summaryrefslogtreecommitdiff
path: root/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll')
-rw-r--r--llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll3565
1 files changed, 1809 insertions, 1756 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
index 35d135b12396..ede44e738fe0 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgcn.bitcast.512bit.ll
@@ -5873,287 +5873,298 @@ define <16 x i32> @bitcast_v32bf16_to_v16i32(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-LABEL: bitcast_v32bf16_to_v16i32:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
+; GFX11-TRUE16-NEXT: ; implicit-def: $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)
-; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_2
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v15
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v15, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v15
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v15.h
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v15 :: v_dual_mov_b32 v30, v14
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v13 :: v_dual_mov_b32 v28, v12
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v11 :: v_dual_mov_b32 v26, v10
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v9 :: v_dual_mov_b32 v24, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v7 :: v_dual_mov_b32 v22, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v5 :: v_dual_mov_b32 v20, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v3 :: v_dual_mov_b32 v18, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v1 :: v_dual_mov_b32 v16, v0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: .LBB22_2: ; %Flow
+; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB22_4
+; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v14
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v15
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v17, 16, 1
; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v16, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v16
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v16, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v19, v21 :: v_dual_and_b32 v17, 0xffff0000, v14
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_lshlrev_b32 v14, 16, v14
-; GFX11-TRUE16-NEXT: v_bfi_b32 v15, 0xffff, v15, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v17, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, 0x400000, v14
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v19, v22 :: v_dual_and_b32 v15, 0xffff0000, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 0x40c00000, v15 :: v_dual_lshlrev_b32 v14, 16, v14
+; GFX11-TRUE16-NEXT: v_add3_u32 v19, v23, v17, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v15
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v31, v20, v21
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v14
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v23, v17, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v25, v14, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v17
+; GFX11-TRUE16-NEXT: v_add3_u32 v22, v24, v14, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v22, v26, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v22, v25, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v20, v24 :: v_dual_and_b32 v18, 0xffff0000, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_lshlrev_b32 v13, 16, v13
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v14.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v18
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 0x40c00000, v18 :: v_dual_cndmask_b32 v30, v19, v16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v14.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v13, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_add3_u32 v17, v18, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v14, v17
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v19, v18, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v15
; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v13, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v13, v16, v20 :: v_dual_and_b32 v20, 0xffff0000, v11
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_lshlrev_b32 v12, 16, v12
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v16, v18, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v29, v17, v19
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v11
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v11, 16, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v13, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v12, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v12
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v19, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v14, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v12, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v18, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v11, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v20, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v11, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v11
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v15, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v15, v17, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v13, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v16, v18, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v11
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v12, v17
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v12.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v15, v13, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v11, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v11, v14, v16 :: v_dual_and_b32 v12, 0xffff0000, v10
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_cndmask_b32 v27, v15, v17
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v18, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v11, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v20, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_lshlrev_b32 v10, 16, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v10, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v9
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v9, 16, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v10
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v10, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v16, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v9, 16, 1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v8
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v9, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v10, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v13, v15, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v11
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v14, v16, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v9
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v17, v22 :: v_dual_and_b32 v22, 0xffff0000, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v10.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v11, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v9, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v9, v12, v14 :: v_dual_and_b32 v10, 0xffff0000, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v19, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v9, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v7 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_cndmask_b32 v25, v13, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v9.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v7
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v7, 16, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v8
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v8, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v8
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v10, v16
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v23, 0xffff0000, v4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v10, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v8, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v11, v13, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v24, v12, v14, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v8.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v7, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v10, v12 :: v_dual_and_b32 v8, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v20, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v7, v16, v20
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v6, 16, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v6, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_add_f32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_cndmask_b32 v23, v11, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v5
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v5, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v7, v16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v19, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v5
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v20
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v22, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v6, v17
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v4, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v23
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v4
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v4, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v18, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v9, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_add_f32 v4, 0x40c00000, v4
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v6, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v2
-; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v5, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v3
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v17, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v2
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v16, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v3, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v3, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_cndmask_b32 v4, v7, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v3, 0x40c00000, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v23, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v17, v22 :: v_dual_and_b32 v22, 0xffff0000, v1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v2, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v16
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v22 :: v_dual_cndmask_b32 v18, v18, v20
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v0
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v4.h
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v8, v3, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v4, v9, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v10, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v3.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, 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
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v2, v4, v5
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
+; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v2.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v3, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v17
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v1, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: 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_bfe_u32 v9, v0, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v22, v1, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v23, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v22, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_bfe_u32 v26, v0, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v26, v0, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v20.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v19, v24, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v19
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v22, v25, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v20, v0
-; GFX11-TRUE16-NEXT: .LBB22_2: ; %end
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v9, v0, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v4, v7, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v1.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.h
+; GFX11-TRUE16-NEXT: .LBB22_4: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, v16 :: v_dual_mov_b32 v1, v17
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, v18 :: v_dual_mov_b32 v3, v19
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v20 :: v_dual_mov_b32 v5, v21
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, v22 :: v_dual_mov_b32 v7, v23
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, v24 :: v_dual_mov_b32 v9, v25
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, v26 :: v_dual_mov_b32 v11, v27
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, v28 :: v_dual_mov_b32 v13, v29
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v30 :: v_dual_mov_b32 v15, v31
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v32bf16_to_v16i32:
@@ -20624,287 +20635,298 @@ define <16 x float> @bitcast_v32bf16_to_v16f32(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-LABEL: bitcast_v32bf16_to_v16f32:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
+; GFX11-TRUE16-NEXT: ; implicit-def: $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)
-; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB46_2
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v15
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v15, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v15
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v15.h
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v15 :: v_dual_mov_b32 v30, v14
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v13 :: v_dual_mov_b32 v28, v12
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v11 :: v_dual_mov_b32 v26, v10
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v9 :: v_dual_mov_b32 v24, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v7 :: v_dual_mov_b32 v22, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v5 :: v_dual_mov_b32 v20, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v3 :: v_dual_mov_b32 v18, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v1 :: v_dual_mov_b32 v16, v0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: .LBB46_2: ; %Flow
+; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB46_4
+; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v14
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v15
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v17, 16, 1
; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v16, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v16
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v16, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v19, v21 :: v_dual_and_b32 v17, 0xffff0000, v14
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_lshlrev_b32 v14, 16, v14
-; GFX11-TRUE16-NEXT: v_bfi_b32 v15, 0xffff, v15, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v17, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, 0x400000, v14
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v19, v22 :: v_dual_and_b32 v15, 0xffff0000, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 0x40c00000, v15 :: v_dual_lshlrev_b32 v14, 16, v14
+; GFX11-TRUE16-NEXT: v_add3_u32 v19, v23, v17, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v15
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v31, v20, v21
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v14
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v23, v17, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v25, v14, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v17
+; GFX11-TRUE16-NEXT: v_add3_u32 v22, v24, v14, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v22, v26, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v22, v25, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v20, v24 :: v_dual_and_b32 v18, 0xffff0000, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_lshlrev_b32 v13, 16, v13
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v14.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v18
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 0x40c00000, v18 :: v_dual_cndmask_b32 v30, v19, v16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v14.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v13, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_add3_u32 v17, v18, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v14, v17
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v19, v18, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v15
; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v13, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v13, v16, v20 :: v_dual_and_b32 v20, 0xffff0000, v11
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_lshlrev_b32 v12, 16, v12
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v16, v18, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v29, v17, v19
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v11
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v11, 16, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v13, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v12, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v12
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v19, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v14, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v12, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v18, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v11, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v20, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v11, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v11
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v15, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v15, v17, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v13, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v16, v18, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v11
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v12, v17
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v12.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v15, v13, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v11, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v11, v14, v16 :: v_dual_and_b32 v12, 0xffff0000, v10
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_cndmask_b32 v27, v15, v17
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v18, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v11, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v20, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_lshlrev_b32 v10, 16, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v10, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v9
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v9, 16, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v10
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v10, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v16, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v9, 16, 1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v8
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v9, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v10, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v13, v15, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v11
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v14, v16, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v9
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v17, v22 :: v_dual_and_b32 v22, 0xffff0000, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v10.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v11, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v9, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v9, v12, v14 :: v_dual_and_b32 v10, 0xffff0000, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v19, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v9, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v7 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_cndmask_b32 v25, v13, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v9.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v7
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v7, 16, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v8
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v8, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v8
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v10, v16
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v23, 0xffff0000, v4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v10, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v8, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v11, v13, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v24, v12, v14, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v8.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v7, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v10, v12 :: v_dual_and_b32 v8, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v20, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v7, v16, v20
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v6, 16, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v6, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_add_f32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_cndmask_b32 v23, v11, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v5
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v5, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v7, v16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v19, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v5
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v20
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v22, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v6, v17
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v4, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v23
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v4
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v4, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v18, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v9, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_add_f32 v4, 0x40c00000, v4
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v6, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v2
-; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v5, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v3
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v17, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v2
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v16, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v3, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v3, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_cndmask_b32 v4, v7, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v3, 0x40c00000, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v23, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v17, v22 :: v_dual_and_b32 v22, 0xffff0000, v1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v2, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v16
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v22 :: v_dual_cndmask_b32 v18, v18, v20
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v0
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v4.h
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v8, v3, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v4, v9, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v10, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v3.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, 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
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v2, v4, v5
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
+; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v2.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v3, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v17
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v1, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: 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_bfe_u32 v9, v0, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v22, v1, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v23, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v22, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_bfe_u32 v26, v0, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v26, v0, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v20.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v19, v24, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v19
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v22, v25, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v20, v0
-; GFX11-TRUE16-NEXT: .LBB46_2: ; %end
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v9, v0, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v4, v7, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v1.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.h
+; GFX11-TRUE16-NEXT: .LBB46_4: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, v16 :: v_dual_mov_b32 v1, v17
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, v18 :: v_dual_mov_b32 v3, v19
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v20 :: v_dual_mov_b32 v5, v21
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, v22 :: v_dual_mov_b32 v7, v23
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, v24 :: v_dual_mov_b32 v9, v25
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, v26 :: v_dual_mov_b32 v11, v27
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, v28 :: v_dual_mov_b32 v13, v29
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v30 :: v_dual_mov_b32 v15, v31
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v32bf16_to_v16f32:
@@ -34883,287 +34905,298 @@ define <8 x i64> @bitcast_v32bf16_to_v8i64(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-LABEL: bitcast_v32bf16_to_v8i64:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
+; GFX11-TRUE16-NEXT: ; implicit-def: $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)
-; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB66_2
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v15
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v15, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v15
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v15.h
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v15 :: v_dual_mov_b32 v30, v14
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v13 :: v_dual_mov_b32 v28, v12
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v11 :: v_dual_mov_b32 v26, v10
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v9 :: v_dual_mov_b32 v24, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v7 :: v_dual_mov_b32 v22, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v5 :: v_dual_mov_b32 v20, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v3 :: v_dual_mov_b32 v18, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v1 :: v_dual_mov_b32 v16, v0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: .LBB66_2: ; %Flow
+; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB66_4
+; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v14
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v15
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v17, 16, 1
; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v16, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v16
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v16, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v19, v21 :: v_dual_and_b32 v17, 0xffff0000, v14
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_lshlrev_b32 v14, 16, v14
-; GFX11-TRUE16-NEXT: v_bfi_b32 v15, 0xffff, v15, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v17, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, 0x400000, v14
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v19, v22 :: v_dual_and_b32 v15, 0xffff0000, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 0x40c00000, v15 :: v_dual_lshlrev_b32 v14, 16, v14
+; GFX11-TRUE16-NEXT: v_add3_u32 v19, v23, v17, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v15
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v31, v20, v21
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v14
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v23, v17, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v25, v14, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v17
+; GFX11-TRUE16-NEXT: v_add3_u32 v22, v24, v14, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v22, v26, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v22, v25, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v20, v24 :: v_dual_and_b32 v18, 0xffff0000, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_lshlrev_b32 v13, 16, v13
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v14.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v18
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 0x40c00000, v18 :: v_dual_cndmask_b32 v30, v19, v16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v14.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v13, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_add3_u32 v17, v18, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v14, v17
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v19, v18, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v15
; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v13, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v13, v16, v20 :: v_dual_and_b32 v20, 0xffff0000, v11
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_lshlrev_b32 v12, 16, v12
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v16, v18, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v29, v17, v19
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v11
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v11, 16, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v13, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v12, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v12
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v19, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v14, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v12, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v18, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v11, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v20, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v11, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v11
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v15, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v15, v17, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v13, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v16, v18, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v11
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v12, v17
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v12.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v15, v13, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v11, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v11, v14, v16 :: v_dual_and_b32 v12, 0xffff0000, v10
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_cndmask_b32 v27, v15, v17
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v18, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v11, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v20, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_lshlrev_b32 v10, 16, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v10, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v9
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v9, 16, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v10
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v10, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v16, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v9, 16, 1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v8
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v9, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v10, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v13, v15, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v11
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v14, v16, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v9
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v17, v22 :: v_dual_and_b32 v22, 0xffff0000, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v10.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v11, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v9, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v9, v12, v14 :: v_dual_and_b32 v10, 0xffff0000, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v19, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v9, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v7 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_cndmask_b32 v25, v13, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v9.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v7
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v7, 16, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v8
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v8, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v8
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v10, v16
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v23, 0xffff0000, v4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v10, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v8, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v11, v13, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v24, v12, v14, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v8.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v7, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v10, v12 :: v_dual_and_b32 v8, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v20, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v7, v16, v20
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v6, 16, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v6, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_add_f32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_cndmask_b32 v23, v11, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v5
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v5, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v7, v16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v19, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v5
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v20
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v22, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v6, v17
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v4, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v23
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v4
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v4, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v18, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v9, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_add_f32 v4, 0x40c00000, v4
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v6, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v2
-; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v5, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v3
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v17, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v2
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v16, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v3, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v3, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_cndmask_b32 v4, v7, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v3, 0x40c00000, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v23, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v17, v22 :: v_dual_and_b32 v22, 0xffff0000, v1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v2, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v16
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v22 :: v_dual_cndmask_b32 v18, v18, v20
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v0
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v4.h
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v8, v3, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v4, v9, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v10, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v3.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, 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
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v2, v4, v5
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
+; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v2.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v3, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v17
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v1, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: 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_bfe_u32 v9, v0, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v22, v1, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v23, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v22, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_bfe_u32 v26, v0, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v26, v0, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v20.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v19, v24, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v19
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v22, v25, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v20, v0
-; GFX11-TRUE16-NEXT: .LBB66_2: ; %end
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v9, v0, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v4, v7, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v1.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.h
+; GFX11-TRUE16-NEXT: .LBB66_4: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, v16 :: v_dual_mov_b32 v1, v17
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, v18 :: v_dual_mov_b32 v3, v19
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v20 :: v_dual_mov_b32 v5, v21
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, v22 :: v_dual_mov_b32 v7, v23
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, v24 :: v_dual_mov_b32 v9, v25
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, v26 :: v_dual_mov_b32 v11, v27
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, v28 :: v_dual_mov_b32 v13, v29
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v30 :: v_dual_mov_b32 v15, v31
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v32bf16_to_v8i64:
@@ -48217,287 +48250,298 @@ define <8 x double> @bitcast_v32bf16_to_v8f64(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-LABEL: bitcast_v32bf16_to_v8f64:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; GFX11-TRUE16-NEXT: s_mov_b32 s0, exec_lo
-; GFX11-TRUE16-NEXT: v_cmpx_ne_u32_e32 0, v16
-; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
+; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v16
+; GFX11-TRUE16-NEXT: ; implicit-def: $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)
-; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB82_2
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v15
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v15
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v15, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v15
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v15, v20, v22 :: v_dual_add_f32 v16, 0x40c00000, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v15.h
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.false
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v31, v15 :: v_dual_mov_b32 v30, v14
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v29, v13 :: v_dual_mov_b32 v28, v12
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v27, v11 :: v_dual_mov_b32 v26, v10
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v25, v9 :: v_dual_mov_b32 v24, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v23, v7 :: v_dual_mov_b32 v22, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v21, v5 :: v_dual_mov_b32 v20, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v19, v3 :: v_dual_mov_b32 v18, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v17, v1 :: v_dual_mov_b32 v16, v0
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr0
+; GFX11-TRUE16-NEXT: .LBB82_2: ; %Flow
+; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB82_4
+; GFX11-TRUE16-NEXT: ; %bb.3: ; %cmp.true
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v14
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v15
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_add_f32 v16, 0x40c00000, v16
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v17, 16, 1
; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v16, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v16
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v16
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v16, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v19, v21 :: v_dual_and_b32 v17, 0xffff0000, v14
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_lshlrev_b32 v14, 16, v14
-; GFX11-TRUE16-NEXT: v_bfi_b32 v15, 0xffff, v15, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v17, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v26, 0x400000, v14
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v19, v22 :: v_dual_and_b32 v15, 0xffff0000, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 0x40c00000, v15 :: v_dual_lshlrev_b32 v14, 16, v14
+; GFX11-TRUE16-NEXT: v_add3_u32 v19, v23, v17, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v15
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v31, v20, v21
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v14
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v23, v17, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v25, v14, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.l, v16.h
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v17
+; GFX11-TRUE16-NEXT: v_add3_u32 v22, v24, v14, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v22, v26, vcc_lo
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v22, v25, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v20, v24 :: v_dual_and_b32 v18, 0xffff0000, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_lshlrev_b32 v13, 16, v13
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v14.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v18
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v15, 0x40c00000, v18 :: v_dual_cndmask_b32 v30, v19, v16
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.l, v14.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v12
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v12
; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v13, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_add3_u32 v17, v18, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v13
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v14, v17
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v19, v18, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v15
; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v13, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v13, v16, v20 :: v_dual_and_b32 v20, 0xffff0000, v11
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v19, 0x40c00000, v19 :: v_dual_lshlrev_b32 v12, 16, v12
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v16, v18, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_cndmask_b32 v29, v17, v19
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.l, v13.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v11
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_lshlrev_b32 v11, 16, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v18, 0x400000, v14
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v13, v16
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v12, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v12
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v19, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v14, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v12, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v18, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v11, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v20, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v11, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v11
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v15, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v15, v17, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v13, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v16, v18, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v11
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v12, v17
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.l, v12.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v15, v13, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v11, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v11, v14, v16 :: v_dual_and_b32 v12, 0xffff0000, v10
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v12, 0x40c00000, v12 :: v_dual_cndmask_b32 v27, v15, v17
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v18, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v11, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v9
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v9
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v20, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_lshlrev_b32 v10, 16, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v10, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.l, v11.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v9
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_lshlrev_b32 v9, 16, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v10
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v10, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v16, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v9, 16, 1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v8
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v9, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v10, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v13, v15, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v11
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v14, v16, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v9
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v17, v22 :: v_dual_and_b32 v22, 0xffff0000, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.l, v10.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v11, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v9, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v9, v12, v14 :: v_dual_and_b32 v10, 0xffff0000, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v19, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v9, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v7 :: v_dual_add_f32 v8, 0x40c00000, v8
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_cndmask_b32 v25, v13, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v9.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v7
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_lshlrev_b32 v7, 16, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v8
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v8, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v8
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v10, v16
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v18, v18, v23 :: v_dual_and_b32 v23, 0xffff0000, v4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v10, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v8, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v11, v13, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v24, v12, v14, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v7
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.l, v8.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v11, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v7, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v10, v12 :: v_dual_and_b32 v8, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v20, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_cndmask_b32 v7, v16, v20
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v6, 16, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v6, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v20, 0x40c00000, v20 :: v_dual_add_f32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v8, 0x40c00000, v8 :: v_dual_cndmask_b32 v23, v11, v13
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v7.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v5
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_lshlrev_b32 v5, 16, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v6
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v5, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v7, v16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v19, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v5, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v5
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v20
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v8, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v9, v9, v6, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v9, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v22, v10, v12, vcc_lo
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v6, v17
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v4, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v23
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v4
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v4, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v18, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v16, v19, 16, 1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v6.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v9, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
+; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v9 :: v_dual_add_f32 v4, 0x40c00000, v4
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v8, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v4, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v6, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v5.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v2
-; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v5, v18
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v3
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT: v_add3_u32 v16, v16, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v17, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_add_f32 v3, 0x40c00000, v3
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v2
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v2
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v16, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v3, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v3
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v3, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v5 :: v_dual_cndmask_b32 v4, v7, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_add_f32 v3, 0x40c00000, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v20, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v23, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v17, v22 :: v_dual_and_b32 v22, 0xffff0000, v1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v2, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v16
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v22 :: v_dual_cndmask_b32 v18, v18, v20
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v2
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v22, 16, v0
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v4.h
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v8, v3, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v2, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v4, v9, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v10, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v3.h
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v3, 0xffff0000, v1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, 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
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, 0x40c00000, v1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_cndmask_b32 v2, v4, v5
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v6
+; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v2.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v3, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v5, 0x7fff
; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v17
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v19, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v1, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: 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_bfe_u32 v9, v0, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v1, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v22, v1, 0x7fff
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v0, 0x40c00000, v0
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v23, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v22, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_bfe_u32 v26, v0, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v0
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v26, v0, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v18
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v20.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v19, v24, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
-; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v19
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v22, v25, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v1, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v20, v0
-; GFX11-TRUE16-NEXT: .LBB82_2: ; %end
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v9, v0, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v0
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v4, v7, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v1.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v16, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v5.h
+; GFX11-TRUE16-NEXT: .LBB82_4: ; %end
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, v16 :: v_dual_mov_b32 v1, v17
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, v18 :: v_dual_mov_b32 v3, v19
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v20 :: v_dual_mov_b32 v5, v21
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, v22 :: v_dual_mov_b32 v7, v23
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, v24 :: v_dual_mov_b32 v9, v25
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, v26 :: v_dual_mov_b32 v11, v27
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, v28 :: v_dual_mov_b32 v13, v29
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v30 :: v_dual_mov_b32 v15, v31
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
; GFX11-FAKE16-LABEL: bitcast_v32bf16_to_v8f64:
@@ -73032,283 +73076,291 @@ define <32 x half> @bitcast_v32bf16_to_v32f16(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16-LABEL: bitcast_v32bf16_to_v32f16:
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v32, v15 :: v_dual_mov_b32 v31, v14
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v30, v13 :: v_dual_mov_b32 v29, v12
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v28, v11 :: v_dual_mov_b32 v27, v10
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v26, v9 :: v_dual_mov_b32 v25, v8
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v24, v7 :: v_dual_mov_b32 v23, v6
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v22, v5 :: v_dual_mov_b32 v21, v4
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v20, v3 :: v_dual_mov_b32 v19, v2
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v18, v1 :: v_dual_mov_b32 v17, v0
; 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_cmpx_ne_u32_e32 0, v16
; GFX11-TRUE16-NEXT: s_xor_b32 s0, exec_lo, s0
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB102_3
+; GFX11-TRUE16-NEXT: ; %bb.1: ; %Flow
+; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
+; GFX11-TRUE16-NEXT: s_cbranch_execnz .LBB102_4
+; GFX11-TRUE16-NEXT: .LBB102_2: ; %end
+; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
+; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
+; GFX11-TRUE16-NEXT: .LBB102_3: ; %cmp.false
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v0, v17 :: v_dual_mov_b32 v1, v18
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v2, v19 :: v_dual_mov_b32 v3, v20
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v4, v21 :: v_dual_mov_b32 v5, v22
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v6, v23 :: v_dual_mov_b32 v7, v24
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v8, v25 :: v_dual_mov_b32 v9, v26
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v10, v27 :: v_dual_mov_b32 v11, v28
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v12, v29 :: v_dual_mov_b32 v13, v30
+; GFX11-TRUE16-NEXT: v_dual_mov_b32 v14, v31 :: v_dual_mov_b32 v15, v32
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17
; GFX11-TRUE16-NEXT: s_and_not1_saveexec_b32 s0, s0
; GFX11-TRUE16-NEXT: s_cbranch_execz .LBB102_2
-; GFX11-TRUE16-NEXT: ; %bb.1: ; %cmp.true
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v16, 0xffff0000, v0
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff0000, v6
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff0000, v8
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff0000, v10
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v32, 0xffff0000, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v10
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v20, 0x400000, v16
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v17, 0xffff0000, v1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v17, 0x40c00000, v17 :: v_dual_lshlrev_b32 v12, 16, v12
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v0
-; GFX11-TRUE16-NEXT: v_bfe_u32 v0, v16, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v17, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v0, v0, v16, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v17
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v22, v17, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v21, v21, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v0, v0, v20, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v3
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v21, v23 :: v_dual_add_f32 v1, 0x40c00000, v1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v18, 0xffff0000, v2
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_lshlrev_b32 v2, 16, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v1, 16, 1
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v17, v22, v24 :: v_dual_add_f32 v18, 0x40c00000, v18
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v1, 0x7fff
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v4
+; GFX11-TRUE16-NEXT: .LBB102_4: ; %cmp.true
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v18
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v4, 0xffff0000, v19
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v23
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v14, 0xffff0000, v25
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v34, 0xffff0000, v31
; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v22, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v2, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v1.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v16.l, v16.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v22, v2, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v2
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v11, 16, v22
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v24
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v15, 16, v26
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v2
+; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v2, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 0xffff0000, v17
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v0, 0x40c00000, v0 :: v_dual_lshlrev_b32 v1, 16, v17
+; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v0, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v0
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v0, v0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v0, 0x7fff
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_cndmask_b32 v0, v5, v6
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v6, 0xffff0000, v20
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v1, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v18
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v16, v7, v9 :: v_dual_add_f32 v3, 0x40c00000, v3
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v23, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v1, v17
-; GFX11-TRUE16-NEXT: v_bfi_b32 v0, 0xffff, v16, v0
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v3, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v3, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v22, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v3
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v9, 16, v20
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v0.l, v16.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v1, v8, v10 :: v_dual_and_b32 v10, 0xffff0000, v21
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v4
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v19
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, 0x400000, v3
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v19, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v24 :: v_dual_add_f32 v22, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v3.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v22, 16, 1
-; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v3, v20
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v24, v21, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v19, v19, v23 :: v_dual_lshlrev_b32 v4, 16, v4
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v4, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v4
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v2, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v4, 0x40c00000, v4 :: v_dual_cndmask_b32 v17, v5, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add3_u32 v3, v8, v2, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v2
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v17.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v3, v5, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v3, v7, v4, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v4
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT: v_add3_u32 v21, v24, v4, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v24, v25, v22, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v22
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v21, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v5, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v22, v22
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v23, 0x40c00000, v26
-; GFX11-TRUE16-NEXT: v_add3_u32 v21, v21, v5, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v22, v24, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v7
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v5
-; GFX11-TRUE16-NEXT: v_bfe_u32 v26, v23, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v24, 0x40c00000, v24 :: v_dual_lshlrev_b32 v7, 16, v7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v2.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v21, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v21, v26, v23, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v23
-; GFX11-TRUE16-NEXT: v_bfe_u32 v26, v6, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v23, v23
-; GFX11-TRUE16-NEXT: v_bfe_u32 v27, v24, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v7
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v5.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v26, v6, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v21, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v8, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v9
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v18, v3, v5, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v26, v27, v24, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v24
-; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v5, v22
-; GFX11-TRUE16-NEXT: v_bfi_b32 v2, 0xffff, v2, v18
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v23, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v25, 0x40c00000, v28
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v4.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v5, 0x40c00000, v10 :: v_dual_lshlrev_b32 v6, 16, v21
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v7, v9, vcc_lo
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v7, 0xffff0000, v22
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v4, v8, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v8
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v5, 16, 1
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v23, v7, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v24, v26, v27, vcc_lo
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v26, 0xffff0000, v9
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_bfe_u32 v28, v25, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v7, 0x40c00000, v7
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v19, v4, v9, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v10, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v9, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v19.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add3_u32 v5, v9, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v9, v10, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v4, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v10, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v7
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v5, v8, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v26, 0x40c00000, v26 :: v_dual_lshlrev_b32 v9, 16, v9
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v4, v19
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v28, v25, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v25
-; GFX11-TRUE16-NEXT: v_bfe_u32 v28, v8, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v25, v25
-; GFX11-TRUE16-NEXT: v_bfe_u32 v29, v26, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v12 :: v_dual_lshlrev_b32 v8, 16, v23
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v10, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v9, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v9, 0xffff0000, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v7, 16, 1
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
; GFX11-TRUE16-NEXT: v_add_f32_e32 v9, 0x40c00000, v9
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v7.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v25, v28, v8, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v23, v23, v27, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v8
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v7
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v24, 0xffff0000, v27
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v18.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v6, v11, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v6, v12, v7, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v11, v8, 16, 1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v9, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v21.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add3_u32 v7, v11, v8, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v6, v6, v10, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, 0x400000, v8
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT: v_add3_u32 v28, v29, v26, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, 0x400000, v26
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v25, v27, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v9, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v26, v26
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v27, 0x40c00000, v30
-; GFX11-TRUE16-NEXT: v_add3_u32 v25, v25, v9, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v28, v29, vcc_lo
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff0000, v11
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, 0x400000, v9
-; GFX11-TRUE16-NEXT: v_bfe_u32 v30, v27, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v12, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v22, v7, v10, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v8, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v9, 0x40c00000, v14 :: v_dual_lshlrev_b32 v10, 16, v25
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v7, v11, v13, vcc_lo
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v11, 0xffff0000, v26
+; GFX11-TRUE16-NEXT: v_add3_u32 v8, v8, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v9, 16, 1
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v28, 0x40c00000, v28 :: v_dual_lshlrev_b32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v25, v29, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v25, v30, v27, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, 0x400000, v27
-; GFX11-TRUE16-NEXT: v_bfe_u32 v30, v10, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v27, v27
-; GFX11-TRUE16-NEXT: v_bfe_u32 v31, v28, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v11
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v9.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v27, v30, v10, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v25, v25, v29, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v29, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v10, 0x40c00000, v10 :: v_dual_add_f32 v11, 0x40c00000, v11
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v9
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v20.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v23, v8, v13, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v8, v14, v9, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v10, 16, 1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v23.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add3_u32 v9, v13, v10, 0x7fff
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v14, v11, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v8, v8, v12, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, 0x400000, v10
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v15
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT: v_add3_u32 v30, v31, v28, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, 0x400000, v28
-; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v6, v21
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v27, v29, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v27, v11, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v29, 0x40c00000, v32
-; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v7, v24
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v11
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v27, v27, v11, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v30, v31, vcc_lo
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v30, 0xffff0000, v13
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, 0x400000, v11
-; GFX11-TRUE16-NEXT: v_bfe_u32 v32, v29, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v25, v9, v12, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v12, 16, v27
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v27, 16, v28
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v11, 0x40c00000, v24
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v13, v15, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v15, 0x400000, v14
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v10, v14, 0x7fff
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v13, 0xffff0000, v28
+; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v11, 16, 1
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v11
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v30, 0x40c00000, v30 :: v_dual_lshlrev_b32 v13, 16, v13
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v27, v31, vcc_lo
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v26, v10, v15 :: v_dual_add_f32 v13, 0x40c00000, v13
+; GFX11-TRUE16-NEXT: v_add3_u32 v10, v24, v11, 0x7fff
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add3_u32 v27, v32, v29, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, 0x400000, v29
-; GFX11-TRUE16-NEXT: v_bfe_u32 v32, v12, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
-; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v30, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v11.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v29, v32, v12, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v27, v31, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_bfe_u32 v15, v12, 16, 1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v11, v11
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v28, 0xffff0000, v29
+; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v13, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v26.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v11, v15, v12, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v10, v10, v14, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, 0x400000, v12
+; GFX11-TRUE16-NEXT: v_add3_u32 v15, v24, v13, 0x7fff
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v24, 0x40c00000, v27
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT: v_add3_u32 v32, v33, v30, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v33, 0x400000, v30
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v34, 0x400000, v13
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v29, v31, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v31, 0xffff0000, v14
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v14, 16, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v29, v13, 16, 1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v30, v32, v33, vcc_lo
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v32, 16, v15
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v14, 0x40c00000, v14 :: v_dual_add_f32 v31, 0x40c00000, v31
-; GFX11-TRUE16-NEXT: v_add3_u32 v29, v29, v13, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v32, 0x40c00000, v32 :: v_dual_add_f32 v15, 0x40c00000, v15
-; GFX11-TRUE16-NEXT: v_bfe_u32 v35, v14, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, 0x400000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v33, v31, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v25.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v22.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v12, v24, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v11, v14, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-TRUE16-NEXT: v_dual_add_f32 v13, 0x40c00000, v28 :: v_dual_lshlrev_b32 v14, 16, v29
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v12, v24, 0x7fff
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v33.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v15, v27, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v24
+; GFX11-TRUE16-NEXT: v_bfe_u32 v28, v13, 16, 1
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v24, v24
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v30
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v30
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v24, v12, v27, vcc_lo
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
+; GFX11-TRUE16-NEXT: v_add3_u32 v12, v28, v13, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v13
+; GFX11-TRUE16-NEXT: v_bfe_u32 v28, v14, 16, 1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
+; GFX11-TRUE16-NEXT: v_bfe_u32 v29, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v36, v32, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v39, 0x400000, v32
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
-; GFX11-TRUE16-NEXT: v_bfe_u32 v38, v15, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v35, v35, v14, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v36, v36, v32, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v15
-; GFX11-TRUE16-NEXT: v_add3_u32 v33, v33, v31, 0x7fff
-; GFX11-TRUE16-NEXT: v_add3_u32 v38, v38, v15, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v31
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v32, v36, v39, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v13, v28, v14, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v12, v12, v27, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v27, 0x400000, v14
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v12, v27
-; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v11, v28
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v32.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v35, v48, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v28, v29, v15, 0x7fff
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v29, 0x40c00000, v30
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v30, 0x400000, v15
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v13, v27, vcc_lo
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v10, v25
-; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v9, v26
-; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v23
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v15, v38, v49, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v15, 0xffff, v32, v15
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v29, v34, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v31, v31
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v13.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v29, v33, v37, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v13, v30
-; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v14, v29
-; GFX11-TRUE16-NEXT: .LBB102_2: ; %end
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v29, 16, 1
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v34
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v27.h
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v13, v28, v30 :: v_dual_lshlrev_b32 v28, 16, v31
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v29, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v30, 16, v32
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v31, 0x400000, v29
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v29, v29
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v28, 0x40c00000, v28
+; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v15, 16, 1
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v30, 0x40c00000, v30
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v29, v14, v31 :: v_dual_and_b32 v32, 0xffff0000, v32
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v14, v28, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v31, v34, v15, 0x7fff
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v34, v30, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v36, 0x400000, v28
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v28, v28
+; GFX11-TRUE16-NEXT: v_add3_u32 v14, v14, v28, 0x7fff
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v32, 0x40c00000, v32
+; GFX11-TRUE16-NEXT: v_add3_u32 v34, v34, v30, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v38, 0x400000, v30
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v35, 0x400000, v15
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v14, v36, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v30, v30
+; GFX11-TRUE16-NEXT: v_bfe_u32 v37, v32, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v29.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v30, v34, v38, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_add3_u32 v36, v37, v32, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v37, 0x400000, v32
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v14, v31, v35, vcc_lo
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v32, v32
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v28.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v15, v36, v37, vcc_lo
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v30.h
; GFX11-TRUE16-NEXT: s_or_b32 exec_lo, exec_lo, s0
; GFX11-TRUE16-NEXT: s_setpc_b64 s[30:31]
;
@@ -85371,559 +85423,560 @@ define <64 x i8> @bitcast_v32bf16_to_v64i8(<32 x bfloat> %a, i32 %b) {
; GFX11-TRUE16: ; %bb.0:
; GFX11-TRUE16-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX11-TRUE16-NEXT: v_cmp_ne_u32_e32 vcc_lo, 0, v17
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr28_hi16
+; 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: $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: $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: $vgpr21_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_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: $vgpr113_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr24_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr26_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr112_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr27_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr31_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr102_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr23_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr101_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr100_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr34_hi16
+; 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: $vgpr99_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr22_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr98_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr33_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr38_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr21_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr35_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr97_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: $vgpr82_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: $vgpr84_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr36_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr82_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr55_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr32_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr30_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr96_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr81_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr20_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr80_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr39_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr114_hi16
; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr71_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr70_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr68_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr19_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr52_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr66_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr53_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr65_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr85_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr64_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr18_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr69_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr54_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr67_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr51_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr87_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr50_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr17_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr83_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr49_lo16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr86_hi16
-; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr48_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr37_hi16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr29_lo16
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr103_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
; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1)
; 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[17:18], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 24, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 8, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 8, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v10
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v9
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v3
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v1
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v28.h, v1.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.h, v2.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v31.h, v3.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v29.h, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v30.h, v4.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v34.h, v5.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.h, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v6.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v38.h, v7.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v35.h, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v36.h, v8.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v55.h, v9.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v37.h, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v39.h, v10.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v70.h, v11.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v52.h, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v53.h, v12.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v85.h, v13.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v69.h, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v67.h, v14.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v87.h, v15.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v83.h, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v86.h, v16.h
+; 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_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 v80, 24, v14
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 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 v87, 8, v11
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 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 v113, 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 v118, 24, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v4
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v3
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v129, 24, v2
+; 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_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 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
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v54.h, v4.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.h, v4.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v65.h, v5.l
+; 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 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 v83.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 v26.h, v10.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v97.h, v11.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v27.h, v11.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v82.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 v32.h, v13.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v96.h, v14.l
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v33.h, v14.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v114.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: ; implicit-def: $vgpr1
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr3
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr5
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr7
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr9
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr11
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr13
+; GFX11-TRUE16-NEXT: ; implicit-def: $vgpr15
; GFX11-TRUE16-NEXT: .LBB108_2: ; %Flow
; 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 v18, 0xffff0000, v1
; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v4
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4
; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v17, 16, v2
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v2, 0xffff0000, v2
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v18, 0x40c00000, v18 :: v_dual_lshlrev_b32 v1, 16, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v49, 0xffff0000, v11
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v2, 0x40c00000, v2 :: v_dual_lshlrev_b32 v11, 16, v11
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v1, 0x40c00000, v1 :: v_dual_add_f32 v4, 0x40c00000, v4
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v17, 0x40c00000, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v2, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v2
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v17, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v17
+; 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: v_bfe_u32 v18, v17, 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 v21, v21, v2, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v17, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v26, v20, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v1, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v17, 0x7fff
+; 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_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_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v18, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v26.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v1, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v27, v21, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v19
+; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v20, 16, 1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v1, 16, 1
+; GFX11-TRUE16-NEXT: v_add3_u32 v21, v21, v2, 0x7fff
+; 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_bfi_b32 v2, 0xffff, v2, v27
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v28, v20, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v4, 16, 1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v3
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v28.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v17, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v4, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v4
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_bfi_b32 v1, 0xffff, v20, v1
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v20, 0xffff0000, v5
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v5
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v29, v18, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v6
+; GFX11-TRUE16-NEXT: v_add3_u32 v17, v24, v20, 0x7fff
+; 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_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_add3_u32 v17, v17, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v21, 0x40c00000, v21 :: v_dual_lshlrev_b32 v6, 16, v6
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v3, 0x40c00000, v3 :: v_dual_add_f32 v20, 0x40c00000, v20
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v30, v17, v23, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v4, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v3, 16, 1
+; 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_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_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_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_add3_u32 v4, v4, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v17.l, v29.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v3, 0x7fff
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v6, 0x40c00000, v6 :: v_dual_add_f32 v5, 0x40c00000, v5
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v103, 24, v2
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v112, 8, v2
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v31, v18, v19, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v18, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v1
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v31.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v4, v23, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v21, v18, 16, 1
-; GFX11-TRUE16-NEXT: v_bfi_b32 v4, 0xffff, v17, v30
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v6, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
-; GFX11-TRUE16-NEXT: v_bfi_b32 v3, 0xffff, v19, v3
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v21, v18, 0x7fff
+; 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: 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_cmp_u_f32_e32 vcc_lo, v22, v22
+; 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
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v55.h
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v118, 24, v20
+; 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, v4 :: v_dual_add_f32 v5, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v119, 8, v20
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v128, 8, v19
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v6, 16, 1
; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v6
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v6, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v18
-; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v20, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v100, 24, v4
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 8, v4
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v32, v17, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v18, v18
-; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v102, 8, v3
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v33, v19, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v19, 0xffff0000, v8
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v8, 16, v8
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v5, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v6, v6
+; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v6, 16, v8
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v64, v3, v21 :: v_dual_and_b32 v3, 0xffff0000, v8
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, 0x40c00000, v2
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v5
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v4
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
+; GFX11-TRUE16-NEXT: v_bfe_u32 v1, 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: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v2, 0x7fff
+; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v5, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v22, v1, v22, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfe_u32 v1, v4, 16, 1
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3)
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v5, 0x7fff
; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v8, 0x40c00000, v8
-; GFX11-TRUE16-NEXT: v_add3_u32 v17, v17, v5, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v32.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v34, v17, v21 :: v_dual_add_f32 v19, 0x40c00000, v19
-; GFX11-TRUE16-NEXT: v_bfe_u32 v17, v8, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v34.h
-; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v20, v6, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v6, v17, v8, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v17, 0x400000, v8
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v8, v8
-; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v35, v6, v17, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v7
-; GFX11-TRUE16-NEXT: v_bfi_b32 v6, 0xffff, v18, v33
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v36, v5, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfi_b32 v5, 0xffff, v21, v20
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v20, 16, v10
+; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v64.h
+; GFX11-TRUE16-NEXT: v_add3_u32 v1, v1, v4, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v65, v2, v8, vcc_lo
+; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v5, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v21, v1, v21, vcc_lo
+; GFX11-TRUE16-NEXT: v_add3_u32 v1, v2, v3, 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_mov_b16_e32 v8.l, v35.h
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v97, 24, v6
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v98, 8, v6
+; 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_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 v20, 0x40c00000, v20 :: v_dual_add_f32 v7, 0x40c00000, v7
-; GFX11-TRUE16-NEXT: v_bfi_b32 v8, 0xffff, v8, v36
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v99, 8, v5
+; 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_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_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_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_bfe_u32 v19, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v82, 24, v8
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v8
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v7, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v20, 16, 1
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v38, v19, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v20, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v21, 0x400000, v20
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v10, 0xffff0000, v10
-; 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, v23 :: v_dual_add_f32 v10, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: v_bfe_u32 v18, v17, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v17
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v17, v17
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v10, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v18, v18, v17, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v10, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v17, v18, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v20, v20
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v18.l, v38.h
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v37, v7, v21, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v10, v10
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v10, 16, v12
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v12, 0xffff0000, v12
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v7, 16, v9
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v39, v19, v22, vcc_lo
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_3)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v10
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v37.h
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v12, 0x40c00000, v12
-; GFX11-TRUE16-NEXT: v_bfe_u32 v22, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v48, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_bfi_b32 v10, 0xffff, v19, v39
-; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v12, 16, 1
-; GFX11-TRUE16-NEXT: v_add3_u32 v22, v22, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v9
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v50, 0x400000, v12
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v49
-; GFX11-TRUE16-NEXT: v_add3_u32 v24, v24, v12, 0x7fff
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v52, v22, v48 :: v_dual_add_f32 v9, 0x40c00000, v23
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v22, 0xffff0000, v14
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v7 :: v_dual_lshlrev_b32 v14, 16, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 24, v10
-; GFX11-TRUE16-NEXT: v_bfe_u32 v20, v9, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v9
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4)
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v14, 0x40c00000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_add3_u32 v20, v20, v9, 0x7fff
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v80, 8, v10
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v19, v7, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v52.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v55, v19, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v12, v12
-; GFX11-TRUE16-NEXT: v_bfe_u32 v19, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v14, 16, 1
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v53, v24, v50, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v9, v9
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
-; GFX11-TRUE16-NEXT: v_bfi_b32 v12, 0xffff, v7, v53
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v9, v20, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v7, 0x40c00000, v11
-; GFX11-TRUE16-NEXT: v_add3_u32 v11, v19, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v20.l, v55.h
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v65, 24, v12
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v66, 8, v12
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v11, v11, v19, vcc_lo
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v19, 0x40c00000, v22
-; GFX11-TRUE16-NEXT: v_add3_u32 v21, v23, v7, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v23, 0xffff0000, v13
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v19, 16, 1
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v13, 16, v13
-; GFX11-TRUE16-NEXT: v_bfi_b32 v9, 0xffff, v20, v9
-; GFX11-TRUE16-NEXT: v_dual_add_f32 v7, 0x40c00000, v23 :: v_dual_cndmask_b32 v70, v21, 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_cmp_u_f32_e32 vcc_lo, v5, v5
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v12
+; 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_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_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_mov_b16_e32 v23.l, v68.h
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v26.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_lshrrev_b64 v[48:49], 24, v[23:24]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[49:50], 24, v[21:22]
+; 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[50:51], 24, v[19:20]
+; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[51:52], 24, v[17:18]
+; GFX11-TRUE16-NEXT: v_add3_u32 v1, v6, v5, 0x7fff
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v83, 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_add3_u32 v21, v24, v19, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v22, 0x400000, v19
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v19, v19
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v25, v14, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v25, v7, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v13, 0x40c00000, v13
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v67, v21, v22, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v19, 0x400000, v7
-; GFX11-TRUE16-NEXT: v_add3_u32 v14, v25, v7, 0x7fff
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v21, 0xffff0000, v16
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v16, 16, v16
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v69, v23, v24, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v7, v7
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v13, 16, 1
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v9
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v16, 0x40c00000, v16
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v22.l, v69.h
-; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v7, v14, v19 :: v_dual_add_f32 v14, 0x40c00000, v21
-; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v21, 16, v15
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v23, v13, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v23, 0x400000, v13
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v13, v13
-; GFX11-TRUE16-NEXT: v_bfe_u32 v13, v16, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v21, 0x40c00000, v21
-; GFX11-TRUE16-NEXT: v_and_b32_e32 v15, 0xffff0000, v15
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v25, 0x400000, v16
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v85, v19, v23, vcc_lo
-; GFX11-TRUE16-NEXT: v_add3_u32 v13, v13, v16, 0x7fff
-; GFX11-TRUE16-NEXT: v_bfe_u32 v23, v21, 16, 1
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v16, v16
-; GFX11-TRUE16-NEXT: v_bfe_u32 v24, v14, 16, 1
-; GFX11-TRUE16-NEXT: v_add_f32_e32 v15, 0x40c00000, v15
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v49, 0x400000, v21
-; GFX11-TRUE16-NEXT: v_add3_u32 v23, v23, v21, 0x7fff
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v83, v13, v25, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v21, v21
-; GFX11-TRUE16-NEXT: v_add3_u32 v19, v24, v14, 0x7fff
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v24, 0x400000, v14
-; GFX11-TRUE16-NEXT: v_bfe_u32 v48, v15, 16, 1
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, 0x400000, v15
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v87, v23, v49, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v14, v14
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v21.l, v85.h
-; GFX11-TRUE16-NEXT: v_add3_u32 v13, v48, v15, 0x7fff
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v23.l, v70.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v14, 0xffff, v22, v67
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v86, v19, v24, vcc_lo
-; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v15, v15
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v19.l, v83.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v87.h
-; GFX11-TRUE16-NEXT: v_bfi_b32 v11, 0xffff, v23, v11
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v51, 24, v14
-; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v13, v13, v16, vcc_lo
-; GFX11-TRUE16-NEXT: v_bfi_b32 v16, 0xffff, v19, v86
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v54, 8, v14
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v68, 8, v11
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v25.l, v83.h
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v101, 24, v24
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v82, v1, v2 :: v_dual_lshlrev_b32 v1, 16, v11
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v4, 16, 1
+; 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_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_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_lshrrev_b32_e32 v102, 8, v24
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v113, 8, v23
+; 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_mov_b16_e32 v28.l, v82.h
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v97, 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_mov_b16_e32 v27.l, v97.h
+; 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_lshlrev_b32_e32 v9, 16, v13
+; GFX11-TRUE16-NEXT: v_add3_u32 v3, v8, v1, 0x7fff
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1
+; GFX11-TRUE16-NEXT: v_and_b32_e32 v8, 0xffff0000, v15
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v96, v6, v7, vcc_lo
+; 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_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_mov_b16_e32 v33.l, v96.h
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v85, 24, v28
+; 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_b32_e32 v80, 24, v33
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v81, 8, v33
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, 0x40c00000, v3
+; 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 v86, 8, v28
+; GFX11-TRUE16-NEXT: v_bfe_u32 v2, v3, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v3
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-TRUE16-NEXT: v_dual_cndmask_b32 v112, v4, v6 :: v_dual_add_f32 v1, 0x40c00000, v5
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v6, 0x40c00000, v8
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v2, v3, 0x7fff
+; 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_bfi_b32 v15, 0xffff, v15, v13
-; GFX11-TRUE16-NEXT: v_bfi_b32 v13, 0xffff, v21, v7
-; GFX11-TRUE16-NEXT: v_bfi_b32 v7, 0xffff, v18, v17
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v48, 24, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v49, 8, v16
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[17:18], 24, v[15:16]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[18:19], 24, v[13:14]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[19:20], 24, v[11:12]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[20:21], 24, v[9:10]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[21:22], 24, v[7:8]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[22:23], 24, v[5:6]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[23:24], 24, v[3:4]
-; GFX11-TRUE16-NEXT: v_lshrrev_b64 v[24:25], 24, v[1:2]
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v50, 8, v15
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v64, 8, v13
-; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v96, 8, v7
+; GFX11-TRUE16-NEXT: v_mov_b16_e32 v32.l, v112.h
+; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v1, 16, 1
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v1
+; GFX11-TRUE16-NEXT: v_bfe_u32 v10, v6, 16, 1
+; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v103, v2, v9, vcc_lo
+; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, 0x400000, v6
+; GFX11-TRUE16-NEXT: v_add3_u32 v4, v7, v1, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v5, 16, v15
+; GFX11-TRUE16-NEXT: v_add3_u32 v2, v10, v6, 0x7fff
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v84, 8, v32
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v87, 8, v27
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_add_f32_e32 v5, 0x40c00000, v5
+; 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 v114, 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_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, v114.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_b64 v[30:31], 24, v[32:33]
+; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v71, 8, v37
; 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, v28.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.l, 8, v113.l
-; GFX11-TRUE16-NEXT: v_and_b16 v1.h, 0xff, v1.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v24.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v24.h, 0
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v27.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v1.l, v2.l
-; GFX11-TRUE16-NEXT: v_and_b16 v2.l, 0xff, v26.h
-; GFX11-TRUE16-NEXT: v_or_b16 v1.h, v1.h, v2.h
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.l, v24.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v2.h, 8, v112.l
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v103.l
-; GFX11-TRUE16-NEXT: v_and_b16 v3.h, 0xff, v3.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v23.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v1, v24, v1
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v2.l, v2.h
-; GFX11-TRUE16-NEXT: v_or_b16 v2.h, v3.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v3.l, 0xff, v31.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.l, 8, v102.l
-; GFX11-TRUE16-NEXT: v_or_b16 v3.h, v3.h, v4.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v4.h, 8, v101.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v2, v24, v2
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v30.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v3.l, v4.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v4.l, 0xff, v29.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v100.l
-; GFX11-TRUE16-NEXT: v_and_b16 v5.h, 0xff, v5.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v22.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v3, v24, v3
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v4.l, v4.h
-; GFX11-TRUE16-NEXT: v_or_b16 v4.h, v5.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v5.l, 0xff, v34.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.l, 8, v99.l
-; GFX11-TRUE16-NEXT: v_or_b16 v5.h, v5.h, v6.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v6.h, 8, v98.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, v24, v4
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v33.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v5.l, v6.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v6.l, 0xff, v32.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v97.l
-; GFX11-TRUE16-NEXT: v_and_b16 v7.h, 0xff, v7.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v21.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v5, v24, v5
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v6.l, v6.h
-; GFX11-TRUE16-NEXT: v_or_b16 v6.h, v7.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v6.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v7.l, 0xff, v38.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.l, 8, v96.l
-; GFX11-TRUE16-NEXT: v_or_b16 v7.h, v7.h, v8.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v8.h, 8, v84.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, v24, v6
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v36.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v7.l, v8.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v7.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v8.l, 0xff, v35.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v82.l
-; GFX11-TRUE16-NEXT: v_and_b16 v9.h, 0xff, v9.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v20.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v7, v24, v7
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v8.l, v8.h
-; GFX11-TRUE16-NEXT: v_or_b16 v8.h, v9.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v8.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v9.l, 0xff, v55.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.l, 8, v81.l
-; GFX11-TRUE16-NEXT: v_or_b16 v9.h, v9.h, v10.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v10.h, 8, v80.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, v24, v8
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v39.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v9.l, v10.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v9.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v10.l, 0xff, v37.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v71.l
-; GFX11-TRUE16-NEXT: v_and_b16 v11.h, 0xff, v11.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v19.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, v24, v9
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v10.l, v10.h
-; GFX11-TRUE16-NEXT: v_or_b16 v10.h, v11.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v10.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v11.l, 0xff, v70.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.l, 8, v68.l
-; GFX11-TRUE16-NEXT: v_or_b16 v11.h, v11.h, v12.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v12.h, 8, v66.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v10, v24, v10
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v53.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v11.l, v12.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v11.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v12.l, 0xff, v52.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v65.l
-; GFX11-TRUE16-NEXT: v_and_b16 v13.h, 0xff, v13.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v18.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v11, v24, v11
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v12.l, v12.h
-; GFX11-TRUE16-NEXT: v_or_b16 v12.h, v13.l, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v12.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v13.l, 0xff, v85.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.l, 8, v64.l
-; GFX11-TRUE16-NEXT: v_or_b16 v13.h, v13.h, v14.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v54.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v12, v24, v12
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v67.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v13.l, v14.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v13.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v14.l, 0xff, v69.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v51.l
-; GFX11-TRUE16-NEXT: v_and_b16 v15.h, 0xff, v15.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v17.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v13, v24, v13
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v14.l, v14.h
-; GFX11-TRUE16-NEXT: v_or_b16 v14.h, v15.l, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v14.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v15.l, 0xff, v87.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.l, 8, v50.l
-; GFX11-TRUE16-NEXT: v_or_b16 v15.h, v15.h, v16.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v16.h, 8, v49.l
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v14, v24, v14
-; GFX11-TRUE16-NEXT: v_and_b16 v17.l, 0xff, v86.h
-; GFX11-TRUE16-NEXT: v_or_b16 v24.l, v15.l, v16.l
-; GFX11-TRUE16-NEXT: v_mov_b16_e32 v15.l, v24.h
-; GFX11-TRUE16-NEXT: v_and_b16 v16.l, 0xff, v83.h
-; GFX11-TRUE16-NEXT: v_lshlrev_b16 v17.h, 8, v48.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, v16.h
-; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
+; 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_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.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_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_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_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, v113.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_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_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, v83.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, v80.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, v87.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, v82.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_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, v96.h
+; GFX11-TRUE16-NEXT: v_lshlrev_b16 v14.h, 8, v81.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, v114.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, v24.h
-; GFX11-TRUE16-NEXT: v_or_b32_e32 v16, v24, v16
+; 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: 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