This is an archive of the discontinued LLVM Phabricator instance.

Differential D81606

[AMDGPU] Fixed promote alloca with ptr/int casts
ClosedPublic

Authored by rampitec on Jun 10 2020, 11:25 AM.

Details

Reviewers
arsenm
Commits
rG6e1eee6034f7: [AMDGPU] Fixed promote alloca with ptr/int casts
Summary

There is an invalid cast produced when a pointee is a pointer
and the alloca type is cast to a pointer to int.

Diff Detail

Event Timeline

rampitec created this revision.Jun 10 2020, 11:25 AM
Herald added a project: Restricted Project. · View Herald TranscriptJun 10 2020, 11:25 AM
Herald added subscribers: kerbowa, hiraditya, t-tye and 7 others. · View Herald Transcript
arsenm accepted this revision.Jun 10 2020, 11:30 AM
This revision is now accepted and ready to land.Jun 10 2020, 11:30 AM
Closed by commit rG6e1eee6034f7: [AMDGPU] Fixed promote alloca with ptr/int casts (authored by rampitec). · Explain WhyJun 10 2020, 11:47 AM
This revision was automatically updated to reflect the committed changes.

Revision Contents

PathSize
llvm/
lib/
Target/
AMDGPU/
4 lines
test/
CodeGen/
AMDGPU/
28 lines

Diff 269920

llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp

Show First 20 Lines • Show All 502 Lines • ▼ Show 20 Lines case Instruction::Load: {
if (!Index) if (!Index)
break; break;
Type *VecPtrTy = VectorTy->getPointerTo(AMDGPUAS::PRIVATE_ADDRESS); Type *VecPtrTy = VectorTy->getPointerTo(AMDGPUAS::PRIVATE_ADDRESS);
Value *BitCast = Builder.CreateBitCast(Alloca, VecPtrTy); Value *BitCast = Builder.CreateBitCast(Alloca, VecPtrTy);
Value *VecValue = Builder.CreateLoad(VectorTy, BitCast); Value *VecValue = Builder.CreateLoad(VectorTy, BitCast);
Value *ExtractElement = Builder.CreateExtractElement(VecValue, Index); Value *ExtractElement = Builder.CreateExtractElement(VecValue, Index);
if (Inst->getType() != VecEltTy) if (Inst->getType() != VecEltTy)
ExtractElement = Builder.CreateBitCast(ExtractElement, Inst->getType()); ExtractElement = Builder.CreateBitOrPointerCast(ExtractElement, Inst->getType());
Inst->replaceAllUsesWith(ExtractElement); Inst->replaceAllUsesWith(ExtractElement);
Inst->eraseFromParent(); Inst->eraseFromParent();
break; break;
} }
case Instruction::Store: { case Instruction::Store: {
StoreInst *SI = cast<StoreInst>(Inst); StoreInst *SI = cast<StoreInst>(Inst);
if (SI->getValueOperand()->getType() == AllocaTy || if (SI->getValueOperand()->getType() == AllocaTy ||
SI->getValueOperand()->getType()->isVectorTy()) SI->getValueOperand()->getType()->isVectorTy())
break; break;
Value *Ptr = SI->getPointerOperand(); Value *Ptr = SI->getPointerOperand();
Value *Index = calculateVectorIndex(Ptr, GEPVectorIdx); Value *Index = calculateVectorIndex(Ptr, GEPVectorIdx);
if (!Index) if (!Index)
break; break;
Type *VecPtrTy = VectorTy->getPointerTo(AMDGPUAS::PRIVATE_ADDRESS); Type *VecPtrTy = VectorTy->getPointerTo(AMDGPUAS::PRIVATE_ADDRESS);
Value *BitCast = Builder.CreateBitCast(Alloca, VecPtrTy); Value *BitCast = Builder.CreateBitCast(Alloca, VecPtrTy);
Value *VecValue = Builder.CreateLoad(VectorTy, BitCast); Value *VecValue = Builder.CreateLoad(VectorTy, BitCast);
Value *Elt = SI->getValueOperand(); Value *Elt = SI->getValueOperand();
if (Elt->getType() != VecEltTy) if (Elt->getType() != VecEltTy)
Elt = Builder.CreateBitCast(Elt, VecEltTy); Elt = Builder.CreateBitOrPointerCast(Elt, VecEltTy);
Value *NewVecValue = Builder.CreateInsertElement(VecValue, Elt, Index); Value *NewVecValue = Builder.CreateInsertElement(VecValue, Elt, Index);
Builder.CreateStore(NewVecValue, BitCast); Builder.CreateStore(NewVecValue, BitCast);
Inst->eraseFromParent(); Inst->eraseFromParent();
break; break;
} }
default: default:
llvm_unreachable("Inconsistency in instructions promotable to vector"); llvm_unreachable("Inconsistency in instructions promotable to vector");
▲ Show 20 Lines • Show All 516 Lines • Show Last 20 Lines

llvm/test/CodeGen/AMDGPU/promote-alloca-pointer-array.ll

  • This file was added.
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -S -mtriple=amdgcn-- -mcpu=fiji -data-layout=A5 -amdgpu-promote-alloca < %s | FileCheck -check-prefix=OPT %s
define i64 @test_pointer_array(i64 %v) {
; OPT-LABEL: @test_pointer_array(
; OPT-NEXT: entry:
; OPT-NEXT: [[A:%.*]] = alloca [3 x i8*], align 16, addrspace(5)
; OPT-NEXT: [[GEP:%.*]] = getelementptr inbounds [3 x i8*], [3 x i8*] addrspace(5)* [[A]], i32 0, i32 0
; OPT-NEXT: [[CAST:%.*]] = bitcast i8* addrspace(5)* [[GEP]] to i64 addrspace(5)*
; OPT-NEXT: [[TMP0:%.*]] = bitcast [3 x i8*] addrspace(5)* [[A]] to <3 x i8*> addrspace(5)*
; OPT-NEXT: [[TMP1:%.*]] = load <3 x i8*>, <3 x i8*> addrspace(5)* [[TMP0]], align 32
; OPT-NEXT: [[TMP2:%.*]] = inttoptr i64 [[V:%.*]] to i8*
; OPT-NEXT: [[TMP3:%.*]] = insertelement <3 x i8*> [[TMP1]], i8* [[TMP2]], i32 0
; OPT-NEXT: store <3 x i8*> [[TMP3]], <3 x i8*> addrspace(5)* [[TMP0]], align 32
; OPT-NEXT: [[TMP4:%.*]] = bitcast [3 x i8*] addrspace(5)* [[A]] to <3 x i8*> addrspace(5)*
; OPT-NEXT: [[TMP5:%.*]] = load <3 x i8*>, <3 x i8*> addrspace(5)* [[TMP4]], align 32
; OPT-NEXT: [[TMP6:%.*]] = extractelement <3 x i8*> [[TMP5]], i32 0
; OPT-NEXT: [[TMP7:%.*]] = ptrtoint i8* [[TMP6]] to i64
; OPT-NEXT: ret i64 [[TMP7]]
;
entry:
%a = alloca [3 x i8*], align 16, addrspace(5)
%gep = getelementptr inbounds [3 x i8*], [3 x i8*] addrspace(5)* %a, i32 0, i32 0
%cast = bitcast i8* addrspace(5)* %gep to i64 addrspace(5)*
store i64 %v, i64 addrspace(5)* %cast, align 16
%ld = load i64, i64 addrspace(5)* %cast, align 16
ret i64 %ld
}