diff options
Diffstat (limited to 'llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll')
| -rw-r--r-- | llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll | 2806 |
1 files changed, 1424 insertions, 1382 deletions
diff --git a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll index 7f95e7df3f6b..929bb61ddabc 100644 --- a/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll +++ b/llvm/test/CodeGen/AMDGPU/local-atomicrmw-fsub.ll @@ -50,17 +50,17 @@ define float @local_atomic_fsub_ret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX942-LABEL: local_atomic_fsub_ret_f32: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v2, v0 +; GFX942-NEXT: ds_read_b32 v1, v0 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB0_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB0_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -119,17 +119,17 @@ define float @local_atomic_fsub_ret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX90A-LABEL: local_atomic_fsub_ret_f32: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v2, v0 +; GFX90A-NEXT: ds_read_b32 v1, v0 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB0_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB0_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -262,17 +262,17 @@ define float @local_atomic_fsub_ret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX942-LABEL: local_atomic_fsub_ret_f32__offset: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v2, v0 offset:65532 +; GFX942-NEXT: ds_read_b32 v1, v0 offset:65532 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB1_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB1_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -331,17 +331,17 @@ define float @local_atomic_fsub_ret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX90A-LABEL: local_atomic_fsub_ret_f32__offset: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532 +; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65532 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB1_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB1_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -453,13 +453,14 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX12-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-NEXT: s_wait_dscnt 0x0 -; GFX12-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX12-NEXT: v_mov_b32_e32 v2, v1 +; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX12-NEXT: s_wait_storecnt 0x0 -; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 +; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 ; GFX12-NEXT: s_wait_dscnt 0x0 ; GFX12-NEXT: global_inv scope:SCOPE_SE -; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX12-NEXT: s_wait_alu 0xfffe ; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-NEXT: s_wait_alu 0xfffe @@ -478,12 +479,12 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX942-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX942-NEXT: v_mov_b32_e32 v2, v1 +; GFX942-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v1, v2 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB2_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -498,13 +499,14 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX11-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX11-NEXT: v_mov_b32_e32 v2, v1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 +; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-NEXT: buffer_gl0_inv -; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -521,13 +523,13 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX10-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX10-NEXT: v_mov_b32_e32 v2, v1 +; GFX10-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX10-NEXT: v_mov_b32_e32 v1, v2 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB2_1 @@ -543,12 +545,12 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX90A-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 +; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v1, v2 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB2_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -563,12 +565,12 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX908-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX908-NEXT: v_mov_b32_e32 v2, v1 +; GFX908-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v1, v2 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB2_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -584,12 +586,12 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX8-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX8-NEXT: v_mov_b32_e32 v2, v1 +; GFX8-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v1, v2 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB2_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -605,12 +607,12 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX7-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX7-NEXT: v_mov_b32_e32 v2, v1 +; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v1, v2 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB2_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -626,12 +628,12 @@ define void @local_atomic_fsub_noret_f32(ptr addrspace(3) %ptr) nounwind { ; GFX6-NEXT: .LBB2_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX6-NEXT: v_mov_b32_e32 v2, v1 +; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v1, v2 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB2_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -654,13 +656,14 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX12-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-NEXT: s_wait_dscnt 0x0 -; GFX12-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX12-NEXT: v_mov_b32_e32 v2, v1 +; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX12-NEXT: s_wait_storecnt 0x0 -; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532 +; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532 ; GFX12-NEXT: s_wait_dscnt 0x0 ; GFX12-NEXT: global_inv scope:SCOPE_SE -; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX12-NEXT: s_wait_alu 0xfffe ; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-NEXT: s_wait_alu 0xfffe @@ -679,12 +682,12 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX942-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532 +; GFX942-NEXT: v_mov_b32_e32 v2, v1 +; GFX942-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v1, v2 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB3_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -699,13 +702,14 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX11-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX11-NEXT: v_mov_b32_e32 v2, v1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65532 +; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65532 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-NEXT: buffer_gl0_inv -; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -722,13 +726,13 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX10-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX10-NEXT: v_mov_b32_e32 v2, v1 +; GFX10-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532 +; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX10-NEXT: v_mov_b32_e32 v1, v2 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB3_1 @@ -744,12 +748,12 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX90A-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532 +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 +; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v1, v2 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB3_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -764,12 +768,12 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX908-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532 +; GFX908-NEXT: v_mov_b32_e32 v2, v1 +; GFX908-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v1, v2 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB3_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -785,12 +789,12 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX8-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532 +; GFX8-NEXT: v_mov_b32_e32 v2, v1 +; GFX8-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v1, v2 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB3_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -806,12 +810,12 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX7-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65532 +; GFX7-NEXT: v_mov_b32_e32 v2, v1 +; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65532 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v1, v2 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB3_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -828,12 +832,12 @@ define void @local_atomic_fsub_noret_f32__offset(ptr addrspace(3) %ptr) nounwind ; GFX6-NEXT: .LBB3_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX6-NEXT: v_mov_b32_e32 v2, v1 +; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v1, v2 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB3_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -883,18 +887,18 @@ define double @local_atomic_fsub_ret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX942-LABEL: local_atomic_fsub_ret_f64: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b64 v[4:5], v0 ; GFX942-NEXT: v_mov_b32_e32 v2, v0 +; GFX942-NEXT: ds_read_b64 v[0:1], v0 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB4_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b64_e32 v[4:5], v[0:1] ; GFX942-NEXT: v_add_f64 v[0:1], v[4:5], -4.0 ; GFX942-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5] ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b64_e32 v[4:5], v[0:1] ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB4_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -953,18 +957,18 @@ define double @local_atomic_fsub_ret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX90A-LABEL: local_atomic_fsub_ret_f64: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b64 v[4:5], v0 ; GFX90A-NEXT: v_mov_b32_e32 v2, v0 +; GFX90A-NEXT: ds_read_b64 v[0:1], v0 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB4_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1] ; GFX90A-NEXT: v_add_f64 v[0:1], v[4:5], -4.0 ; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5] ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1] ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB4_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1100,18 +1104,18 @@ define double @local_atomic_fsub_ret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX942-LABEL: local_atomic_fsub_ret_f64__offset: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b64 v[4:5], v0 offset:65528 ; GFX942-NEXT: v_mov_b32_e32 v2, v0 +; GFX942-NEXT: ds_read_b64 v[0:1], v0 offset:65528 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB5_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b64_e32 v[4:5], v[0:1] ; GFX942-NEXT: v_add_f64 v[0:1], v[4:5], -4.0 ; GFX942-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5] ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b64_e32 v[4:5], v[0:1] ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB5_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1170,18 +1174,18 @@ define double @local_atomic_fsub_ret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX90A-LABEL: local_atomic_fsub_ret_f64__offset: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b64 v[4:5], v0 offset:65528 ; GFX90A-NEXT: v_mov_b32_e32 v2, v0 +; GFX90A-NEXT: ds_read_b64 v[0:1], v0 offset:65528 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB5_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1] ; GFX90A-NEXT: v_add_f64 v[0:1], v[4:5], -4.0 ; GFX90A-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[4:5], v[0:1] offset:65528 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[4:5] ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[0:1], v[0:1] op_sel:[0,1] ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB5_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1296,13 +1300,14 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX12-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-NEXT: s_wait_dscnt 0x0 -; GFX12-NEXT: v_add_f64_e32 v[3:4], -4.0, v[1:2] +; GFX12-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-NEXT: v_add_f64_e32 v[1:2], -4.0, v[3:4] ; GFX12-NEXT: s_wait_storecnt 0x0 -; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] +; GFX12-NEXT: ds_cmpstore_rtn_b64 v[1:2], v0, v[1:2], v[3:4] ; GFX12-NEXT: s_wait_dscnt 0x0 ; GFX12-NEXT: global_inv scope:SCOPE_SE -; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2] -; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4 +; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[1:2], v[3:4] ; GFX12-NEXT: s_wait_alu 0xfffe ; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-NEXT: s_wait_alu 0xfffe @@ -1321,12 +1326,12 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX942-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_add_f64 v[4:5], v[2:3], -4.0 -; GFX942-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] +; GFX942-NEXT: v_mov_b64_e32 v[4:5], v[2:3] +; GFX942-NEXT: v_add_f64 v[2:3], v[4:5], -4.0 +; GFX942-NEXT: ds_cmpst_rtn_b64 v[2:3], v0, v[4:5], v[2:3] ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3] +; GFX942-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5] ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b64_e32 v[2:3], v[4:5] ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB6_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1341,13 +1346,14 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX11-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 +; GFX11-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 ; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] +; GFX11-NEXT: ds_cmpstore_rtn_b64 v[1:2], v0, v[1:2], v[3:4] ; GFX11-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-NEXT: buffer_gl0_inv -; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2] -; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4 +; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[1:2], v[3:4] ; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -1364,14 +1370,14 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX10-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 +; GFX10-NEXT: v_mov_b32_e32 v4, v2 +; GFX10-NEXT: v_mov_b32_e32 v3, v1 +; GFX10-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] +; GFX10-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2] -; GFX10-NEXT: v_mov_b32_e32 v1, v3 -; GFX10-NEXT: v_mov_b32_e32 v2, v4 +; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[1:2], v[3:4] ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB6_1 @@ -1387,12 +1393,12 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX90A-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_add_f64 v[4:5], v[2:3], -4.0 -; GFX90A-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] +; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1] +; GFX90A-NEXT: v_add_f64 v[2:3], v[4:5], -4.0 +; GFX90A-NEXT: ds_cmpst_rtn_b64 v[2:3], v0, v[4:5], v[2:3] ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3] +; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5] ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1] ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB6_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1407,13 +1413,13 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX908-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 -; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] +; GFX908-NEXT: v_mov_b32_e32 v4, v2 +; GFX908-NEXT: v_mov_b32_e32 v3, v1 +; GFX908-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 +; GFX908-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2] -; GFX908-NEXT: v_mov_b32_e32 v1, v3 +; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[3:4] ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v2, v4 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB6_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1429,13 +1435,13 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX8-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 -; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] +; GFX8-NEXT: v_mov_b32_e32 v4, v2 +; GFX8-NEXT: v_mov_b32_e32 v3, v1 +; GFX8-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 +; GFX8-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2] -; GFX8-NEXT: v_mov_b32_e32 v1, v3 +; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[3:4] ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v2, v4 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB6_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1451,13 +1457,13 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX7-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 -; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] +; GFX7-NEXT: v_mov_b32_e32 v4, v2 +; GFX7-NEXT: v_mov_b32_e32 v3, v1 +; GFX7-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 +; GFX7-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2] -; GFX7-NEXT: v_mov_b32_e32 v1, v3 +; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[3:4] ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v2, v4 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB6_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1473,13 +1479,13 @@ define void @local_atomic_fsub_noret_f64(ptr addrspace(3) %ptr) nounwind { ; GFX6-NEXT: .LBB6_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 -; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] +; GFX6-NEXT: v_mov_b32_e32 v4, v2 +; GFX6-NEXT: v_mov_b32_e32 v3, v1 +; GFX6-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 +; GFX6-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2] -; GFX6-NEXT: v_mov_b32_e32 v1, v3 +; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[3:4] ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v2, v4 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB6_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1502,13 +1508,14 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX12-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-NEXT: s_wait_dscnt 0x0 -; GFX12-NEXT: v_add_f64_e32 v[3:4], -4.0, v[1:2] +; GFX12-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-NEXT: v_add_f64_e32 v[1:2], -4.0, v[3:4] ; GFX12-NEXT: s_wait_storecnt 0x0 -; GFX12-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528 +; GFX12-NEXT: ds_cmpstore_rtn_b64 v[1:2], v0, v[1:2], v[3:4] offset:65528 ; GFX12-NEXT: s_wait_dscnt 0x0 ; GFX12-NEXT: global_inv scope:SCOPE_SE -; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2] -; GFX12-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4 +; GFX12-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[1:2], v[3:4] ; GFX12-NEXT: s_wait_alu 0xfffe ; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-NEXT: s_wait_alu 0xfffe @@ -1527,12 +1534,12 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX942-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_add_f64 v[4:5], v[2:3], -4.0 -; GFX942-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528 +; GFX942-NEXT: v_mov_b64_e32 v[4:5], v[2:3] +; GFX942-NEXT: v_add_f64 v[2:3], v[4:5], -4.0 +; GFX942-NEXT: ds_cmpst_rtn_b64 v[2:3], v0, v[4:5], v[2:3] offset:65528 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3] +; GFX942-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5] ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b64_e32 v[2:3], v[4:5] ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB7_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1547,13 +1554,14 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX11-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 +; GFX11-NEXT: v_dual_mov_b32 v4, v2 :: v_dual_mov_b32 v3, v1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 ; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-NEXT: ds_cmpstore_rtn_b64 v[3:4], v0, v[3:4], v[1:2] offset:65528 +; GFX11-NEXT: ds_cmpstore_rtn_b64 v[1:2], v0, v[1:2], v[3:4] offset:65528 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-NEXT: buffer_gl0_inv -; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2] -; GFX11-NEXT: v_dual_mov_b32 v1, v3 :: v_dual_mov_b32 v2, v4 +; GFX11-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[1:2], v[3:4] ; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -1570,14 +1578,14 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX10-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 +; GFX10-NEXT: v_mov_b32_e32 v4, v2 +; GFX10-NEXT: v_mov_b32_e32 v3, v1 +; GFX10-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528 +; GFX10-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] offset:65528 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[3:4], v[1:2] -; GFX10-NEXT: v_mov_b32_e32 v1, v3 -; GFX10-NEXT: v_mov_b32_e32 v2, v4 +; GFX10-NEXT: v_cmp_eq_u64_e32 vcc_lo, v[1:2], v[3:4] ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB7_1 @@ -1593,12 +1601,12 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX90A-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_add_f64 v[4:5], v[2:3], -4.0 -; GFX90A-NEXT: ds_cmpst_rtn_b64 v[4:5], v0, v[2:3], v[4:5] offset:65528 +; GFX90A-NEXT: v_pk_mov_b32 v[4:5], v[2:3], v[2:3] op_sel:[0,1] +; GFX90A-NEXT: v_add_f64 v[2:3], v[4:5], -4.0 +; GFX90A-NEXT: ds_cmpst_rtn_b64 v[2:3], v0, v[4:5], v[2:3] offset:65528 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[4:5], v[2:3] +; GFX90A-NEXT: v_cmp_eq_u64_e32 vcc, v[2:3], v[4:5] ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_pk_mov_b32 v[2:3], v[4:5], v[4:5] op_sel:[0,1] ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB7_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1613,13 +1621,13 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX908-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 -; GFX908-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528 +; GFX908-NEXT: v_mov_b32_e32 v4, v2 +; GFX908-NEXT: v_mov_b32_e32 v3, v1 +; GFX908-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 +; GFX908-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] offset:65528 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2] -; GFX908-NEXT: v_mov_b32_e32 v1, v3 +; GFX908-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[3:4] ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v2, v4 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB7_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1635,13 +1643,13 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX8-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 -; GFX8-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528 +; GFX8-NEXT: v_mov_b32_e32 v4, v2 +; GFX8-NEXT: v_mov_b32_e32 v3, v1 +; GFX8-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 +; GFX8-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] offset:65528 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2] -; GFX8-NEXT: v_mov_b32_e32 v1, v3 +; GFX8-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[3:4] ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v2, v4 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB7_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1657,13 +1665,13 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX7-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_add_f64 v[3:4], v[1:2], -4.0 -; GFX7-NEXT: ds_cmpst_rtn_b64 v[3:4], v0, v[1:2], v[3:4] offset:65528 +; GFX7-NEXT: v_mov_b32_e32 v4, v2 +; GFX7-NEXT: v_mov_b32_e32 v3, v1 +; GFX7-NEXT: v_add_f64 v[1:2], v[3:4], -4.0 +; GFX7-NEXT: ds_cmpst_rtn_b64 v[1:2], v0, v[3:4], v[1:2] offset:65528 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[1:2] -; GFX7-NEXT: v_mov_b32_e32 v1, v3 +; GFX7-NEXT: v_cmp_eq_u64_e32 vcc, v[1:2], v[3:4] ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v2, v4 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB7_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1680,13 +1688,13 @@ define void @local_atomic_fsub_noret_f64__offset(ptr addrspace(3) %ptr) nounwind ; GFX6-NEXT: .LBB7_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_add_f64 v[3:4], v[0:1], -4.0 -; GFX6-NEXT: ds_cmpst_rtn_b64 v[3:4], v2, v[0:1], v[3:4] +; GFX6-NEXT: v_mov_b32_e32 v4, v1 +; GFX6-NEXT: v_mov_b32_e32 v3, v0 +; GFX6-NEXT: v_add_f64 v[0:1], v[3:4], -4.0 +; GFX6-NEXT: ds_cmpst_rtn_b64 v[0:1], v2, v[3:4], v[0:1] ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[3:4], v[0:1] -; GFX6-NEXT: v_mov_b32_e32 v0, v3 +; GFX6-NEXT: v_cmp_eq_u64_e32 vcc, v[0:1], v[3:4] ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v1, v4 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB7_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -1791,30 +1799,30 @@ define half @local_atomic_fsub_ret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-NEXT: v_and_b32_e32 v1, -4, v0 -; GFX942-NEXT: ds_read_b32 v3, v1 -; GFX942-NEXT: v_lshlrev_b32_e32 v2, 3, v0 +; GFX942-NEXT: ds_read_b32 v2, v1 +; GFX942-NEXT: v_lshlrev_b32_e32 v3, 3, v0 ; GFX942-NEXT: s_mov_b32 s0, 0xffff -; GFX942-NEXT: v_and_b32_e32 v0, 24, v2 -; GFX942-NEXT: v_lshlrev_b32_e64 v2, v2, s0 -; GFX942-NEXT: v_not_b32_e32 v2, v2 +; GFX942-NEXT: v_and_b32_e32 v0, 24, v3 +; GFX942-NEXT: v_lshlrev_b32_e64 v3, v3, s0 +; GFX942-NEXT: v_not_b32_e32 v3, v3 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB8_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX942-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX942-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX942-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX942-NEXT: v_mov_b32_e32 v4, v2 +; GFX942-NEXT: v_lshrrev_b32_e32 v2, v0, v4 +; GFX942-NEXT: v_add_f16_e32 v2, -4.0, v2 +; GFX942-NEXT: v_lshlrev_b32_e32 v2, v0, v2 +; GFX942-NEXT: v_and_or_b32 v2, v4, v3, v2 +; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB8_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX942-NEXT: s_or_b64 exec, exec, s[0:1] -; GFX942-NEXT: v_lshrrev_b32_e32 v0, v0, v4 +; GFX942-NEXT: v_lshrrev_b32_e32 v0, v0, v2 ; GFX942-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-TRUE16-LABEL: local_atomic_fsub_ret_f16: @@ -1925,30 +1933,30 @@ define half @local_atomic_fsub_ret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0 -; GFX90A-NEXT: ds_read_b32 v3, v1 -; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 3, v0 +; GFX90A-NEXT: ds_read_b32 v2, v1 +; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 3, v0 ; GFX90A-NEXT: s_mov_b32 s4, 0xffff -; GFX90A-NEXT: v_and_b32_e32 v0, 24, v2 -; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v2, s4 -; GFX90A-NEXT: v_not_b32_e32 v2, v2 +; GFX90A-NEXT: v_and_b32_e32 v0, 24, v3 +; GFX90A-NEXT: v_lshlrev_b32_e64 v3, v3, s4 +; GFX90A-NEXT: v_not_b32_e32 v3, v3 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB8_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX90A-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX90A-NEXT: v_mov_b32_e32 v4, v2 +; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v0, v4 +; GFX90A-NEXT: v_add_f16_e32 v2, -4.0, v2 +; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v0, v2 +; GFX90A-NEXT: v_and_or_b32 v2, v4, v3, v2 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB8_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5] -; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v4 +; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v2 ; GFX90A-NEXT: s_setpc_b64 s[30:31] ; ; GFX908-LABEL: local_atomic_fsub_ret_f16: @@ -2179,30 +2187,30 @@ define half @local_atomic_fsub_ret_f16__offset(ptr addrspace(3) %ptr) nounwind { ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX942-NEXT: v_add_u32_e32 v0, 0xfffe, v0 ; GFX942-NEXT: v_and_b32_e32 v1, -4, v0 -; GFX942-NEXT: ds_read_b32 v3, v1 +; GFX942-NEXT: ds_read_b32 v2, v1 ; GFX942-NEXT: v_and_b32_e32 v0, 3, v0 ; GFX942-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX942-NEXT: s_mov_b32 s0, 0xffff -; GFX942-NEXT: v_lshlrev_b32_e64 v2, v0, s0 -; GFX942-NEXT: v_not_b32_e32 v2, v2 +; GFX942-NEXT: v_lshlrev_b32_e64 v3, v0, s0 +; GFX942-NEXT: v_not_b32_e32 v3, v3 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB9_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX942-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX942-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX942-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX942-NEXT: v_mov_b32_e32 v4, v2 +; GFX942-NEXT: v_lshrrev_b32_e32 v2, v0, v4 +; GFX942-NEXT: v_add_f16_e32 v2, -4.0, v2 +; GFX942-NEXT: v_lshlrev_b32_e32 v2, v0, v2 +; GFX942-NEXT: v_and_or_b32 v2, v4, v3, v2 +; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB9_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX942-NEXT: s_or_b64 exec, exec, s[0:1] -; GFX942-NEXT: v_lshrrev_b32_e32 v0, v0, v4 +; GFX942-NEXT: v_lshrrev_b32_e32 v0, v0, v2 ; GFX942-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-TRUE16-LABEL: local_atomic_fsub_ret_f16__offset: @@ -2319,30 +2327,30 @@ define half @local_atomic_fsub_ret_f16__offset(ptr addrspace(3) %ptr) nounwind { ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) ; GFX90A-NEXT: v_add_u32_e32 v0, 0xfffe, v0 ; GFX90A-NEXT: v_and_b32_e32 v1, -4, v0 -; GFX90A-NEXT: ds_read_b32 v3, v1 +; GFX90A-NEXT: ds_read_b32 v2, v1 ; GFX90A-NEXT: v_and_b32_e32 v0, 3, v0 ; GFX90A-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX90A-NEXT: s_mov_b32 s4, 0xffff -; GFX90A-NEXT: v_lshlrev_b32_e64 v2, v0, s4 -; GFX90A-NEXT: v_not_b32_e32 v2, v2 +; GFX90A-NEXT: v_lshlrev_b32_e64 v3, v0, s4 +; GFX90A-NEXT: v_not_b32_e32 v3, v3 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB9_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX90A-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX90A-NEXT: v_mov_b32_e32 v4, v2 +; GFX90A-NEXT: v_lshrrev_b32_e32 v2, v0, v4 +; GFX90A-NEXT: v_add_f16_e32 v2, -4.0, v2 +; GFX90A-NEXT: v_lshlrev_b32_e32 v2, v0, v2 +; GFX90A-NEXT: v_and_or_b32 v2, v4, v3, v2 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v1, v4, v2 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB9_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5] -; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v4 +; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v2 ; GFX90A-NEXT: s_setpc_b64 s[30:31] ; ; GFX908-LABEL: local_atomic_fsub_ret_f16__offset: @@ -2494,27 +2502,27 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: ds_load_b32 v2, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX12-TRUE16-NEXT: ds_load_b32 v3, v1 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, 24, v0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v3, v3 +; GFX12-TRUE16-NEXT: v_not_b32_e32 v2, v2 ; GFX12-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX12-TRUE16-NEXT: v_add_f16_e32 v4.l, -4.0, v4.l +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 +; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, -4.0, v3.l ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2 +; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -2535,28 +2543,28 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX12-FAKE16-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX12-FAKE16-NEXT: s_mov_b32 s0, 0 -; GFX12-FAKE16-NEXT: ds_load_b32 v2, v1 -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX12-FAKE16-NEXT: ds_load_b32 v3, v1 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX12-FAKE16-NEXT: v_and_b32_e32 v0, 24, v0 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX12-FAKE16-NEXT: v_not_b32_e32 v3, v3 +; GFX12-FAKE16-NEXT: v_not_b32_e32 v2, v2 ; GFX12-FAKE16-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX12-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 -; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX12-FAKE16-NEXT: v_add_f16_e32 v4, -4.0, v4 +; GFX12-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4 +; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX12-FAKE16-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX12-FAKE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX12-FAKE16-NEXT: s_wait_storecnt 0x0 -; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2 +; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 ; GFX12-FAKE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX12-FAKE16-NEXT: v_mov_b32_e32 v2, v4 +; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX12-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe @@ -2581,15 +2589,15 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX942-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX942-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX942-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX942-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX942-NEXT: v_mov_b32_e32 v4, v3 +; GFX942-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX942-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX942-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX942-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB10_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -2602,27 +2610,27 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: ds_load_b32 v2, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX11-TRUE16-NEXT: ds_load_b32 v3, v1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 24, v0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_not_b32_e32 v3, v3 +; GFX11-TRUE16-NEXT: v_not_b32_e32 v2, v2 ; GFX11-TRUE16-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX11-TRUE16-NEXT: v_add_f16_e32 v4.l, -4.0, v4.l +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 +; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, -4.0, v3.l ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2 +; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -2637,28 +2645,28 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX11-FAKE16-NEXT: s_mov_b32 s0, 0 -; GFX11-FAKE16-NEXT: ds_load_b32 v2, v1 -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX11-FAKE16-NEXT: ds_load_b32 v3, v1 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 24, v0 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-FAKE16-NEXT: v_not_b32_e32 v3, v3 +; GFX11-FAKE16-NEXT: v_not_b32_e32 v2, v2 ; GFX11-FAKE16-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX11-FAKE16-NEXT: v_add_f16_e32 v4, -4.0, v4 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX11-FAKE16-NEXT: v_add_f16_e32 v3, -4.0, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4 +; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX11-FAKE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX11-FAKE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2 +; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-FAKE16-NEXT: buffer_gl0_inv -; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, v4 +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -2673,23 +2681,23 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX10-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX10-NEXT: s_mov_b32 s4, 0 -; GFX10-NEXT: ds_read_b32 v2, v1 -; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX10-NEXT: ds_read_b32 v3, v1 +; GFX10-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX10-NEXT: v_and_b32_e32 v0, 24, v0 -; GFX10-NEXT: v_not_b32_e32 v3, v3 +; GFX10-NEXT: v_not_b32_e32 v2, v2 ; GFX10-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX10-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 -; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX10-NEXT: v_mov_b32_e32 v4, v3 +; GFX10-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX10-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 +; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4 +; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX10-NEXT: v_mov_b32_e32 v2, v4 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB10_1 @@ -2711,15 +2719,15 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX90A-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX90A-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX90A-NEXT: v_mov_b32_e32 v4, v3 +; GFX90A-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX90A-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX90A-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB10_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -2740,15 +2748,15 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX908-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX908-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX908-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX908-NEXT: v_mov_b32_e32 v4, v3 +; GFX908-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX908-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX908-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v3, v4 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB10_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -2770,16 +2778,16 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX8-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX8-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX8-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX8-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX8-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX8-NEXT: v_mov_b32_e32 v4, v3 +; GFX8-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX8-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX8-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX8-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX8-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v3, v4 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB10_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -2800,18 +2808,18 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX7-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4 -; GFX7-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX7-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4 -; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX7-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX7-NEXT: v_mov_b32_e32 v4, v3 +; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3 +; GFX7-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX7-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3 +; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX7-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v3, v4 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB10_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -2832,18 +2840,18 @@ define void @local_atomic_fsub_noret_f16(ptr addrspace(3) %ptr) nounwind { ; GFX6-NEXT: .LBB10_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4 -; GFX6-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX6-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4 -; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX6-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX6-NEXT: v_mov_b32_e32 v4, v3 +; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3 +; GFX6-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX6-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3 +; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX6-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v3, v4 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB10_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -2874,19 +2882,19 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX12-TRUE16-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f16_e32 v4.l, -4.0, v4.l -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 +; GFX12-TRUE16-NEXT: v_add_f16_e32 v3.l, -4.0, v3.l +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -2917,19 +2925,20 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX12-FAKE16-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX12-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 -; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3 +; GFX12-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX12-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX12-FAKE16-NEXT: v_add_f16_e32 v3, -4.0, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX12-FAKE16-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX12-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX12-FAKE16-NEXT: s_wait_storecnt 0x0 -; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 ; GFX12-FAKE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX12-FAKE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX12-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe @@ -2955,15 +2964,15 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX942-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX942-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX942-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX942-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX942-NEXT: v_mov_b32_e32 v4, v3 +; GFX942-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX942-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX942-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX942-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB11_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -2986,19 +2995,19 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX11-TRUE16-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v4.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f16_e32 v4.l, -4.0, v4.l -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 +; GFX11-TRUE16-NEXT: v_add_f16_e32 v3.l, -4.0, v3.l +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -3023,19 +3032,20 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX11-FAKE16-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX11-FAKE16-NEXT: v_and_b32_e32 v4, 0xffff, v4 +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX11-FAKE16-NEXT: v_add_f16_e32 v3, -4.0, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX11-FAKE16-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX11-FAKE16-NEXT: v_and_b32_e32 v3, 0xffff, v3 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX11-FAKE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-FAKE16-NEXT: buffer_gl0_inv -; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -3058,16 +3068,16 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX10-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX10-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 -; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX10-NEXT: v_mov_b32_e32 v4, v3 +; GFX10-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX10-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_0 +; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX10-NEXT: v_mov_b32_e32 v3, v4 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB11_1 @@ -3090,15 +3100,15 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX90A-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX90A-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX90A-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX90A-NEXT: v_mov_b32_e32 v4, v3 +; GFX90A-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX90A-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX90A-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB11_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3120,15 +3130,15 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX908-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX908-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX908-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX908-NEXT: v_mov_b32_e32 v4, v3 +; GFX908-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX908-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX908-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v3, v4 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB11_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3151,16 +3161,16 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX8-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX8-NEXT: v_add_f16_e32 v4, -4.0, v4 -; GFX8-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX8-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX8-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX8-NEXT: v_mov_b32_e32 v4, v3 +; GFX8-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX8-NEXT: v_add_f16_e32 v3, -4.0, v3 +; GFX8-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX8-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX8-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v3, v4 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB11_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3182,18 +3192,18 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX7-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX7-NEXT: v_cvt_f32_f16_e32 v4, v4 -; GFX7-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX7-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX7-NEXT: v_cvt_f16_f32_e32 v4, v4 -; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX7-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX7-NEXT: v_mov_b32_e32 v4, v3 +; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX7-NEXT: v_cvt_f32_f16_e32 v3, v3 +; GFX7-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX7-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX7-NEXT: v_cvt_f16_f32_e32 v3, v3 +; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX7-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v3, v4 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB11_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3215,18 +3225,18 @@ define void @local_atomic_fsub_noret_f16__offset(ptr addrspace(3) %ptr) nounwind ; GFX6-NEXT: .LBB11_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX6-NEXT: v_cvt_f32_f16_e32 v4, v4 -; GFX6-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX6-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX6-NEXT: v_cvt_f16_f32_e32 v4, v4 -; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX6-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX6-NEXT: v_mov_b32_e32 v4, v3 +; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX6-NEXT: v_cvt_f32_f16_e32 v3, v3 +; GFX6-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX6-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX6-NEXT: v_cvt_f16_f32_e32 v3, v3 +; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX6-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v3, v4 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB11_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3308,19 +3318,19 @@ define half @local_atomic_fsub_ret_f16__offset__align4(ptr addrspace(3) %ptr) no ; GFX942-LABEL: local_atomic_fsub_ret_f16__offset__align4: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v2, v0 offset:65534 +; GFX942-NEXT: ds_read_b32 v1, v0 offset:65534 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: s_mov_b32 s2, 0xffff0000 ; GFX942-NEXT: .LBB12_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: v_add_f16_e32 v1, -4.0, v2 ; GFX942-NEXT: v_and_or_b32 v1, v2, s2, v1 ; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB12_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3411,19 +3421,19 @@ define half @local_atomic_fsub_ret_f16__offset__align4(ptr addrspace(3) %ptr) no ; GFX90A-LABEL: local_atomic_fsub_ret_f16__offset__align4: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65534 +; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: s_mov_b32 s6, 0xffff0000 ; GFX90A-NEXT: .LBB12_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: v_add_f16_e32 v1, -4.0, v2 ; GFX90A-NEXT: v_and_or_b32 v1, v2, s6, v1 ; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB12_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3548,16 +3558,16 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX12-TRUE16-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_add_f16_e32 v2.l, -4.0, v1.l -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v1 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add_f16_e32 v1.l, -4.0, v2.l +; GFX12-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534 +; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -3580,16 +3590,17 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX12-FAKE16-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX12-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 -; GFX12-FAKE16-NEXT: v_add_f16_e32 v2, -4.0, v1 +; GFX12-FAKE16-NEXT: v_mov_b32_e32 v2, v1 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX12-FAKE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2 +; GFX12-FAKE16-NEXT: v_add_f16_e32 v1, -4.0, v2 +; GFX12-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1 ; GFX12-FAKE16-NEXT: s_wait_storecnt 0x0 -; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534 +; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 ; GFX12-FAKE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-FAKE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX12-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe @@ -3609,13 +3620,13 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX942-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_add_f16_e32 v2, -4.0, v1 -; GFX942-NEXT: v_and_or_b32 v2, v1, s2, v2 -; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX942-NEXT: v_mov_b32_e32 v2, v1 +; GFX942-NEXT: v_add_f16_e32 v1, -4.0, v2 +; GFX942-NEXT: v_and_or_b32 v1, v2, s2, v1 +; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v1, v2 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB13_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3630,16 +3641,16 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX11-TRUE16-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_add_f16_e32 v2.l, -4.0, v1.l -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v2.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v1 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v1.h, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add_f16_e32 v1.l, -4.0, v2.l +; GFX11-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534 +; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -3656,16 +3667,17 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX11-FAKE16-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FAKE16-NEXT: v_add_f16_e32 v2, -4.0, v1 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, v1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX11-FAKE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2 +; GFX11-FAKE16-NEXT: v_add_f16_e32 v1, -4.0, v2 +; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, 0xffff, v1 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1 ; GFX11-FAKE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534 +; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-FAKE16-NEXT: buffer_gl0_inv -; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX11-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -3682,15 +3694,15 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX10-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_add_f16_e32 v2, -4.0, v1 -; GFX10-NEXT: v_and_b32_e32 v2, 0xffff, v2 -; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2 +; GFX10-NEXT: v_mov_b32_e32 v2, v1 +; GFX10-NEXT: v_add_f16_e32 v1, -4.0, v2 +; GFX10-NEXT: v_and_b32_e32 v1, 0xffff, v1 +; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX10-NEXT: v_mov_b32_e32 v1, v2 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB13_1 @@ -3707,13 +3719,13 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX90A-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_add_f16_e32 v2, -4.0, v1 -; GFX90A-NEXT: v_and_or_b32 v2, v1, s6, v2 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 +; GFX90A-NEXT: v_add_f16_e32 v1, -4.0, v2 +; GFX90A-NEXT: v_and_or_b32 v1, v2, s6, v1 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v1, v2 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB13_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3729,13 +3741,13 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX908-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_add_f16_e32 v2, -4.0, v1 -; GFX908-NEXT: v_and_or_b32 v2, v1, s6, v2 -; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX908-NEXT: v_mov_b32_e32 v2, v1 +; GFX908-NEXT: v_add_f16_e32 v1, -4.0, v2 +; GFX908-NEXT: v_and_or_b32 v1, v2, s6, v1 +; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v1, v2 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB13_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3751,14 +3763,14 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX8-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_add_f16_e32 v2, -4.0, v1 -; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX8-NEXT: v_or_b32_e32 v2, v3, v2 -; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX8-NEXT: v_mov_b32_e32 v2, v1 +; GFX8-NEXT: v_add_f16_e32 v1, -4.0, v2 +; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX8-NEXT: v_or_b32_e32 v1, v3, v1 +; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v1, v2 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB13_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3774,16 +3786,16 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX7-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cvt_f32_f16_e32 v2, v1 -; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX7-NEXT: v_cvt_f16_f32_e32 v2, v2 -; GFX7-NEXT: v_or_b32_e32 v2, v3, v2 -; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX7-NEXT: v_mov_b32_e32 v2, v1 +; GFX7-NEXT: v_cvt_f32_f16_e32 v1, v2 +; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX7-NEXT: v_cvt_f16_f32_e32 v1, v1 +; GFX7-NEXT: v_or_b32_e32 v1, v3, v1 +; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v1, v2 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB13_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3800,16 +3812,16 @@ define void @local_atomic_fsub_noret_f16__offset__align4(ptr addrspace(3) %ptr) ; GFX6-NEXT: .LBB13_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cvt_f32_f16_e32 v2, v1 -; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX6-NEXT: v_cvt_f16_f32_e32 v2, v2 -; GFX6-NEXT: v_or_b32_e32 v2, v3, v2 -; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX6-NEXT: v_mov_b32_e32 v2, v1 +; GFX6-NEXT: v_cvt_f32_f16_e32 v1, v2 +; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX6-NEXT: v_cvt_f16_f32_e32 v1, v1 +; GFX6-NEXT: v_or_b32_e32 v1, v3, v1 +; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v1, v2 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB13_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -3945,27 +3957,27 @@ define bfloat @local_atomic_fsub_ret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX942-NEXT: .LBB14_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX942-NEXT: v_mov_b32_e32 v4, v3 +; GFX942-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD ; GFX942-NEXT: s_nop 0 -; GFX942-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX942-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX942-NEXT: v_add3_u32 v5, v5, v4, s2 -; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 +; GFX942-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX942-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX942-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX942-NEXT: v_add3_u32 v5, v5, v3, s2 +; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 ; GFX942-NEXT: s_nop 1 -; GFX942-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX942-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX942-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX942-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX942-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX942-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB14_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX942-NEXT: s_or_b64 exec, exec, s[0:1] -; GFX942-NEXT: v_lshrrev_b32_e32 v0, v0, v4 +; GFX942-NEXT: v_lshrrev_b32_e32 v0, v0, v3 ; GFX942-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-TRUE16-LABEL: local_atomic_fsub_ret_bf16: @@ -4112,25 +4124,25 @@ define bfloat @local_atomic_fsub_ret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX90A-NEXT: .LBB14_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX90A-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6 -; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 -; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX90A-NEXT: v_mov_b32_e32 v4, v3 +; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX90A-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6 +; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 +; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB14_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5] -; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v4 +; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3 ; GFX90A-NEXT: s_setpc_b64 s[30:31] ; ; GFX908-LABEL: local_atomic_fsub_ret_bf16: @@ -4404,27 +4416,27 @@ define bfloat @local_atomic_fsub_ret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX942-NEXT: .LBB15_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX942-NEXT: v_mov_b32_e32 v4, v3 +; GFX942-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD ; GFX942-NEXT: s_nop 0 -; GFX942-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX942-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX942-NEXT: v_add3_u32 v5, v5, v4, s2 -; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 +; GFX942-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX942-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX942-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX942-NEXT: v_add3_u32 v5, v5, v3, s2 +; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 ; GFX942-NEXT: s_nop 1 -; GFX942-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX942-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX942-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX942-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX942-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX942-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB15_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX942-NEXT: s_or_b64 exec, exec, s[0:1] -; GFX942-NEXT: v_lshrrev_b32_e32 v0, v0, v4 +; GFX942-NEXT: v_lshrrev_b32_e32 v0, v0, v3 ; GFX942-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-TRUE16-LABEL: local_atomic_fsub_ret_bf16__offset: @@ -4577,25 +4589,25 @@ define bfloat @local_atomic_fsub_ret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX90A-NEXT: .LBB15_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX90A-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6 -; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 -; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX90A-NEXT: v_mov_b32_e32 v4, v3 +; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX90A-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6 +; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 +; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB15_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX90A-NEXT: s_or_b64 exec, exec, s[4:5] -; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v4 +; GFX90A-NEXT: v_lshrrev_b32_e32 v0, v0, v3 ; GFX90A-NEXT: s_setpc_b64 s[30:31] ; ; GFX908-LABEL: local_atomic_fsub_ret_bf16__offset: @@ -4759,38 +4771,38 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX12-TRUE16-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX12-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX12-TRUE16-NEXT: ds_load_b32 v2, v1 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX12-TRUE16-NEXT: ds_load_b32 v3, v1 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX12-TRUE16-NEXT: v_and_b32_e32 v0, 24, v0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_not_b32_e32 v3, v3 +; GFX12-TRUE16-NEXT: v_not_b32_e32 v2, v2 ; GFX12-TRUE16-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff +; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v5 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v5 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2 +; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -4811,37 +4823,37 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX12-FAKE16-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX12-FAKE16-NEXT: s_mov_b32 s0, 0 -; GFX12-FAKE16-NEXT: ds_load_b32 v2, v1 -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX12-FAKE16-NEXT: ds_load_b32 v3, v1 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX12-FAKE16-NEXT: v_and_b32_e32 v0, 24, v0 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX12-FAKE16-NEXT: v_not_b32_e32 v3, v3 +; GFX12-FAKE16-NEXT: v_not_b32_e32 v2, v2 ; GFX12-FAKE16-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX12-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 -; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 +; GFX12-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX12-FAKE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX12-FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX12-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX12-FAKE16-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX12-FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX12-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff +; GFX12-FAKE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffd -; GFX12-FAKE16-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo +; GFX12-FAKE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4 +; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX12-FAKE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX12-FAKE16-NEXT: s_wait_storecnt 0x0 -; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2 +; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 ; GFX12-FAKE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX12-FAKE16-NEXT: v_mov_b32_e32 v2, v4 +; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX12-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe @@ -4867,22 +4879,22 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX942-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX942-NEXT: v_mov_b32_e32 v4, v3 +; GFX942-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD ; GFX942-NEXT: s_nop 0 -; GFX942-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX942-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX942-NEXT: v_add3_u32 v5, v5, v4, s2 -; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 +; GFX942-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX942-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX942-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX942-NEXT: v_add3_u32 v5, v5, v3, s2 +; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 ; GFX942-NEXT: s_nop 1 -; GFX942-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX942-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX942-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX942-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX942-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX942-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB16_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -4895,38 +4907,38 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX11-TRUE16-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX11-TRUE16-NEXT: s_mov_b32 s0, 0 -; GFX11-TRUE16-NEXT: ds_load_b32 v2, v1 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX11-TRUE16-NEXT: ds_load_b32 v3, v1 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX11-TRUE16-NEXT: v_and_b32_e32 v0, 24, v0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_not_b32_e32 v3, v3 +; GFX11-TRUE16-NEXT: v_not_b32_e32 v2, v2 ; GFX11-TRUE16-NEXT: .p2align 6 ; GFX11-TRUE16-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v0, v5 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v0, v5 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2 +; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -4941,37 +4953,37 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX11-FAKE16-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX11-FAKE16-NEXT: s_mov_b32 s0, 0 -; GFX11-FAKE16-NEXT: ds_load_b32 v2, v1 -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX11-FAKE16-NEXT: ds_load_b32 v3, v1 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX11-FAKE16-NEXT: v_and_b32_e32 v0, 24, v0 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) -; GFX11-FAKE16-NEXT: v_not_b32_e32 v3, v3 +; GFX11-FAKE16-NEXT: v_not_b32_e32 v2, v2 ; GFX11-FAKE16-NEXT: .p2align 6 ; GFX11-FAKE16-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v4, v0, v2 -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX11-FAKE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX11-FAKE16-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo +; GFX11-FAKE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v4, v0, v4 +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v3, v0, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX11-FAKE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX11-FAKE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v1, v4, v2 +; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v1, v3, v4 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-FAKE16-NEXT: buffer_gl0_inv -; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, v4 +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -4986,28 +4998,28 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX10-NEXT: v_and_b32_e32 v1, -4, v0 ; GFX10-NEXT: v_lshlrev_b32_e32 v0, 3, v0 ; GFX10-NEXT: s_mov_b32 s4, 0 -; GFX10-NEXT: ds_read_b32 v2, v1 -; GFX10-NEXT: v_lshlrev_b32_e64 v3, v0, 0xffff +; GFX10-NEXT: ds_read_b32 v3, v1 +; GFX10-NEXT: v_lshlrev_b32_e64 v2, v0, 0xffff ; GFX10-NEXT: v_and_b32_e32 v0, 24, v0 -; GFX10-NEXT: v_not_b32_e32 v3, v3 +; GFX10-NEXT: v_not_b32_e32 v2, v2 ; GFX10-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v0, v2 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX10-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo -; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX10-NEXT: v_and_or_b32 v4, v2, v3, v4 +; GFX10-NEXT: v_mov_b32_e32 v4, v3 +; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX10-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo +; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v1, v2, v4 +; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v2 -; GFX10-NEXT: v_mov_b32_e32 v2, v4 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB16_1 @@ -5030,20 +5042,20 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX90A-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX90A-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6 -; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 -; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX90A-NEXT: v_mov_b32_e32 v4, v3 +; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX90A-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6 +; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 +; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB16_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5065,20 +5077,20 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX908-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX908-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6 -; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 -; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX908-NEXT: v_mov_b32_e32 v4, v3 +; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX908-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6 +; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 +; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v3, v4 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB16_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5100,22 +5112,22 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX8-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v0, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX8-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4 +; GFX8-NEXT: v_mov_b32_e32 v4, v3 +; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v0, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX8-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3 ; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6 -; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4 -; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 -; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc -; GFX8-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX8-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3 +; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 +; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc +; GFX8-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX8-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v3, v4 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB16_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5136,18 +5148,18 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX7-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX7-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; GFX7-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX7-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX7-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX7-NEXT: v_mov_b32_e32 v4, v3 +; GFX7-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX7-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX7-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX7-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX7-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v3, v4 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB16_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5168,18 +5180,18 @@ define void @local_atomic_fsub_noret_bf16(ptr addrspace(3) %ptr) nounwind { ; GFX6-NEXT: .LBB16_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_lshrrev_b32_e32 v4, v0, v3 -; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX6-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; GFX6-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX6-NEXT: v_lshlrev_b32_e32 v4, v0, v4 -; GFX6-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v1, v3, v4 +; GFX6-NEXT: v_mov_b32_e32 v4, v3 +; GFX6-NEXT: v_lshrrev_b32_e32 v3, v0, v4 +; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX6-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX6-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX6-NEXT: v_lshlrev_b32_e32 v3, v0, v3 +; GFX6-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v1, v4, v3 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v3, v4 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB16_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5210,29 +5222,30 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX12-TRUE16-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-TRUE16-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff +; GFX12-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX12-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v5 -; GFX12-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -5263,28 +5276,29 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX12-FAKE16-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX12-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 -; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3 +; GFX12-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX12-FAKE16-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX12-FAKE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX12-FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX12-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX12-FAKE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX12-FAKE16-NEXT: s_wait_alu 0xfffd +; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo -; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v4, 16, v4 +; GFX12-FAKE16-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX12-FAKE16-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX12-FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX12-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX12-FAKE16-NEXT: s_wait_alu 0xfffd +; GFX12-FAKE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX12-FAKE16-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX12-FAKE16-NEXT: s_wait_storecnt 0x0 -; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 ; GFX12-FAKE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX12-FAKE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX12-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe @@ -5311,22 +5325,22 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX942-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX942-NEXT: v_mov_b32_e32 v4, v3 +; GFX942-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD ; GFX942-NEXT: s_nop 0 -; GFX942-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX942-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX942-NEXT: v_add3_u32 v5, v5, v4, s2 -; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 +; GFX942-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX942-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX942-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX942-NEXT: v_add3_u32 v5, v5, v3, s2 +; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 ; GFX942-NEXT: s_nop 1 -; GFX942-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX942-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX942-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX942-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX942-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX942-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB17_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5350,28 +5364,29 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX11-TRUE16-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-TRUE16-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX11-TRUE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v4, v1, v5 -; GFX11-TRUE16-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX11-TRUE16-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.h, 0 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, v1, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -5397,27 +5412,28 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX11-FAKE16-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v4, v1, v3 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX11-FAKE16-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) -; GFX11-FAKE16-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX11-FAKE16-NEXT: v_add3_u32 v5, v5, v4, 0x7fff +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo -; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v4, 16, v4 +; GFX11-FAKE16-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX11-FAKE16-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX11-FAKE16-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX11-FAKE16-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX11-FAKE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-FAKE16-NEXT: buffer_gl0_inv -; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -5440,21 +5456,21 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX10-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX10-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX10-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 -; GFX10-NEXT: v_add3_u32 v5, v5, v4, 0x7fff -; GFX10-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc_lo -; GFX10-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX10-NEXT: v_and_or_b32 v4, v3, v2, v4 +; GFX10-NEXT: v_mov_b32_e32 v4, v3 +; GFX10-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX10-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX10-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX10-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX10-NEXT: v_add3_u32 v5, v5, v3, 0x7fff +; GFX10-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc_lo +; GFX10-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX10-NEXT: v_and_or_b32 v3, v4, v2, v3 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX10-NEXT: v_mov_b32_e32 v3, v4 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB17_1 @@ -5478,20 +5494,20 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX90A-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX90A-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX90A-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX90A-NEXT: v_add3_u32 v5, v5, v4, s6 -; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 -; GFX90A-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX90A-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX90A-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX90A-NEXT: v_mov_b32_e32 v4, v3 +; GFX90A-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX90A-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX90A-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX90A-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX90A-NEXT: v_add3_u32 v5, v5, v3, s6 +; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 +; GFX90A-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX90A-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX90A-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB17_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5514,20 +5530,20 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX908-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX908-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX908-NEXT: v_bfe_u32 v5, v4, 16, 1 -; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v4 -; GFX908-NEXT: v_add3_u32 v5, v5, v4, s6 -; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 -; GFX908-NEXT: v_cndmask_b32_e32 v4, v5, v6, vcc -; GFX908-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX908-NEXT: v_and_or_b32 v4, v3, v2, v4 -; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX908-NEXT: v_mov_b32_e32 v4, v3 +; GFX908-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX908-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX908-NEXT: v_bfe_u32 v5, v3, 16, 1 +; GFX908-NEXT: v_or_b32_e32 v6, 0x400000, v3 +; GFX908-NEXT: v_add3_u32 v5, v5, v3, s6 +; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 +; GFX908-NEXT: v_cndmask_b32_e32 v3, v5, v6, vcc +; GFX908-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX908-NEXT: v_and_or_b32 v3, v4, v2, v3 +; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v3, v4 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB17_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5550,22 +5566,22 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX8-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_lshrrev_b32_sdwa v4, v1, v3 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD -; GFX8-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4 +; GFX8-NEXT: v_mov_b32_e32 v4, v3 +; GFX8-NEXT: v_lshrrev_b32_sdwa v3, v1, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD +; GFX8-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3 ; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6 -; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4 -; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v4, v4 -; GFX8-NEXT: v_cndmask_b32_e32 v4, v6, v7, vcc -; GFX8-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX8-NEXT: v_lshlrev_b32_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX8-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3 +; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v3, v3 +; GFX8-NEXT: v_cndmask_b32_e32 v3, v6, v7, vcc +; GFX8-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX8-NEXT: v_lshlrev_b32_sdwa v3, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX8-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v3, v4 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB17_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5587,18 +5603,18 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX7-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX7-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX7-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX7-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; GFX7-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX7-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX7-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX7-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX7-NEXT: v_mov_b32_e32 v4, v3 +; GFX7-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX7-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX7-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX7-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX7-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX7-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX7-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX7-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v3, v4 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB17_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5620,18 +5636,18 @@ define void @local_atomic_fsub_noret_bf16__offset(ptr addrspace(3) %ptr) nounwin ; GFX6-NEXT: .LBB17_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_lshrrev_b32_e32 v4, v1, v3 -; GFX6-NEXT: v_lshlrev_b32_e32 v4, 16, v4 -; GFX6-NEXT: v_add_f32_e32 v4, -4.0, v4 -; GFX6-NEXT: v_lshrrev_b32_e32 v4, 16, v4 -; GFX6-NEXT: v_and_b32_e32 v5, v3, v2 -; GFX6-NEXT: v_lshlrev_b32_e32 v4, v1, v4 -; GFX6-NEXT: v_or_b32_e32 v4, v5, v4 -; GFX6-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX6-NEXT: v_mov_b32_e32 v4, v3 +; GFX6-NEXT: v_lshrrev_b32_e32 v3, v1, v4 +; GFX6-NEXT: v_lshlrev_b32_e32 v3, 16, v3 +; GFX6-NEXT: v_add_f32_e32 v3, -4.0, v3 +; GFX6-NEXT: v_lshrrev_b32_e32 v3, 16, v3 +; GFX6-NEXT: v_and_b32_e32 v5, v4, v2 +; GFX6-NEXT: v_lshlrev_b32_e32 v3, v1, v3 +; GFX6-NEXT: v_or_b32_e32 v3, v5, v3 +; GFX6-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v3, v4 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB17_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5733,13 +5749,14 @@ define bfloat @local_atomic_fsub_ret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX942-LABEL: local_atomic_fsub_ret_bf16__offset__align4: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v2, v0 offset:65534 +; GFX942-NEXT: ds_read_b32 v1, v0 offset:65534 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: s_movk_i32 s2, 0x7fff ; GFX942-NEXT: s_mov_b32 s3, 0xffff0000 ; GFX942-NEXT: .LBB18_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: v_lshlrev_b32_e32 v1, 16, v2 ; GFX942-NEXT: v_add_f32_e32 v1, -4.0, v1 ; GFX942-NEXT: v_bfe_u32 v3, v1, 16, 1 @@ -5754,7 +5771,6 @@ define bfloat @local_atomic_fsub_ret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB18_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -5871,13 +5887,14 @@ define bfloat @local_atomic_fsub_ret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX90A-LABEL: local_atomic_fsub_ret_bf16__offset__align4: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65534 +; GFX90A-NEXT: ds_read_b32 v1, v0 offset:65534 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: s_movk_i32 s6, 0x7fff ; GFX90A-NEXT: s_mov_b32 s7, 0xffff0000 ; GFX90A-NEXT: .LBB18_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2 ; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v1 ; GFX90A-NEXT: v_bfe_u32 v3, v1, 16, 1 @@ -5891,7 +5908,6 @@ define bfloat @local_atomic_fsub_ret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB18_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6031,26 +6047,27 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX12-TRUE16-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v2, v1 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX12-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v2 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX12-TRUE16-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX12-TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo ; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.h -; GFX12-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v3 +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v3 ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534 +; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -6073,25 +6090,26 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX12-FAKE16-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX12-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 -; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 +; GFX12-FAKE16-NEXT: v_mov_b32_e32 v2, v1 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX12-FAKE16-NEXT: v_bfe_u32 v3, v2, 16, 1 -; GFX12-FAKE16-NEXT: v_or_b32_e32 v4, 0x400000, v2 -; GFX12-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX12-FAKE16-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX12-FAKE16-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX12-FAKE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX12-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX12-FAKE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffd -; GFX12-FAKE16-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX12-FAKE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2 +; GFX12-FAKE16-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo +; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1 ; GFX12-FAKE16-NEXT: s_wait_storecnt 0x0 -; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534 +; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 ; GFX12-FAKE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-FAKE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX12-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe @@ -6112,21 +6130,21 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX942-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX942-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX942-NEXT: v_bfe_u32 v3, v2, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v4, 0x400000, v2 -; GFX942-NEXT: v_add3_u32 v3, v3, v2, s2 -; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 +; GFX942-NEXT: v_mov_b32_e32 v2, v1 +; GFX942-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX942-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX942-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX942-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX942-NEXT: v_add3_u32 v3, v3, v1, s2 +; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 ; GFX942-NEXT: s_nop 1 -; GFX942-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc -; GFX942-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX942-NEXT: v_and_or_b32 v2, v1, s3, v2 -; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX942-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc +; GFX942-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; GFX942-NEXT: v_and_or_b32 v1, v2, s3, v1 +; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v1, v2 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB19_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6142,25 +6160,26 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX11-TRUE16-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v2, v1 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v2, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 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(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff -; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX11-TRUE16-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-TRUE16-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-TRUE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2) +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo ; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.h, 0 -; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v2.h -; GFX11-TRUE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v3 +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v3.l, v1.h +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v3 ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534 +; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -6178,24 +6197,25 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX11-FAKE16-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v2, 16, v1 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v2, v1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX11-FAKE16-NEXT: v_bfe_u32 v3, v2, 16, 1 -; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, 0x400000, v2 -; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_add3_u32 v3, v3, v2, 0x7fff -; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX11-FAKE16-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_3) +; GFX11-FAKE16-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX11-FAKE16-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX11-FAKE16-NEXT: v_add3_u32 v3, v3, v1, 0x7fff ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX11-FAKE16-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2 +; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo +; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1 ; GFX11-FAKE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 offset:65534 +; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:65534 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-FAKE16-NEXT: buffer_gl0_inv -; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-FAKE16-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX11-FAKE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -6212,21 +6232,21 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX10-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX10-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX10-NEXT: v_bfe_u32 v3, v2, 16, 1 -; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v2 -; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v2, v2 -; GFX10-NEXT: v_add3_u32 v3, v3, v2, 0x7fff -; GFX10-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc_lo -; GFX10-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX10-NEXT: v_and_or_b32 v2, 0xffff0000, v1, v2 +; GFX10-NEXT: v_mov_b32_e32 v2, v1 +; GFX10-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX10-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX10-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX10-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v1, v1 +; GFX10-NEXT: v_add3_u32 v3, v3, v1, 0x7fff +; GFX10-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc_lo +; GFX10-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; GFX10-NEXT: v_and_or_b32 v1, 0xffff0000, v2, v1 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX10-NEXT: v_mov_b32_e32 v1, v2 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB19_1 @@ -6244,20 +6264,20 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX90A-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX90A-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX90A-NEXT: v_bfe_u32 v3, v2, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v2 -; GFX90A-NEXT: v_add3_u32 v3, v3, v2, s6 -; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 -; GFX90A-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc -; GFX90A-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX90A-NEXT: v_and_or_b32 v2, v1, s7, v2 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 +; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX90A-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX90A-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX90A-NEXT: v_add3_u32 v3, v3, v1, s6 +; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 +; GFX90A-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc +; GFX90A-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; GFX90A-NEXT: v_and_or_b32 v1, v2, s7, v1 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v1, v2 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB19_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6274,20 +6294,20 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX908-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX908-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX908-NEXT: v_bfe_u32 v3, v2, 16, 1 -; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v2 -; GFX908-NEXT: v_add3_u32 v3, v3, v2, s6 -; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 -; GFX908-NEXT: v_cndmask_b32_e32 v2, v3, v4, vcc -; GFX908-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX908-NEXT: v_and_or_b32 v2, v1, s7, v2 -; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX908-NEXT: v_mov_b32_e32 v2, v1 +; GFX908-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX908-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX908-NEXT: v_bfe_u32 v3, v1, 16, 1 +; GFX908-NEXT: v_or_b32_e32 v4, 0x400000, v1 +; GFX908-NEXT: v_add3_u32 v3, v3, v1, s6 +; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 +; GFX908-NEXT: v_cndmask_b32_e32 v1, v3, v4, vcc +; GFX908-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; GFX908-NEXT: v_and_or_b32 v1, v2, s7, v1 +; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v1, v2 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB19_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6303,21 +6323,21 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX8-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX8-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX8-NEXT: v_bfe_u32 v4, v2, 16, 1 -; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v2 +; GFX8-NEXT: v_mov_b32_e32 v2, v1 +; GFX8-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX8-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX8-NEXT: v_bfe_u32 v4, v1, 16, 1 +; GFX8-NEXT: v_add_u32_e32 v4, vcc, v4, v1 ; GFX8-NEXT: v_add_u32_e32 v4, vcc, 0x7fff, v4 -; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v2 -; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v2, v2 -; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX8-NEXT: v_cndmask_b32_e32 v2, v4, v5, vcc -; GFX8-NEXT: v_or_b32_sdwa v2, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 -; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX8-NEXT: v_or_b32_e32 v5, 0x400000, v1 +; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v1, v1 +; GFX8-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX8-NEXT: v_cndmask_b32_e32 v1, v4, v5, vcc +; GFX8-NEXT: v_or_b32_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:WORD_1 +; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v1, v2 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB19_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6333,16 +6353,16 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX7-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX7-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX7-NEXT: v_or_b32_e32 v2, v3, v2 -; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 offset:65534 +; GFX7-NEXT: v_mov_b32_e32 v2, v1 +; GFX7-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX7-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX7-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; GFX7-NEXT: v_or_b32_e32 v1, v3, v1 +; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 offset:65534 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v1, v2 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB19_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6359,16 +6379,16 @@ define void @local_atomic_fsub_noret_bf16__offset__align4(ptr addrspace(3) %ptr) ; GFX6-NEXT: .LBB19_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_lshlrev_b32_e32 v2, 16, v1 -; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v2 -; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v1 -; GFX6-NEXT: v_lshrrev_b32_e32 v2, 16, v2 -; GFX6-NEXT: v_or_b32_e32 v2, v3, v2 -; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX6-NEXT: v_mov_b32_e32 v2, v1 +; GFX6-NEXT: v_lshlrev_b32_e32 v1, 16, v2 +; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v1 +; GFX6-NEXT: v_and_b32_e32 v3, 0xffff0000, v2 +; GFX6-NEXT: v_lshrrev_b32_e32 v1, 16, v1 +; GFX6-NEXT: v_or_b32_e32 v1, v3, v1 +; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v1, v2 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB19_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6418,17 +6438,17 @@ define <2 x half> @local_atomic_fsub_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> ; GFX942-LABEL: local_atomic_fsub_ret_v2f16: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v3, v0 +; GFX942-NEXT: ds_read_b32 v2, v0 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB20_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b32_e32 v3, v2 ; GFX942-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v2 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB20_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6487,17 +6507,17 @@ define <2 x half> @local_atomic_fsub_ret_v2f16(ptr addrspace(3) %ptr, <2 x half> ; GFX90A-LABEL: local_atomic_fsub_ret_v2f16: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v3, v0 +; GFX90A-NEXT: ds_read_b32 v2, v0 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB20_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_mov_b32_e32 v3, v2 ; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v2 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB20_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6673,17 +6693,17 @@ define <2 x half> @local_atomic_fsub_ret_v2f16__offset(ptr addrspace(3) %ptr, <2 ; GFX942-LABEL: local_atomic_fsub_ret_v2f16__offset: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v3, v0 offset:65532 +; GFX942-NEXT: ds_read_b32 v2, v0 offset:65532 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB21_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b32_e32 v3, v2 ; GFX942-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v3, v2 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB21_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6742,17 +6762,17 @@ define <2 x half> @local_atomic_fsub_ret_v2f16__offset(ptr addrspace(3) %ptr, <2 ; GFX90A-LABEL: local_atomic_fsub_ret_v2f16__offset: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v3, v0 offset:65532 +; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB21_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_mov_b32_e32 v3, v2 ; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v3, v2 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB21_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6906,13 +6926,14 @@ define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %va ; GFX12-NEXT: .LBB22_1: ; %atomicrmw.start ; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-NEXT: s_wait_dscnt 0x0 -; GFX12-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX12-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX12-NEXT: s_wait_storecnt 0x0 -; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 +; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 ; GFX12-NEXT: s_wait_dscnt 0x0 ; GFX12-NEXT: global_inv scope:SCOPE_SE -; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 -; GFX12-NEXT: v_mov_b32_e32 v2, v3 +; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 ; GFX12-NEXT: s_wait_alu 0xfffe ; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-NEXT: s_wait_alu 0xfffe @@ -6931,12 +6952,12 @@ define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %va ; GFX942-NEXT: .LBB22_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] -; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 +; GFX942-NEXT: v_mov_b32_e32 v3, v2 +; GFX942-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v2, v3 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB22_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -6951,13 +6972,14 @@ define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %va ; GFX11-NEXT: .LBB22_1: ; %atomicrmw.start ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX11-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 +; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-NEXT: buffer_gl0_inv -; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 -; GFX11-NEXT: v_mov_b32_e32 v2, v3 +; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 ; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -6974,13 +6996,13 @@ define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %va ; GFX10-NEXT: .LBB22_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX10-NEXT: v_mov_b32_e32 v3, v2 +; GFX10-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 +; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 -; GFX10-NEXT: v_mov_b32_e32 v2, v3 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB22_1 @@ -6996,12 +7018,12 @@ define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %va ; GFX90A-NEXT: .LBB22_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] -; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 +; GFX90A-NEXT: v_mov_b32_e32 v3, v2 +; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v2, v3 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB22_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -7016,12 +7038,12 @@ define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %va ; GFX908-NEXT: .LBB22_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] -; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 +; GFX908-NEXT: v_mov_b32_e32 v3, v2 +; GFX908-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v2, v3 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB22_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -7037,14 +7059,14 @@ define void @local_atomic_fsub_noret_v2f16(ptr addrspace(3) %ptr, <2 x half> %va ; GFX8-NEXT: .LBB22_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_sub_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 -; GFX8-NEXT: v_sub_f16_e32 v4, v2, v1 -; GFX8-NEXT: v_or_b32_e32 v3, v4, v3 -; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 +; GFX8-NEXT: v_mov_b32_e32 v3, v2 +; GFX8-NEXT: v_sub_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 +; GFX8-NEXT: v_sub_f16_e32 v4, v3, v1 +; GFX8-NEXT: v_or_b32_e32 v2, v4, v2 +; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v2, v3 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB22_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -7149,13 +7171,14 @@ define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x h ; GFX12-NEXT: .LBB23_1: ; %atomicrmw.start ; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-NEXT: s_wait_dscnt 0x0 -; GFX12-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX12-NEXT: v_mov_b32_e32 v3, v2 +; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX12-NEXT: s_wait_storecnt 0x0 -; GFX12-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532 +; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532 ; GFX12-NEXT: s_wait_dscnt 0x0 ; GFX12-NEXT: global_inv scope:SCOPE_SE -; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 -; GFX12-NEXT: v_mov_b32_e32 v2, v3 +; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 ; GFX12-NEXT: s_wait_alu 0xfffe ; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-NEXT: s_wait_alu 0xfffe @@ -7174,12 +7197,12 @@ define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x h ; GFX942-NEXT: .LBB23_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] -; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532 +; GFX942-NEXT: v_mov_b32_e32 v3, v2 +; GFX942-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v2, v3 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB23_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -7194,13 +7217,14 @@ define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x h ; GFX11-NEXT: .LBB23_1: ; %atomicrmw.start ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX11-NEXT: v_mov_b32_e32 v3, v2 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v2 offset:65532 +; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v3 offset:65532 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-NEXT: buffer_gl0_inv -; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 -; GFX11-NEXT: v_mov_b32_e32 v2, v3 +; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 ; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -7217,13 +7241,13 @@ define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x h ; GFX10-NEXT: .LBB23_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX10-NEXT: v_mov_b32_e32 v3, v2 +; GFX10-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532 +; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v2 -; GFX10-NEXT: v_mov_b32_e32 v2, v3 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v3 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB23_1 @@ -7239,12 +7263,12 @@ define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x h ; GFX90A-NEXT: .LBB23_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] -; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532 +; GFX90A-NEXT: v_mov_b32_e32 v3, v2 +; GFX90A-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v2, v3 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB23_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -7259,12 +7283,12 @@ define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x h ; GFX908-NEXT: .LBB23_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_pk_add_f16 v3, v2, v1 neg_lo:[0,1] neg_hi:[0,1] -; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532 +; GFX908-NEXT: v_mov_b32_e32 v3, v2 +; GFX908-NEXT: v_pk_add_f16 v2, v3, v1 neg_lo:[0,1] neg_hi:[0,1] +; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v2, v3 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB23_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -7280,14 +7304,14 @@ define void @local_atomic_fsub_noret_v2f16__offset(ptr addrspace(3) %ptr, <2 x h ; GFX8-NEXT: .LBB23_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_sub_f16_sdwa v3, v2, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 -; GFX8-NEXT: v_sub_f16_e32 v4, v2, v1 -; GFX8-NEXT: v_or_b32_e32 v3, v4, v3 -; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v2, v3 offset:65532 +; GFX8-NEXT: v_mov_b32_e32 v3, v2 +; GFX8-NEXT: v_sub_f16_sdwa v2, v3, v1 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 +; GFX8-NEXT: v_sub_f16_e32 v4, v3, v1 +; GFX8-NEXT: v_or_b32_e32 v2, v4, v2 +; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v3, v2 offset:65532 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v2 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v3 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v2, v3 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB23_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -7494,41 +7518,41 @@ define <2 x bfloat> @local_atomic_fsub_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bf ; GFX942-LABEL: local_atomic_fsub_ret_v2bf16: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v3, v0 +; GFX942-NEXT: ds_read_b32 v2, v0 ; GFX942-NEXT: s_mov_b64 s[2:3], 0 -; GFX942-NEXT: v_lshlrev_b32_e32 v2, 16, v1 +; GFX942-NEXT: v_lshlrev_b32_e32 v3, 16, v1 ; GFX942-NEXT: s_movk_i32 s4, 0x7fff -; GFX942-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX942-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 ; GFX942-NEXT: s_mov_b32 s5, 0x7060302 ; GFX942-NEXT: .LBB24_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshlrev_b32_e32 v1, 16, v3 -; GFX942-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX942-NEXT: v_sub_f32_e32 v1, v1, v2 -; GFX942-NEXT: v_sub_f32_e32 v5, v5, v4 -; GFX942-NEXT: v_bfe_u32 v6, v1, 16, 1 +; GFX942-NEXT: v_mov_b32_e32 v4, v2 +; GFX942-NEXT: v_lshlrev_b32_e32 v2, 16, v4 +; GFX942-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX942-NEXT: v_sub_f32_e32 v2, v2, v3 +; GFX942-NEXT: v_sub_f32_e32 v5, v5, v1 +; GFX942-NEXT: v_bfe_u32 v6, v2, 16, 1 ; GFX942-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v7, 0x400000, v1 +; GFX942-NEXT: v_or_b32_e32 v7, 0x400000, v2 ; GFX942-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX942-NEXT: v_add3_u32 v6, v6, v1, s4 +; GFX942-NEXT: v_add3_u32 v6, v6, v2, s4 ; GFX942-NEXT: v_add3_u32 v8, v8, v5, s4 ; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX942-NEXT: v_cmp_u_f32_e64 s[0:1], v1, v1 +; GFX942-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2 ; GFX942-NEXT: s_nop 0 ; GFX942-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX942-NEXT: v_cndmask_b32_e64 v1, v6, v7, s[0:1] -; GFX942-NEXT: v_perm_b32 v1, v5, v1, s5 -; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v3, v1 +; GFX942-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1] +; GFX942-NEXT: v_perm_b32 v2, v5, v2, s5 +; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4 ; GFX942-NEXT: s_or_b64 s[2:3], vcc, s[2:3] -; GFX942-NEXT: v_mov_b32_e32 v3, v1 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-NEXT: s_cbranch_execnz .LBB24_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX942-NEXT: s_or_b64 exec, exec, s[2:3] -; GFX942-NEXT: v_mov_b32_e32 v0, v1 +; GFX942-NEXT: v_mov_b32_e32 v0, v2 ; GFX942-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-TRUE16-LABEL: local_atomic_fsub_ret_v2bf16: @@ -7666,40 +7690,40 @@ define <2 x bfloat> @local_atomic_fsub_ret_v2bf16(ptr addrspace(3) %ptr, <2 x bf ; GFX90A-LABEL: local_atomic_fsub_ret_v2bf16: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v3, v0 +; GFX90A-NEXT: ds_read_b32 v2, v0 ; GFX90A-NEXT: s_mov_b64 s[6:7], 0 -; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1 +; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1 ; GFX90A-NEXT: s_movk_i32 s8, 0x7fff -; GFX90A-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 ; GFX90A-NEXT: s_mov_b32 s9, 0x7060302 ; GFX90A-NEXT: .LBB24_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v3 -; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX90A-NEXT: v_sub_f32_e32 v1, v1, v2 -; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v4 -; GFX90A-NEXT: v_bfe_u32 v6, v1, 16, 1 +; GFX90A-NEXT: v_mov_b32_e32 v4, v2 +; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4 +; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v3 +; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v1 +; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1 ; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v1 +; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2 ; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX90A-NEXT: v_add3_u32 v6, v6, v1, s8 +; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8 ; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8 ; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v1, v1 -; GFX90A-NEXT: v_cndmask_b32_e64 v1, v6, v7, s[4:5] +; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2 +; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5] ; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX90A-NEXT: v_perm_b32 v1, v5, v1, s9 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v3, v1 +; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4 ; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7] -; GFX90A-NEXT: v_mov_b32_e32 v3, v1 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX90A-NEXT: s_cbranch_execnz .LBB24_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7] -; GFX90A-NEXT: v_mov_b32_e32 v0, v1 +; GFX90A-NEXT: v_mov_b32_e32 v0, v2 ; GFX90A-NEXT: s_setpc_b64 s[30:31] ; ; GFX908-LABEL: local_atomic_fsub_ret_v2bf16: @@ -7972,41 +7996,41 @@ define <2 x bfloat> @local_atomic_fsub_ret_v2bf16__offset(ptr addrspace(3) %ptr, ; GFX942-LABEL: local_atomic_fsub_ret_v2bf16__offset: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v3, v0 offset:65532 +; GFX942-NEXT: ds_read_b32 v2, v0 offset:65532 ; GFX942-NEXT: s_mov_b64 s[2:3], 0 -; GFX942-NEXT: v_lshlrev_b32_e32 v2, 16, v1 +; GFX942-NEXT: v_lshlrev_b32_e32 v3, 16, v1 ; GFX942-NEXT: s_movk_i32 s4, 0x7fff -; GFX942-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX942-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 ; GFX942-NEXT: s_mov_b32 s5, 0x7060302 ; GFX942-NEXT: .LBB25_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshlrev_b32_e32 v1, 16, v3 -; GFX942-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX942-NEXT: v_sub_f32_e32 v1, v1, v2 -; GFX942-NEXT: v_sub_f32_e32 v5, v5, v4 -; GFX942-NEXT: v_bfe_u32 v6, v1, 16, 1 +; GFX942-NEXT: v_mov_b32_e32 v4, v2 +; GFX942-NEXT: v_lshlrev_b32_e32 v2, 16, v4 +; GFX942-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX942-NEXT: v_sub_f32_e32 v2, v2, v3 +; GFX942-NEXT: v_sub_f32_e32 v5, v5, v1 +; GFX942-NEXT: v_bfe_u32 v6, v2, 16, 1 ; GFX942-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v7, 0x400000, v1 +; GFX942-NEXT: v_or_b32_e32 v7, 0x400000, v2 ; GFX942-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX942-NEXT: v_add3_u32 v6, v6, v1, s4 +; GFX942-NEXT: v_add3_u32 v6, v6, v2, s4 ; GFX942-NEXT: v_add3_u32 v8, v8, v5, s4 ; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX942-NEXT: v_cmp_u_f32_e64 s[0:1], v1, v1 +; GFX942-NEXT: v_cmp_u_f32_e64 s[0:1], v2, v2 ; GFX942-NEXT: s_nop 0 ; GFX942-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX942-NEXT: v_cndmask_b32_e64 v1, v6, v7, s[0:1] -; GFX942-NEXT: v_perm_b32 v1, v5, v1, s5 -; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v3, v1 offset:65532 +; GFX942-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[0:1] +; GFX942-NEXT: v_perm_b32 v2, v5, v2, s5 +; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4 ; GFX942-NEXT: s_or_b64 s[2:3], vcc, s[2:3] -; GFX942-NEXT: v_mov_b32_e32 v3, v1 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-NEXT: s_cbranch_execnz .LBB25_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX942-NEXT: s_or_b64 exec, exec, s[2:3] -; GFX942-NEXT: v_mov_b32_e32 v0, v1 +; GFX942-NEXT: v_mov_b32_e32 v0, v2 ; GFX942-NEXT: s_setpc_b64 s[30:31] ; ; GFX11-TRUE16-LABEL: local_atomic_fsub_ret_v2bf16__offset: @@ -8144,40 +8168,40 @@ define <2 x bfloat> @local_atomic_fsub_ret_v2bf16__offset(ptr addrspace(3) %ptr, ; GFX90A-LABEL: local_atomic_fsub_ret_v2bf16__offset: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v3, v0 offset:65532 +; GFX90A-NEXT: ds_read_b32 v2, v0 offset:65532 ; GFX90A-NEXT: s_mov_b64 s[6:7], 0 -; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v1 +; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v1 ; GFX90A-NEXT: s_movk_i32 s8, 0x7fff -; GFX90A-NEXT: v_and_b32_e32 v4, 0xffff0000, v1 +; GFX90A-NEXT: v_and_b32_e32 v1, 0xffff0000, v1 ; GFX90A-NEXT: s_mov_b32 s9, 0x7060302 ; GFX90A-NEXT: .LBB25_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshlrev_b32_e32 v1, 16, v3 -; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX90A-NEXT: v_sub_f32_e32 v1, v1, v2 -; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v4 -; GFX90A-NEXT: v_bfe_u32 v6, v1, 16, 1 +; GFX90A-NEXT: v_mov_b32_e32 v4, v2 +; GFX90A-NEXT: v_lshlrev_b32_e32 v2, 16, v4 +; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX90A-NEXT: v_sub_f32_e32 v2, v2, v3 +; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v1 +; GFX90A-NEXT: v_bfe_u32 v6, v2, 16, 1 ; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v1 +; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v2 ; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX90A-NEXT: v_add3_u32 v6, v6, v1, s8 +; GFX90A-NEXT: v_add3_u32 v6, v6, v2, s8 ; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8 ; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v1, v1 -; GFX90A-NEXT: v_cndmask_b32_e64 v1, v6, v7, s[4:5] +; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v2, v2 +; GFX90A-NEXT: v_cndmask_b32_e64 v2, v6, v7, s[4:5] ; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX90A-NEXT: v_perm_b32 v1, v5, v1, s9 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v3, v1 offset:65532 +; GFX90A-NEXT: v_perm_b32 v2, v5, v2, s9 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v4, v2 offset:65532 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v4 ; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7] -; GFX90A-NEXT: v_mov_b32_e32 v3, v1 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX90A-NEXT: s_cbranch_execnz .LBB25_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end ; GFX90A-NEXT: s_or_b64 exec, exec, s[6:7] -; GFX90A-NEXT: v_mov_b32_e32 v0, v1 +; GFX90A-NEXT: v_mov_b32_e32 v0, v2 ; GFX90A-NEXT: s_setpc_b64 s[30:31] ; ; GFX908-LABEL: local_atomic_fsub_ret_v2bf16__offset: @@ -8357,31 +8381,34 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX12-TRUE16-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_dual_sub_f32 v5, v5, v2 :: v_dual_lshlrev_b32 v4, 16, v3 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v4, v4, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v1 ; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v8, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v5, v3 +; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v5, v4 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -8406,32 +8433,33 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX12-FAKE16-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX12-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 -; GFX12-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 +; GFX12-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3 -; GFX12-FAKE16-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX12-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX12-FAKE16-NEXT: v_sub_f32_e32 v5, v5, v1 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX12-FAKE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-FAKE16-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX12-FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 +; GFX12-FAKE16-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX12-FAKE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-FAKE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX12-FAKE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff -; GFX12-FAKE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 +; GFX12-FAKE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX12-FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-FAKE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffd -; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX12-FAKE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo +; GFX12-FAKE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff ; GFX12-FAKE16-NEXT: s_wait_alu 0xf1ff -; GFX12-FAKE16-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0 -; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_perm_b32 v4, v5, v4, 0x7060302 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_cndmask_b32_e64 v3, v6, v8, s0 +; GFX12-FAKE16-NEXT: v_perm_b32 v3, v5, v3, 0x7060302 ; GFX12-FAKE16-NEXT: s_wait_storecnt 0x0 -; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 ; GFX12-FAKE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX12-FAKE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX12-FAKE16-NEXT: s_or_b32 s1, vcc_lo, s1 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe @@ -8454,27 +8482,27 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX942-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX942-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX942-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX942-NEXT: v_mov_b32_e32 v4, v3 +; GFX942-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX942-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX942-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX942-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX942-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX942-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX942-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX942-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX942-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX942-NEXT: v_add3_u32 v6, v6, v4, s4 +; GFX942-NEXT: v_add3_u32 v6, v6, v3, s4 ; GFX942-NEXT: v_add3_u32 v8, v8, v5, s4 ; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX942-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4 +; GFX942-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3 ; GFX942-NEXT: s_nop 0 ; GFX942-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX942-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1] -; GFX942-NEXT: v_perm_b32 v4, v5, v4, s5 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX942-NEXT: v_cndmask_b32_e64 v3, v6, v7, s[0:1] +; GFX942-NEXT: v_perm_b32 v3, v5, v3, s5 +; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX942-NEXT: s_or_b64 s[2:3], vcc, s[2:3] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-NEXT: s_cbranch_execnz .LBB26_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -8492,30 +8520,32 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX11-TRUE16-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_sub_f32 v5, v5, v2 :: v_dual_lshlrev_b32 v4, 16, v3 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v4, v4, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 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 v4, v6, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v5, v3 +; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v5, v4 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -8536,30 +8566,32 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX11-FAKE16-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3 -; GFX11-FAKE16-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX11-FAKE16-NEXT: v_sub_f32_e32 v5, v5, v1 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-FAKE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-FAKE16-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 +; GFX11-FAKE16-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-FAKE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-FAKE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff -; GFX11-FAKE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 -; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-FAKE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-FAKE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo -; GFX11-FAKE16-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0 -; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_perm_b32 v4, v5, v4, 0x7060302 +; GFX11-FAKE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_cndmask_b32_e64 v3, v6, v8, s0 +; GFX11-FAKE16-NEXT: v_perm_b32 v3, v5, v3, 0x7060302 ; GFX11-FAKE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 +; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-FAKE16-NEXT: buffer_gl0_inv -; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-FAKE16-NEXT: s_or_b32 s1, vcc_lo, s1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1 @@ -8579,27 +8611,27 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX10-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX10-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX10-NEXT: v_mov_b32_e32 v4, v3 +; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX10-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX10-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4 +; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3 ; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff ; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4 +; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3 ; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo -; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4 -; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302 +; GFX10-NEXT: v_cndmask_b32_e64 v3, v6, v8, s4 +; GFX10-NEXT: v_perm_b32 v3, v5, v3, 0x7060302 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX10-NEXT: v_mov_b32_e32 v3, v4 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5 ; GFX10-NEXT: s_cbranch_execnz .LBB26_1 @@ -8619,26 +8651,26 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX90A-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX90A-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX90A-NEXT: v_mov_b32_e32 v4, v3 +; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8 +; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8 ; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8 ; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4 -; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5] +; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3 +; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v7, s[4:5] ; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX90A-NEXT: v_perm_b32 v3, v5, v3, s9 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX90A-NEXT: s_cbranch_execnz .LBB26_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -8657,26 +8689,26 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX908-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX908-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX908-NEXT: v_mov_b32_e32 v4, v3 +; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX908-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX908-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX908-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8 +; GFX908-NEXT: v_add3_u32 v6, v6, v3, s8 ; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8 ; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4 -; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5] +; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3 +; GFX908-NEXT: v_cndmask_b32_e64 v3, v6, v7, s[4:5] ; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9 -; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX908-NEXT: v_perm_b32 v3, v5, v3, s9 +; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7] -; GFX908-NEXT: v_mov_b32_e32 v3, v4 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX908-NEXT: s_cbranch_execnz .LBB26_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -8694,29 +8726,29 @@ define void @local_atomic_fsub_noret_v2bf16(ptr addrspace(3) %ptr, <2 x bfloat> ; GFX8-NEXT: .LBB26_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX8-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX8-NEXT: v_mov_b32_e32 v4, v3 +; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX8-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX8-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4 +; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3 ; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5 ; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6 ; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8 ; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4 -; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4 +; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3 +; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3 ; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5] +; GFX8-NEXT: v_cndmask_b32_e64 v3, v6, v7, s[4:5] ; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5 -; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16 -; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 +; GFX8-NEXT: v_alignbit_b32 v3, v5, v3, 16 +; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7] -; GFX8-NEXT: v_mov_b32_e32 v3, v4 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX8-NEXT: s_cbranch_execnz .LBB26_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -8815,31 +8847,34 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX12-TRUE16-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX12-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 -; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 +; GFX12-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-TRUE16-NEXT: v_dual_sub_f32 v5, v5, v2 :: v_dual_lshlrev_b32 v4, 16, v3 -; GFX12-TRUE16-NEXT: v_sub_f32_e32 v4, v4, v1 +; GFX12-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 +; GFX12-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 ; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX12-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX12-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v1 ; GFX12-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX12-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX12-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX12-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1) +; GFX12-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd -; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_3) -; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v4, v6, v8, vcc_lo +; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo ; GFX12-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffd ; GFX12-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo -; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) +; GFX12-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h ; GFX12-TRUE16-NEXT: s_wait_storecnt 0x0 -; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v5, v3 offset:65532 +; GFX12-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v5, v4 offset:65532 ; GFX12-TRUE16-NEXT: s_wait_dscnt 0x0 ; GFX12-TRUE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX12-TRUE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX12-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe ; GFX12-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-TRUE16-NEXT: s_wait_alu 0xfffe @@ -8864,32 +8899,33 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX12-FAKE16-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX12-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 -; GFX12-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 +; GFX12-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3 -; GFX12-FAKE16-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX12-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX12-FAKE16-NEXT: v_sub_f32_e32 v5, v5, v1 +; GFX12-FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 ; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX12-FAKE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX12-FAKE16-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX12-FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 +; GFX12-FAKE16-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX12-FAKE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX12-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX12-FAKE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX12-FAKE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff -; GFX12-FAKE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 +; GFX12-FAKE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX12-FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX12-FAKE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffd -; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2) ; GFX12-FAKE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo +; GFX12-FAKE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff ; GFX12-FAKE16-NEXT: s_wait_alu 0xf1ff -; GFX12-FAKE16-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0 -; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX12-FAKE16-NEXT: v_perm_b32 v4, v5, v4, 0x7060302 +; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX12-FAKE16-NEXT: v_cndmask_b32_e64 v3, v6, v8, s0 +; GFX12-FAKE16-NEXT: v_perm_b32 v3, v5, v3, 0x7060302 ; GFX12-FAKE16-NEXT: s_wait_storecnt 0x0 -; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532 +; GFX12-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 offset:65532 ; GFX12-FAKE16-NEXT: s_wait_dscnt 0x0 ; GFX12-FAKE16-NEXT: global_inv scope:SCOPE_SE -; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX12-FAKE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX12-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe ; GFX12-FAKE16-NEXT: s_or_b32 s1, vcc_lo, s1 ; GFX12-FAKE16-NEXT: s_wait_alu 0xfffe @@ -8912,27 +8948,27 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX942-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX942-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX942-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX942-NEXT: v_mov_b32_e32 v4, v3 +; GFX942-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX942-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX942-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX942-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX942-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX942-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX942-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX942-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX942-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX942-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX942-NEXT: v_add3_u32 v6, v6, v4, s4 +; GFX942-NEXT: v_add3_u32 v6, v6, v3, s4 ; GFX942-NEXT: v_add3_u32 v8, v8, v5, s4 ; GFX942-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX942-NEXT: v_cmp_u_f32_e64 s[0:1], v4, v4 +; GFX942-NEXT: v_cmp_u_f32_e64 s[0:1], v3, v3 ; GFX942-NEXT: s_nop 0 ; GFX942-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX942-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[0:1] -; GFX942-NEXT: v_perm_b32 v4, v5, v4, s5 -; GFX942-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532 +; GFX942-NEXT: v_cndmask_b32_e64 v3, v6, v7, s[0:1] +; GFX942-NEXT: v_perm_b32 v3, v5, v3, s5 +; GFX942-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 offset:65532 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX942-NEXT: s_or_b64 s[2:3], vcc, s[2:3] -; GFX942-NEXT: v_mov_b32_e32 v3, v4 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[2:3] ; GFX942-NEXT: s_cbranch_execnz .LBB27_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -8950,30 +8986,32 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX11-TRUE16-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX11-TRUE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 +; GFX11-TRUE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-TRUE16-NEXT: v_dual_sub_f32 v5, v5, v2 :: v_dual_lshlrev_b32 v4, 16, v3 -; GFX11-TRUE16-NEXT: v_sub_f32_e32 v4, v4, v1 +; GFX11-TRUE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v5, v5, v2 +; GFX11-TRUE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-TRUE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 -; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v4, v4 +; GFX11-TRUE16-NEXT: v_sub_f32_e32 v3, v3, v1 ; GFX11-TRUE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3) ; GFX11-TRUE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v4, 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 v4, v6, v8, vcc_lo +; GFX11-TRUE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX11-TRUE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v3, v3 +; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-TRUE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff +; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v3, v6, v8, vcc_lo ; GFX11-TRUE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 ; GFX11-TRUE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo ; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_3) -; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v4.h +; GFX11-TRUE16-NEXT: v_mov_b16_e32 v5.l, v3.h ; GFX11-TRUE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v5, v3 offset:65532 +; GFX11-TRUE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v5, v4 offset:65532 ; GFX11-TRUE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-TRUE16-NEXT: buffer_gl0_inv -; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX11-TRUE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX11-TRUE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-TRUE16-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-TRUE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -8994,30 +9032,32 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX11-FAKE16-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX11-FAKE16-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 +; GFX11-FAKE16-NEXT: v_mov_b32_e32 v4, v3 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_dual_sub_f32 v5, v5, v1 :: v_dual_lshlrev_b32 v4, 16, v3 -; GFX11-FAKE16-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX11-FAKE16-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX11-FAKE16-NEXT: v_sub_f32_e32 v5, v5, v1 +; GFX11-FAKE16-NEXT: v_lshlrev_b32_e32 v3, 16, v4 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2) ; GFX11-FAKE16-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX11-FAKE16-NEXT: v_bfe_u32 v6, v4, 16, 1 -; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v4 +; GFX11-FAKE16-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX11-FAKE16-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX11-FAKE16-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-FAKE16-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX11-FAKE16-NEXT: v_add3_u32 v6, v6, v4, 0x7fff -; GFX11-FAKE16-NEXT: v_cmp_u_f32_e64 s0, v4, v4 -; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2) +; GFX11-FAKE16-NEXT: v_bfe_u32 v6, v3, 16, 1 +; GFX11-FAKE16-NEXT: v_or_b32_e32 v8, 0x400000, v3 +; GFX11-FAKE16-NEXT: v_cmp_u_f32_e64 s0, v3, v3 +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4) ; GFX11-FAKE16-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo -; GFX11-FAKE16-NEXT: v_cndmask_b32_e64 v4, v6, v8, s0 -; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) -; GFX11-FAKE16-NEXT: v_perm_b32 v4, v5, v4, 0x7060302 +; GFX11-FAKE16-NEXT: v_add3_u32 v6, v6, v3, 0x7fff +; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1) +; GFX11-FAKE16-NEXT: v_cndmask_b32_e64 v3, v6, v8, s0 +; GFX11-FAKE16-NEXT: v_perm_b32 v3, v5, v3, 0x7060302 ; GFX11-FAKE16-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v4, v0, v4, v3 offset:65532 +; GFX11-FAKE16-NEXT: ds_cmpstore_rtn_b32 v3, v0, v3, v4 offset:65532 ; GFX11-FAKE16-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-FAKE16-NEXT: buffer_gl0_inv -; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX11-FAKE16-NEXT: v_mov_b32_e32 v3, v4 +; GFX11-FAKE16-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX11-FAKE16-NEXT: s_or_b32 s1, vcc_lo, s1 ; GFX11-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-FAKE16-NEXT: s_and_not1_b32 exec_lo, exec_lo, s1 @@ -9037,27 +9077,27 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX10-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX10-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX10-NEXT: v_mov_b32_e32 v4, v3 +; GFX10-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX10-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX10-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX10-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX10-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX10-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX10-NEXT: v_bfe_u32 v7, v5, 16, 1 -; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v4 +; GFX10-NEXT: v_or_b32_e32 v8, 0x400000, v3 ; GFX10-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX10-NEXT: v_cmp_u_f32_e32 vcc_lo, v5, v5 -; GFX10-NEXT: v_add3_u32 v6, v6, v4, 0x7fff +; GFX10-NEXT: v_add3_u32 v6, v6, v3, 0x7fff ; GFX10-NEXT: v_add3_u32 v7, v7, v5, 0x7fff -; GFX10-NEXT: v_cmp_u_f32_e64 s4, v4, v4 +; GFX10-NEXT: v_cmp_u_f32_e64 s4, v3, v3 ; GFX10-NEXT: v_cndmask_b32_e32 v5, v7, v9, vcc_lo -; GFX10-NEXT: v_cndmask_b32_e64 v4, v6, v8, s4 -; GFX10-NEXT: v_perm_b32 v4, v5, v4, 0x7060302 +; GFX10-NEXT: v_cndmask_b32_e64 v3, v6, v8, s4 +; GFX10-NEXT: v_perm_b32 v3, v5, v3, 0x7060302 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532 +; GFX10-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 offset:65532 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v4, v3 -; GFX10-NEXT: v_mov_b32_e32 v3, v4 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v3, v4 ; GFX10-NEXT: s_or_b32 s5, vcc_lo, s5 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s5 ; GFX10-NEXT: s_cbranch_execnz .LBB27_1 @@ -9077,26 +9117,26 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX90A-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX90A-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX90A-NEXT: v_mov_b32_e32 v4, v3 +; GFX90A-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX90A-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX90A-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX90A-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX90A-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX90A-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX90A-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX90A-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX90A-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX90A-NEXT: v_add3_u32 v6, v6, v4, s8 +; GFX90A-NEXT: v_add3_u32 v6, v6, v3, s8 ; GFX90A-NEXT: v_add3_u32 v8, v8, v5, s8 ; GFX90A-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4 -; GFX90A-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5] +; GFX90A-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3 +; GFX90A-NEXT: v_cndmask_b32_e64 v3, v6, v7, s[4:5] ; GFX90A-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX90A-NEXT: v_perm_b32 v4, v5, v4, s9 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532 +; GFX90A-NEXT: v_perm_b32 v3, v5, v3, s9 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 offset:65532 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX90A-NEXT: s_or_b64 s[6:7], vcc, s[6:7] -; GFX90A-NEXT: v_mov_b32_e32 v3, v4 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX90A-NEXT: s_cbranch_execnz .LBB27_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9115,26 +9155,26 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX908-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX908-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX908-NEXT: v_mov_b32_e32 v4, v3 +; GFX908-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX908-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX908-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX908-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX908-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX908-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX908-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v4 +; GFX908-NEXT: v_or_b32_e32 v7, 0x400000, v3 ; GFX908-NEXT: v_or_b32_e32 v9, 0x400000, v5 -; GFX908-NEXT: v_add3_u32 v6, v6, v4, s8 +; GFX908-NEXT: v_add3_u32 v6, v6, v3, s8 ; GFX908-NEXT: v_add3_u32 v8, v8, v5, s8 ; GFX908-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4 -; GFX908-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5] +; GFX908-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3 +; GFX908-NEXT: v_cndmask_b32_e64 v3, v6, v7, s[4:5] ; GFX908-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX908-NEXT: v_perm_b32 v4, v5, v4, s9 -; GFX908-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532 +; GFX908-NEXT: v_perm_b32 v3, v5, v3, s9 +; GFX908-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 offset:65532 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX908-NEXT: s_or_b64 s[6:7], vcc, s[6:7] -; GFX908-NEXT: v_mov_b32_e32 v3, v4 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX908-NEXT: s_cbranch_execnz .LBB27_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9152,29 +9192,29 @@ define void @local_atomic_fsub_noret_v2bf16__ofset(ptr addrspace(3) %ptr, <2 x b ; GFX8-NEXT: .LBB27_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_lshlrev_b32_e32 v4, 16, v3 -; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v3 -; GFX8-NEXT: v_sub_f32_e32 v4, v4, v2 +; GFX8-NEXT: v_mov_b32_e32 v4, v3 +; GFX8-NEXT: v_lshlrev_b32_e32 v3, 16, v4 +; GFX8-NEXT: v_and_b32_e32 v5, 0xffff0000, v4 +; GFX8-NEXT: v_sub_f32_e32 v3, v3, v2 ; GFX8-NEXT: v_sub_f32_e32 v5, v5, v1 -; GFX8-NEXT: v_bfe_u32 v6, v4, 16, 1 +; GFX8-NEXT: v_bfe_u32 v6, v3, 16, 1 ; GFX8-NEXT: v_bfe_u32 v8, v5, 16, 1 -; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v4 +; GFX8-NEXT: v_add_u32_e32 v6, vcc, v6, v3 ; GFX8-NEXT: v_add_u32_e32 v8, vcc, v8, v5 ; GFX8-NEXT: v_add_u32_e32 v6, vcc, 0x7fff, v6 ; GFX8-NEXT: v_add_u32_e32 v8, vcc, 0x7fff, v8 ; GFX8-NEXT: v_or_b32_e32 v9, 0x400000, v5 ; GFX8-NEXT: v_cmp_u_f32_e32 vcc, v5, v5 -; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v4 -; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v4, v4 +; GFX8-NEXT: v_or_b32_e32 v7, 0x400000, v3 +; GFX8-NEXT: v_cmp_u_f32_e64 s[4:5], v3, v3 ; GFX8-NEXT: v_cndmask_b32_e32 v5, v8, v9, vcc -; GFX8-NEXT: v_cndmask_b32_e64 v4, v6, v7, s[4:5] +; GFX8-NEXT: v_cndmask_b32_e64 v3, v6, v7, s[4:5] ; GFX8-NEXT: v_lshrrev_b32_e32 v5, 16, v5 -; GFX8-NEXT: v_alignbit_b32 v4, v5, v4, 16 -; GFX8-NEXT: ds_cmpst_rtn_b32 v4, v0, v3, v4 offset:65532 +; GFX8-NEXT: v_alignbit_b32 v3, v5, v3, 16 +; GFX8-NEXT: ds_cmpst_rtn_b32 v3, v0, v4, v3 offset:65532 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v4, v3 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v3, v4 ; GFX8-NEXT: s_or_b64 s[6:7], vcc, s[6:7] -; GFX8-NEXT: v_mov_b32_e32 v3, v4 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[6:7] ; GFX8-NEXT: s_cbranch_execnz .LBB27_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9299,17 +9339,17 @@ define float @local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode(ptr addrspa ; GFX942-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode: ; GFX942: ; %bb.0: ; GFX942-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX942-NEXT: ds_read_b32 v2, v0 +; GFX942-NEXT: ds_read_b32 v1, v0 ; GFX942-NEXT: s_mov_b64 s[0:1], 0 ; GFX942-NEXT: .LBB28_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) +; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) ; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v2, v1 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB28_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9368,17 +9408,17 @@ define float @local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode(ptr addrspa ; GFX90A-LABEL: local_atomic_fsub_ret_f32__amdgpu_ignore_denormal_mode: ; GFX90A: ; %bb.0: ; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) -; GFX90A-NEXT: ds_read_b32 v2, v0 +; GFX90A-NEXT: ds_read_b32 v1, v0 ; GFX90A-NEXT: s_mov_b64 s[4:5], 0 ; GFX90A-NEXT: .LBB28_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) ; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v2, v1 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB28_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9489,13 +9529,14 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX12-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX12-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX12-NEXT: s_wait_dscnt 0x0 -; GFX12-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX12-NEXT: v_mov_b32_e32 v2, v1 +; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX12-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX12-NEXT: s_wait_storecnt 0x0 -; GFX12-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 +; GFX12-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 ; GFX12-NEXT: s_wait_dscnt 0x0 ; GFX12-NEXT: global_inv scope:SCOPE_SE -; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX12-NEXT: v_mov_b32_e32 v1, v2 +; GFX12-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX12-NEXT: s_wait_alu 0xfffe ; GFX12-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX12-NEXT: s_wait_alu 0xfffe @@ -9514,12 +9555,12 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX942-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX942-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX942-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX942-NEXT: v_mov_b32_e32 v2, v1 +; GFX942-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX942-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX942-NEXT: s_waitcnt lgkmcnt(0) -; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX942-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX942-NEXT: s_or_b64 s[0:1], vcc, s[0:1] -; GFX942-NEXT: v_mov_b32_e32 v1, v2 ; GFX942-NEXT: s_andn2_b64 exec, exec, s[0:1] ; GFX942-NEXT: s_cbranch_execnz .LBB29_1 ; GFX942-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9534,13 +9575,14 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX11-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX11-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) -; GFX11-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX11-NEXT: v_mov_b32_e32 v2, v1 +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) +; GFX11-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX11-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX11-NEXT: ds_cmpstore_rtn_b32 v2, v0, v2, v1 +; GFX11-NEXT: ds_cmpstore_rtn_b32 v1, v0, v1, v2 ; GFX11-NEXT: s_waitcnt lgkmcnt(0) ; GFX11-NEXT: buffer_gl0_inv -; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX11-NEXT: v_mov_b32_e32 v1, v2 +; GFX11-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX11-NEXT: s_or_b32 s0, vcc_lo, s0 ; GFX11-NEXT: s_delay_alu instid0(SALU_CYCLE_1) ; GFX11-NEXT: s_and_not1_b32 exec_lo, exec_lo, s0 @@ -9557,13 +9599,13 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX10-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX10-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) -; GFX10-NEXT: v_add_f32_e32 v2, -4.0, v1 +; GFX10-NEXT: v_mov_b32_e32 v2, v1 +; GFX10-NEXT: v_add_f32_e32 v1, -4.0, v2 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0 -; GFX10-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX10-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX10-NEXT: s_waitcnt lgkmcnt(0) ; GFX10-NEXT: buffer_gl0_inv -; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v2, v1 -; GFX10-NEXT: v_mov_b32_e32 v1, v2 +; GFX10-NEXT: v_cmp_eq_u32_e32 vcc_lo, v1, v2 ; GFX10-NEXT: s_or_b32 s4, vcc_lo, s4 ; GFX10-NEXT: s_andn2_b32 exec_lo, exec_lo, s4 ; GFX10-NEXT: s_cbranch_execnz .LBB29_1 @@ -9579,12 +9621,12 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX90A-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX90A-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX90A-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX90A-NEXT: v_mov_b32_e32 v2, v1 +; GFX90A-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX90A-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX90A-NEXT: s_waitcnt lgkmcnt(0) -; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX90A-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX90A-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX90A-NEXT: v_mov_b32_e32 v1, v2 ; GFX90A-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX90A-NEXT: s_cbranch_execnz .LBB29_1 ; GFX90A-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9599,12 +9641,12 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX908-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX908-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX908-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX908-NEXT: v_mov_b32_e32 v2, v1 +; GFX908-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX908-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX908-NEXT: s_waitcnt lgkmcnt(0) -; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX908-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX908-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX908-NEXT: v_mov_b32_e32 v1, v2 ; GFX908-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX908-NEXT: s_cbranch_execnz .LBB29_1 ; GFX908-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9620,12 +9662,12 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX8-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX8-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX8-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX8-NEXT: v_mov_b32_e32 v2, v1 +; GFX8-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX8-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX8-NEXT: s_waitcnt lgkmcnt(0) -; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX8-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX8-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX8-NEXT: v_mov_b32_e32 v1, v2 ; GFX8-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX8-NEXT: s_cbranch_execnz .LBB29_1 ; GFX8-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9641,12 +9683,12 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX7-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX7-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX7-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX7-NEXT: v_mov_b32_e32 v2, v1 +; GFX7-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX7-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX7-NEXT: s_waitcnt lgkmcnt(0) -; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX7-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX7-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX7-NEXT: v_mov_b32_e32 v1, v2 ; GFX7-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX7-NEXT: s_cbranch_execnz .LBB29_1 ; GFX7-NEXT: ; %bb.2: ; %atomicrmw.end @@ -9662,12 +9704,12 @@ define void @local_atomic_fsub_noret_f32__amdgpu_ignore_denormal_mode(ptr addrsp ; GFX6-NEXT: .LBB29_1: ; %atomicrmw.start ; GFX6-NEXT: ; =>This Inner Loop Header: Depth=1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_add_f32_e32 v2, -4.0, v1 -; GFX6-NEXT: ds_cmpst_rtn_b32 v2, v0, v1, v2 +; GFX6-NEXT: v_mov_b32_e32 v2, v1 +; GFX6-NEXT: v_add_f32_e32 v1, -4.0, v2 +; GFX6-NEXT: ds_cmpst_rtn_b32 v1, v0, v2, v1 ; GFX6-NEXT: s_waitcnt lgkmcnt(0) -; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v2, v1 +; GFX6-NEXT: v_cmp_eq_u32_e32 vcc, v1, v2 ; GFX6-NEXT: s_or_b64 s[4:5], vcc, s[4:5] -; GFX6-NEXT: v_mov_b32_e32 v1, v2 ; GFX6-NEXT: s_andn2_b64 exec, exec, s[4:5] ; GFX6-NEXT: s_cbranch_execnz .LBB29_1 ; GFX6-NEXT: ; %bb.2: ; %atomicrmw.end |