[llvm] [AMDGPU] Prevent VOPD combining of VGPRs with different MSBs (PR #157168)
Stanislav Mekhanoshin via llvm-commits
llvm-commits at lists.llvm.org
Fri Sep 5 12:54:43 PDT 2025
https://github.com/rampitec created https://github.com/llvm/llvm-project/pull/157168
None
>From 6c59a8b2adb7973060556f830eea0f00d2504a5e Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin at amd.com>
Date: Fri, 5 Sep 2025 12:53:13 -0700
Subject: [PATCH] [AMDGPU] Prevent VOPD combining of VGPRs with different MSBs
---
.../Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp | 4 +
.../CodeGen/AMDGPU/vopd-combine-gfx1250.mir | 273 ++++++++++++++++++
2 files changed, 277 insertions(+)
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index 6348d3607878e..9f4f42185d9a0 100644
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -934,6 +934,10 @@ std::optional<unsigned> InstInfo::getInvalidCompOperandIndex(
if (!OpXRegs[CompOprIdx] || !OpYRegs[CompOprIdx])
continue;
+ if (getVGPREncodingMSBs(OpXRegs[CompOprIdx], MRI) !=
+ getVGPREncodingMSBs(OpYRegs[CompOprIdx], MRI))
+ return CompOprIdx;
+
if (SkipSrc && CompOprIdx >= Component::DST_NUM)
continue;
diff --git a/llvm/test/CodeGen/AMDGPU/vopd-combine-gfx1250.mir b/llvm/test/CodeGen/AMDGPU/vopd-combine-gfx1250.mir
index 586ddf627bd9e..fa6c34cf07730 100644
--- a/llvm/test/CodeGen/AMDGPU/vopd-combine-gfx1250.mir
+++ b/llvm/test/CodeGen/AMDGPU/vopd-combine-gfx1250.mir
@@ -27,6 +27,251 @@ body: |
$vgpr6 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
...
+---
+name: vopd_combine_hi_vgprs
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: vopd_combine_hi_vgprs
+ ; SCHED: $vgpr300 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr301 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr303 = V_SUB_F32_e32 $vgpr301, $vgpr301, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr306 = V_MUL_F32_e32 $vgpr300, $vgpr300, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr300, $vgpr301, implicit $exec
+ ;
+ ; PAIR-LABEL: name: vopd_combine_hi_vgprs
+ ; PAIR: $vgpr300 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr301 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr303, $vgpr306 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32_gfx1250 $vgpr301, $vgpr301, $vgpr300, $vgpr300, implicit $mode, implicit $exec, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr300, $vgpr301, implicit $exec
+ $vgpr300 = IMPLICIT_DEF
+ $vgpr301 = IMPLICIT_DEF
+ $vgpr303 = V_SUB_F32_e32 $vgpr301, $vgpr301, implicit $mode, implicit $exec
+ $vgpr4 = V_BFM_B32_e32 $vgpr300, $vgpr301, implicit $exec
+ $vgpr306 = V_MUL_F32_e32 killed $vgpr300, $vgpr300, implicit $mode, implicit $exec
+...
+
+---
+name: vopd_combine_hi_vgprs_above_512
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: vopd_combine_hi_vgprs_above_512
+ ; SCHED: $vgpr812 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr813 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr559 = V_SUB_F32_e32 $vgpr813, $vgpr813, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr562 = V_MUL_F32_e32 $vgpr812, $vgpr812, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr812, $vgpr813, implicit $exec
+ ;
+ ; PAIR-LABEL: name: vopd_combine_hi_vgprs_above_512
+ ; PAIR: $vgpr812 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr813 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr559, $vgpr562 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32_gfx1250 $vgpr813, $vgpr813, $vgpr812, $vgpr812, implicit $mode, implicit $exec, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr812, $vgpr813, implicit $exec
+ $vgpr812 = IMPLICIT_DEF
+ $vgpr813 = IMPLICIT_DEF
+ $vgpr559 = V_SUB_F32_e32 $vgpr813, $vgpr813, implicit $mode, implicit $exec
+ $vgpr4 = V_BFM_B32_e32 $vgpr812, $vgpr813, implicit $exec
+ $vgpr562 = V_MUL_F32_e32 killed $vgpr812, $vgpr812, implicit $mode, implicit $exec
+...
+
+---
+name: mixed_vgprs_low_and_hi_dst
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: mixed_vgprs_low_and_hi_dst
+ ; SCHED: $vgpr0 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr303 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr0, killed $vgpr1, implicit $exec
+ ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+ ;
+ ; PAIR-LABEL: name: mixed_vgprs_low_and_hi_dst
+ ; PAIR: $vgpr0 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr303 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr0, killed $vgpr1, implicit $exec
+ ; PAIR-NEXT: $vgpr6 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+ $vgpr0 = IMPLICIT_DEF
+ $vgpr1 = IMPLICIT_DEF
+ $vgpr303 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ $vgpr4 = V_BFM_B32_e32 $vgpr0, $vgpr1, implicit $exec
+ $vgpr6 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+...
+
+---
+name: mixed_vgprs_low_and_hi_scr0
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: mixed_vgprs_low_and_hi_scr0
+ ; SCHED: $vgpr0 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr300 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr0, killed $vgpr1, implicit $exec
+ ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr300, killed $vgpr0, implicit $mode, implicit $exec
+ ;
+ ; PAIR-LABEL: name: mixed_vgprs_low_and_hi_scr0
+ ; PAIR: $vgpr0 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr300 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr0, killed $vgpr1, implicit $exec
+ ; PAIR-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr300, killed $vgpr0, implicit $mode, implicit $exec
+ $vgpr0 = IMPLICIT_DEF
+ $vgpr1 = IMPLICIT_DEF
+ $vgpr300 = IMPLICIT_DEF
+ $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ $vgpr4 = V_BFM_B32_e32 $vgpr0, $vgpr1, implicit $exec
+ $vgpr6 = V_MUL_F32_e32 killed $vgpr300, $vgpr0, implicit $mode, implicit $exec
+...
+
+---
+name: mixed_vgprs_low_and_hi_scr1
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: mixed_vgprs_low_and_hi_scr1
+ ; SCHED: $vgpr0 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr300 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr301, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr0, killed $vgpr1, implicit $exec
+ ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr300, killed $vgpr0, implicit $mode, implicit $exec
+ ;
+ ; PAIR-LABEL: name: mixed_vgprs_low_and_hi_scr1
+ ; PAIR: $vgpr0 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr300 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr301, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr4 = V_BFM_B32_e32 $vgpr0, killed $vgpr1, implicit $exec
+ ; PAIR-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr300, killed $vgpr0, implicit $mode, implicit $exec
+ $vgpr0 = IMPLICIT_DEF
+ $vgpr1 = IMPLICIT_DEF
+ $vgpr300 = IMPLICIT_DEF
+ $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr301, implicit $mode, implicit $exec
+ $vgpr4 = V_BFM_B32_e32 $vgpr0, $vgpr1, implicit $exec
+ $vgpr6 = V_MUL_F32_e32 killed $vgpr300, $vgpr0, implicit $mode, implicit $exec
+...
+
+---
+name: mixed_vgprs_hi_and_hi_dst_different_msb
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: mixed_vgprs_hi_and_hi_dst_different_msb
+ ; SCHED: $vgpr0 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr559 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr303 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+ ;
+ ; PAIR-LABEL: name: mixed_vgprs_hi_and_hi_dst_different_msb
+ ; PAIR: $vgpr0 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr559 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr303 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+ $vgpr0 = IMPLICIT_DEF
+ $vgpr1 = IMPLICIT_DEF
+ $vgpr559 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+ $vgpr303 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+...
+
+---
+name: mixed_vgprs_low_and_hi_scr0_different_msb
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: mixed_vgprs_low_and_hi_scr0_different_msb
+ ; SCHED: $vgpr0 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr513 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr812 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr513, killed $vgpr1, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr812, killed $vgpr0, implicit $mode, implicit $exec
+ ;
+ ; PAIR-LABEL: name: mixed_vgprs_low_and_hi_scr0_different_msb
+ ; PAIR: $vgpr0 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr513 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr812 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr513, killed $vgpr1, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr812, killed $vgpr0, implicit $mode, implicit $exec
+ $vgpr0 = IMPLICIT_DEF
+ $vgpr1 = IMPLICIT_DEF
+ $vgpr513 = IMPLICIT_DEF
+ $vgpr812 = IMPLICIT_DEF
+ $vgpr3 = V_SUB_F32_e32 $vgpr513, $vgpr1, implicit $mode, implicit $exec
+ $vgpr6 = V_MUL_F32_e32 killed $vgpr812, $vgpr0, implicit $mode, implicit $exec
+...
+
+---
+name: vopd_combine_sgpr_src0
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: vopd_combine_sgpr_src0
+ ; SCHED: $vgpr0 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr300 = IMPLICIT_DEF
+ ; SCHED-NEXT: $sgpr0 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 killed $sgpr0, $vgpr1, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr300, $vgpr0, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr4 = V_BFM_B32_e32 killed $vgpr0, killed $vgpr1, implicit $exec
+ ;
+ ; PAIR-LABEL: name: vopd_combine_sgpr_src0
+ ; PAIR: $vgpr0 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr300 = IMPLICIT_DEF
+ ; PAIR-NEXT: $sgpr0 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr3, $vgpr6 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32_gfx1250 killed $sgpr0, $vgpr1, $vgpr300, $vgpr0, implicit $mode, implicit $exec, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr4 = V_BFM_B32_e32 killed $vgpr0, killed $vgpr1, implicit $exec
+ $vgpr0 = IMPLICIT_DEF
+ $vgpr1 = IMPLICIT_DEF
+ $vgpr300 = IMPLICIT_DEF
+ $sgpr0 = IMPLICIT_DEF
+ $vgpr3 = V_SUB_F32_e32 $sgpr0, $vgpr1, implicit $mode, implicit $exec
+ $vgpr4 = V_BFM_B32_e32 $vgpr0, $vgpr1, implicit $exec
+ $vgpr6 = V_MUL_F32_e32 killed $vgpr300, $vgpr0, implicit $mode, implicit $exec
+...
+
+---
+name: vopd_combine_imm_src0
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: vopd_combine_imm_src0
+ ; SCHED: $vgpr0 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr300 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 0, $vgpr1, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr300, $vgpr0, implicit $mode, implicit $exec
+ ; SCHED-NEXT: $vgpr4 = V_BFM_B32_e32 killed $vgpr0, killed $vgpr1, implicit $exec
+ ;
+ ; PAIR-LABEL: name: vopd_combine_imm_src0
+ ; PAIR: $vgpr0 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr300 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr3, $vgpr6 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32_gfx1250 0, $vgpr1, $vgpr300, $vgpr0, implicit $mode, implicit $exec, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+ ; PAIR-NEXT: $vgpr4 = V_BFM_B32_e32 killed $vgpr0, killed $vgpr1, implicit $exec
+ $vgpr0 = IMPLICIT_DEF
+ $vgpr1 = IMPLICIT_DEF
+ $vgpr300 = IMPLICIT_DEF
+ $vgpr3 = V_SUB_F32_e32 0, $vgpr1, implicit $mode, implicit $exec
+ $vgpr4 = V_BFM_B32_e32 $vgpr0, $vgpr1, implicit $exec
+ $vgpr6 = V_MUL_F32_e32 killed $vgpr300, $vgpr0, implicit $mode, implicit $exec
+...
+
---
name: vopd_mov_max_i32
tracksRegLiveness: true
@@ -657,6 +902,34 @@ body: |
$vgpr5 = V_ASHRREV_I32_e32 $vgpr2, $vgpr3, implicit $mode, implicit $exec
...
+---
+name: vopd_combine_lshr_lshr
+tracksRegLiveness: true
+body: |
+ bb.0:
+
+ ; SCHED-LABEL: name: vopd_combine_lshr_lshr
+ ; SCHED: $vgpr300 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr301 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr302 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr303 = IMPLICIT_DEF
+ ; SCHED-NEXT: $vgpr304 = V_LSHRREV_B32_e32 $vgpr300, $vgpr301, implicit $exec
+ ; SCHED-NEXT: $vgpr305 = V_LSHRREV_B32_e32 $vgpr302, $vgpr303, implicit $mode, implicit $exec
+ ;
+ ; PAIR-LABEL: name: vopd_combine_lshr_lshr
+ ; PAIR: $vgpr300 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr301 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr302 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr303 = IMPLICIT_DEF
+ ; PAIR-NEXT: $vgpr304, $vgpr305 = V_DUAL_LSHRREV_B32_e32_X_LSHRREV_B32_e32_e96_gfx1250 $vgpr300, $vgpr301, $vgpr302, $vgpr303, implicit $exec, implicit $exec, implicit $mode, implicit $exec
+ $vgpr300 = IMPLICIT_DEF
+ $vgpr301 = IMPLICIT_DEF
+ $vgpr302 = IMPLICIT_DEF
+ $vgpr303 = IMPLICIT_DEF
+ $vgpr304 = V_LSHRREV_B32_e32 $vgpr300, $vgpr301, implicit $exec
+ $vgpr305 = V_LSHRREV_B32_e32 $vgpr302, $vgpr303, implicit $mode, implicit $exec
+...
+
---
name: vopd_combine_sub_u32_sub_u32
tracksRegLiveness: true
More information about the llvm-commits
mailing list