[llvm] 002a0a2 - AMDGPU: Fix broken frame index expansion for v_add_co_u32_e64 (#114634)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Nov 4 10:39:55 PST 2024
Author: Matt Arsenault
Date: 2024-11-04T10:39:50-08:00
New Revision: 002a0a27bc4702d6f34434c1838cb1698a0b0098
URL: https://github.com/llvm/llvm-project/commit/002a0a27bc4702d6f34434c1838cb1698a0b0098
DIFF: https://github.com/llvm/llvm-project/commit/002a0a27bc4702d6f34434c1838cb1698a0b0098.diff
LOG: AMDGPU: Fix broken frame index expansion for v_add_co_u32_e64 (#114634)
With an explicit carry out operand, one too many operands were deleted
resulting in a malformed v_mov_b32.
Added:
Modified:
llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
llvm/test/CodeGen/AMDGPU/eliminate-frame-index-v-add-co-u32.mir
Removed:
################################################################################
diff --git a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
index 1929df12124f94..8b8884a903caf4 100644
--- a/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp
@@ -2607,7 +2607,7 @@ bool SIRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator MI,
MI->removeOperand(FIOperandNum);
unsigned NumOps = MI->getNumOperands();
- for (unsigned I = NumOps - 2; I >= 2; --I)
+ for (unsigned I = NumOps - 2; I >= NumDefs + 1; --I)
MI->removeOperand(I);
if (NumDefs == 2)
diff --git a/llvm/test/CodeGen/AMDGPU/eliminate-frame-index-v-add-co-u32.mir b/llvm/test/CodeGen/AMDGPU/eliminate-frame-index-v-add-co-u32.mir
index e7e56c1a3fdaa3..f6b0405d5b0097 100644
--- a/llvm/test/CodeGen/AMDGPU/eliminate-frame-index-v-add-co-u32.mir
+++ b/llvm/test/CodeGen/AMDGPU/eliminate-frame-index-v-add-co-u32.mir
@@ -1661,3 +1661,170 @@ body: |
SI_RETURN implicit $vgpr0
...
+
+---
+name: v_add_co_u32_e64__fi_sgpr_kernel
+tracksRegLiveness: true
+stack:
+ - { id: 0, size: 20, alignment: 4 }
+machineFunctionInfo:
+ scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
+ frameOffsetReg: '$sgpr33'
+ stackPtrOffsetReg: '$sgpr32'
+ isEntryFunction: true
+body: |
+ bb.0:
+ liveins: $sgpr4
+
+ ; MUBUFW64-LABEL: name: v_add_co_u32_e64__fi_sgpr_kernel
+ ; MUBUFW64: liveins: $sgpr4, $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: {{ $}}
+ ; MUBUFW64-NEXT: $sgpr0 = S_ADD_U32 $sgpr0, $noreg, implicit-def $scc, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: $sgpr1 = S_ADDC_U32 $sgpr1, 0, implicit-def dead $scc, implicit $scc, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: renamable $vgpr0 = V_MOV_B32_e32 killed $sgpr4, implicit $exec
+ ; MUBUFW64-NEXT: SI_RETURN implicit $vgpr0
+ ;
+ ; FLATSCRW64-LABEL: name: v_add_co_u32_e64__fi_sgpr_kernel
+ ; FLATSCRW64: liveins: $sgpr4
+ ; FLATSCRW64-NEXT: {{ $}}
+ ; FLATSCRW64-NEXT: renamable $vgpr0 = V_MOV_B32_e32 killed $sgpr4, implicit $exec
+ ; FLATSCRW64-NEXT: SI_RETURN implicit $vgpr0
+ renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 %stack.0, killed $sgpr4, 0, implicit $exec
+ SI_RETURN implicit $vgpr0
+
+...
+
+---
+name: v_add_co_u32_e64__fi_sgpr_func
+tracksRegLiveness: true
+stack:
+ - { id: 0, size: 20, alignment: 4 }
+machineFunctionInfo:
+ scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
+ frameOffsetReg: '$sgpr33'
+ stackPtrOffsetReg: '$sgpr32'
+body: |
+ bb.0:
+ liveins: $sgpr4
+
+ ; MUBUFW64-LABEL: name: v_add_co_u32_e64__fi_sgpr_func
+ ; MUBUFW64: liveins: $sgpr4
+ ; MUBUFW64-NEXT: {{ $}}
+ ; MUBUFW64-NEXT: renamable $vgpr1 = V_LSHRREV_B32_e64 6, $sgpr32, implicit $exec
+ ; MUBUFW64-NEXT: renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 killed $sgpr4, killed $vgpr1, 0, implicit $exec
+ ; MUBUFW64-NEXT: SI_RETURN implicit $vgpr0
+ ;
+ ; GFX940-LABEL: name: v_add_co_u32_e64__fi_sgpr_func
+ ; GFX940: liveins: $sgpr4
+ ; GFX940-NEXT: {{ $}}
+ ; GFX940-NEXT: $vgpr1 = V_MOV_B32_e32 $sgpr32, implicit $exec
+ ; GFX940-NEXT: renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 killed $sgpr4, killed $vgpr1, 0, implicit $exec
+ ; GFX940-NEXT: SI_RETURN implicit $vgpr0
+ ;
+ ; GFX11-LABEL: name: v_add_co_u32_e64__fi_sgpr_func
+ ; GFX11: liveins: $sgpr4
+ ; GFX11-NEXT: {{ $}}
+ ; GFX11-NEXT: renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 $sgpr32, killed $sgpr4, 0, implicit $exec
+ ; GFX11-NEXT: SI_RETURN implicit $vgpr0
+ ;
+ ; GFX12-LABEL: name: v_add_co_u32_e64__fi_sgpr_func
+ ; GFX12: liveins: $sgpr4
+ ; GFX12-NEXT: {{ $}}
+ ; GFX12-NEXT: renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 $sgpr32, killed $sgpr4, 0, implicit $exec
+ ; GFX12-NEXT: SI_RETURN implicit $vgpr0
+ renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 %stack.0, killed $sgpr4, 0, implicit $exec
+ SI_RETURN implicit $vgpr0
+
+...
+
+---
+name: v_add_co_u32_e64__fi_inc_same_vgpr_kernel
+tracksRegLiveness: true
+stack:
+ - { id: 0, size: 20, alignment: 4 }
+machineFunctionInfo:
+ scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
+ frameOffsetReg: '$sgpr33'
+ stackPtrOffsetReg: '$sgpr32'
+ isEntryFunction: true
+body: |
+ bb.0:
+ liveins: $vgpr0
+
+ ; MUBUFW64-LABEL: name: v_add_co_u32_e64__fi_inc_same_vgpr_kernel
+ ; MUBUFW64: liveins: $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: {{ $}}
+ ; MUBUFW64-NEXT: $sgpr0 = S_ADD_U32 $sgpr0, $noreg, implicit-def $scc, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: $sgpr1 = S_ADDC_U32 $sgpr1, 0, implicit-def dead $scc, implicit $scc, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: SI_RETURN implicit $vgpr0
+ ;
+ ; FLATSCRW64-LABEL: name: v_add_co_u32_e64__fi_inc_same_vgpr_kernel
+ ; FLATSCRW64: liveins: $vgpr0
+ ; FLATSCRW64-NEXT: {{ $}}
+ ; FLATSCRW64-NEXT: SI_RETURN implicit $vgpr0
+ renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 %stack.0, killed $vgpr0, 0, implicit $exec
+ SI_RETURN implicit $vgpr0
+
+...
+
+---
+name: v_add_co_u32_e64__fi_inc_same_vgpr_func
+tracksRegLiveness: true
+stack:
+ - { id: 0, size: 20, alignment: 4 }
+machineFunctionInfo:
+ scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
+ frameOffsetReg: '$sgpr33'
+ stackPtrOffsetReg: '$sgpr32'
+body: |
+ bb.0:
+ liveins: $vgpr0
+
+ ; MUBUFW64-LABEL: name: v_add_co_u32_e64__fi_inc_same_vgpr_func
+ ; MUBUFW64: liveins: $vgpr0
+ ; MUBUFW64-NEXT: {{ $}}
+ ; MUBUFW64-NEXT: renamable $vgpr1 = V_LSHRREV_B32_e64 6, $sgpr32, implicit $exec
+ ; MUBUFW64-NEXT: renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 killed $vgpr0, killed $vgpr1, 0, implicit $exec
+ ; MUBUFW64-NEXT: SI_RETURN implicit $vgpr0
+ ;
+ ; FLATSCRW64-LABEL: name: v_add_co_u32_e64__fi_inc_same_vgpr_func
+ ; FLATSCRW64: liveins: $vgpr0
+ ; FLATSCRW64-NEXT: {{ $}}
+ ; FLATSCRW64-NEXT: renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 $sgpr32, killed $vgpr0, 0, implicit $exec
+ ; FLATSCRW64-NEXT: SI_RETURN implicit $vgpr0
+ renamable $vgpr0, dead renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 %stack.0, killed $vgpr0, 0, implicit $exec
+ SI_RETURN implicit $vgpr0
+
+...
+
+---
+name: v_add_co_u32_e64__fi_sgpr_kernel_live_co
+tracksRegLiveness: true
+stack:
+ - { id: 0, size: 20, alignment: 4 }
+machineFunctionInfo:
+ scratchRSrcReg: '$sgpr0_sgpr1_sgpr2_sgpr3'
+ frameOffsetReg: '$sgpr33'
+ stackPtrOffsetReg: '$sgpr32'
+ isEntryFunction: true
+body: |
+ bb.0:
+ liveins: $sgpr4
+
+ ; MUBUFW64-LABEL: name: v_add_co_u32_e64__fi_sgpr_kernel_live_co
+ ; MUBUFW64: liveins: $sgpr4, $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: {{ $}}
+ ; MUBUFW64-NEXT: $sgpr0 = S_ADD_U32 $sgpr0, $noreg, implicit-def $scc, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: $sgpr1 = S_ADDC_U32 $sgpr1, 0, implicit-def dead $scc, implicit $scc, implicit-def $sgpr0_sgpr1_sgpr2_sgpr3
+ ; MUBUFW64-NEXT: renamable $vgpr0, renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 0, killed $sgpr4, 0, implicit $exec
+ ; MUBUFW64-NEXT: SI_RETURN implicit $vgpr0, implicit $sgpr4_sgpr5
+ ;
+ ; FLATSCRW64-LABEL: name: v_add_co_u32_e64__fi_sgpr_kernel_live_co
+ ; FLATSCRW64: liveins: $sgpr4
+ ; FLATSCRW64-NEXT: {{ $}}
+ ; FLATSCRW64-NEXT: renamable $vgpr0, renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 0, killed $sgpr4, 0, implicit $exec
+ ; FLATSCRW64-NEXT: SI_RETURN implicit $vgpr0, implicit $sgpr4_sgpr5
+ renamable $vgpr0, renamable $sgpr4_sgpr5 = V_ADD_CO_U32_e64 %stack.0, killed $sgpr4, 0, implicit $exec
+ SI_RETURN implicit $vgpr0, implicit $sgpr4_sgpr5
+
+...
More information about the llvm-commits
mailing list