/AMDGPU/vector-reduce-umax.ll b/llvm/test/CodeGen/AMDGPU/vector-reduce-umax.ll index 69fd58aadfbc..e62165cb933c 100644 --- a/llvm/test/CodeGen/AMDGPU/vector-reduce-umax.ll +++ b/llvm/test/CodeGen/AMDGPU/vector-reduce-umax.ll @@ -1,21 +1,21 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-SDAG,GFX11-SDAG-TRUE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-SDAG,GFX11-SDAG-FAKE16 %s ; FIXME-TRUE16. enable gisel -; XUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s +; XUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG,GFX12-SDAG-TRUE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG,GFX12-SDAG-FAKE16 %s -; XUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s +; XUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s define i8 @test_vector_reduce_umax_v2i8(<2 x i8> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umax_v2i8: @@ -1525,6 +1525,7 @@ entry: ret i8 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_umax_v2i16(<2 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umax_v2i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1569,7 +1570,7 @@ define i16 @test_vector_reduce_umax_v2i16(<2 x i16> %v) { ; GFX9-GISEL-LABEL: test_vector_reduce_umax_v2i16: ; GFX9-GISEL: ; %bb.0: ; %entry ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1583,7 +1584,7 @@ define i16 @test_vector_reduce_umax_v2i16(<2 x i16> %v) { ; GFX10-GISEL-LABEL: test_vector_reduce_umax_v2i16: ; GFX10-GISEL: ; %bb.0: ; %entry ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1604,7 +1605,7 @@ define i16 @test_vector_reduce_umax_v2i16(<2 x i16> %v) { ; GFX11-GISEL-LABEL: test_vector_reduce_umax_v2i16: ; GFX11-GISEL: ; %bb.0: ; %entry ; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1638,7 +1639,7 @@ define i16 @test_vector_reduce_umax_v2i16(<2 x i16> %v) { ; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0 ; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0 ; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1782,6 +1783,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_umax_v4i16(<4 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umax_v4i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1841,7 +1843,7 @@ define i16 @test_vector_reduce_umax_v4i16(<4 x i16> %v) { ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1857,7 +1859,7 @@ define i16 @test_vector_reduce_umax_v4i16(<4 x i16> %v) { ; GFX10-GISEL: ; %bb.0: ; %entry ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1883,7 +1885,7 @@ define i16 @test_vector_reduce_umax_v4i16(<4 x i16> %v) { ; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1921,7 +1923,7 @@ define i16 @test_vector_reduce_umax_v4i16(<4 x i16> %v) { ; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0 ; GFX12-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] entry: @@ -1929,6 +1931,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_umax_v8i16(<8 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umax_v8i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -2017,7 +2020,7 @@ define i16 @test_vector_reduce_umax_v8i16(<8 x i16> %v) { ; GFX9-GISEL-NEXT: s_nop 0 ; GFX9-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2037,7 +2040,7 @@ define i16 @test_vector_reduce_umax_v8i16(<8 x i16> %v) { ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v2 ; GFX10-GISEL-NEXT: v_pk_max_u16 v1, v1, v3 ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2070,7 +2073,7 @@ define i16 @test_vector_reduce_umax_v8i16(<8 x i16> %v) { ; GFX11-GISEL-NEXT: v_pk_max_u16 v1, v1, v3 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2116,7 +2119,7 @@ define i16 @test_vector_reduce_umax_v8i16(<8 x i16> %v) { ; GFX12-GISEL-NEXT: v_pk_max_u16 v1, v1, v3 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2125,6 +2128,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_umax_v16i16(<16 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umax_v16i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -2267,7 +2271,7 @@ define i16 @test_vector_reduce_umax_v16i16(<16 x i16> %v) { ; GFX9-GISEL-NEXT: s_nop 0 ; GFX9-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2295,7 +2299,7 @@ define i16 @test_vector_reduce_umax_v16i16(<16 x i16> %v) { ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v2 ; GFX10-GISEL-NEXT: v_pk_max_u16 v1, v1, v3 ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2343,7 +2347,7 @@ define i16 @test_vector_reduce_umax_v16i16(<16 x i16> %v) { ; GFX11-GISEL-NEXT: v_pk_max_u16 v1, v1, v3 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2404,7 +2408,7 @@ define i16 @test_vector_reduce_umax_v16i16(<16 x i16> %v) { ; GFX12-GISEL-NEXT: v_pk_max_u16 v1, v1, v3 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_max_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] diff --git a/llvm/test/CodeGen/AMDGPU/vector-reduce-umin.ll b/llvm/test/CodeGen/AMDGPU/vector-reduce-umin.ll index 1d3b42ee43b0..83ecaaa7e084 100644 --- a/llvm/test/CodeGen/AMDGPU/vector-reduce-umin.ll +++ b/llvm/test/CodeGen/AMDGPU/vector-reduce-umin.ll @@ -1,21 +1,21 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx700 < %s | FileCheck -check-prefixes=GFX7,GFX7-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx801 < %s | FileCheck -check-prefixes=GFX8,GFX8-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx942 < %s | FileCheck -check-prefixes=GFX9,GFX9-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-SDAG %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1010 < %s | FileCheck -check-prefixes=GFX10,GFX10-GISEL %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-SDAG,GFX11-SDAG-TRUE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-SDAG,GFX11-SDAG-FAKE16 %s ; FIXME-TRUE16. enable gisel -; XUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s +; XUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-TRUE16 %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX11,GFX11-GISEL,GFX11-GISEL-FAKE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG,GFX12-SDAG-TRUE16 %s ; RUN: llc -global-isel=0 -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG,GFX12-SDAG-FAKE16 %s -; XUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s -; RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s +; XUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1200 -mattr=+real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-TRUE16 %s +; RUN: llc -global-isel=1 -new-reg-bank-select -mtriple=amdgcn -mcpu=gfx1200 -mattr=-real-true16 < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL,GFX12-GISEL-FAKE16 %s define i8 @test_vector_reduce_umin_v2i8(<2 x i8> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umin_v2i8: @@ -1271,6 +1271,7 @@ entry: ret i8 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_umin_v2i16(<2 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umin_v2i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1312,7 +1313,7 @@ define i16 @test_vector_reduce_umin_v2i16(<2 x i16> %v) { ; GFX9-GISEL-LABEL: test_vector_reduce_umin_v2i16: ; GFX9-GISEL: ; %bb.0: ; %entry ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1326,7 +1327,7 @@ define i16 @test_vector_reduce_umin_v2i16(<2 x i16> %v) { ; GFX10-GISEL-LABEL: test_vector_reduce_umin_v2i16: ; GFX10-GISEL: ; %bb.0: ; %entry ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1347,7 +1348,7 @@ define i16 @test_vector_reduce_umin_v2i16(<2 x i16> %v) { ; GFX11-GISEL-LABEL: test_vector_reduce_umin_v2i16: ; GFX11-GISEL: ; %bb.0: ; %entry ; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1381,7 +1382,7 @@ define i16 @test_vector_reduce_umin_v2i16(<2 x i16> %v) { ; GFX12-GISEL-NEXT: s_wait_samplecnt 0x0 ; GFX12-GISEL-NEXT: s_wait_bvhcnt 0x0 ; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1527,6 +1528,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_umin_v4i16(<4 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umin_v4i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1583,7 +1585,7 @@ define i16 @test_vector_reduce_umin_v4i16(<4 x i16> %v) { ; GFX9-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX9-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1599,7 +1601,7 @@ define i16 @test_vector_reduce_umin_v4i16(<4 x i16> %v) { ; GFX10-GISEL: ; %bb.0: ; %entry ; GFX10-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1625,7 +1627,7 @@ define i16 @test_vector_reduce_umin_v4i16(<4 x i16> %v) { ; GFX11-GISEL-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX11-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1663,7 +1665,7 @@ define i16 @test_vector_reduce_umin_v4i16(<4 x i16> %v) { ; GFX12-GISEL-NEXT: s_wait_kmcnt 0x0 ; GFX12-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] entry: @@ -1671,6 +1673,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_umin_v8i16(<8 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umin_v8i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -1756,7 +1759,7 @@ define i16 @test_vector_reduce_umin_v8i16(<8 x i16> %v) { ; GFX9-GISEL-NEXT: s_nop 0 ; GFX9-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1776,7 +1779,7 @@ define i16 @test_vector_reduce_umin_v8i16(<8 x i16> %v) { ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v2 ; GFX10-GISEL-NEXT: v_pk_min_u16 v1, v1, v3 ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -1809,7 +1812,7 @@ define i16 @test_vector_reduce_umin_v8i16(<8 x i16> %v) { ; GFX11-GISEL-NEXT: v_pk_min_u16 v1, v1, v3 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1855,7 +1858,7 @@ define i16 @test_vector_reduce_umin_v8i16(<8 x i16> %v) { ; GFX12-GISEL-NEXT: v_pk_min_u16 v1, v1, v3 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -1864,6 +1867,7 @@ entry: ret i16 %res } +; FIXME: With -new-reg-bank-select, v_alignbit_b32 is regression. Need pattern to look through COPY. define i16 @test_vector_reduce_umin_v16i16(<16 x i16> %v) { ; GFX7-SDAG-LABEL: test_vector_reduce_umin_v16i16: ; GFX7-SDAG: ; %bb.0: ; %entry @@ -2003,7 +2007,7 @@ define i16 @test_vector_reduce_umin_v16i16(<16 x i16> %v) { ; GFX9-GISEL-NEXT: s_nop 0 ; GFX9-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_nop 0 -; GFX9-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX9-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX9-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX9-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2031,7 +2035,7 @@ define i16 @test_vector_reduce_umin_v16i16(<16 x i16> %v) { ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v2 ; GFX10-GISEL-NEXT: v_pk_min_u16 v1, v1, v3 ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 -; GFX10-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX10-GISEL-NEXT: v_alignbit_b32 v1, s4, v0, 16 ; GFX10-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX10-GISEL-NEXT: s_setpc_b64 s[30:31] ; @@ -2079,7 +2083,7 @@ define i16 @test_vector_reduce_umin_v16i16(<16 x i16> %v) { ; GFX11-GISEL-NEXT: v_pk_min_u16 v1, v1, v3 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 -; GFX11-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX11-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX11-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX11-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX11-GISEL-NEXT: s_setpc_b64 s[30:31] @@ -2140,7 +2144,7 @@ define i16 @test_vector_reduce_umin_v16i16(<16 x i16> %v) { ; GFX12-GISEL-NEXT: v_pk_min_u16 v1, v1, v3 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 -; GFX12-GISEL-NEXT: v_lshrrev_b32_e32 v1, 16, v0 +; GFX12-GISEL-NEXT: v_alignbit_b32 v1, s0, v0, 16 ; GFX12-GISEL-NEXT: s_delay_alu instid0(VALU_DEP_1) ; GFX12-GISEL-NEXT: v_pk_min_u16 v0, v0, v1 ; GFX12-GISEL-NEXT: s_setpc_b64 s[30:31] |