[AMDGPU] Fixed promote alloca with ptr/int casts

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

Differential Revision: https://reviews.llvm.org/D81606

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
index 036f544..f26fb9a 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
@@ -508,7 +508,7 @@ static bool tryPromoteAllocaToVector(AllocaInst *Alloca, const DataLayout &DL) {
       Value *VecValue = Builder.CreateLoad(VectorTy, BitCast);
       Value *ExtractElement = Builder.CreateExtractElement(VecValue, Index);
       if (Inst->getType() != VecEltTy)
-        ExtractElement = Builder.CreateBitCast(ExtractElement, Inst->getType());
+        ExtractElement = Builder.CreateBitOrPointerCast(ExtractElement, Inst->getType());
       Inst->replaceAllUsesWith(ExtractElement);
       Inst->eraseFromParent();
       break;
@@ -529,7 +529,7 @@ static bool tryPromoteAllocaToVector(AllocaInst *Alloca, const DataLayout &DL) {
       Value *VecValue = Builder.CreateLoad(VectorTy, BitCast);
       Value *Elt = SI->getValueOperand();
       if (Elt->getType() != VecEltTy)
-        Elt = Builder.CreateBitCast(Elt, VecEltTy);
+        Elt = Builder.CreateBitOrPointerCast(Elt, VecEltTy);
       Value *NewVecValue = Builder.CreateInsertElement(VecValue, Elt, Index);
       Builder.CreateStore(NewVecValue, BitCast);
       Inst->eraseFromParent();

diff --git a/llvm/test/CodeGen/AMDGPU/promote-alloca-pointer-array.ll b/llvm/test/CodeGen/AMDGPU/promote-alloca-pointer-array.ll
new file mode 100644 (file)
index 0000000..14c9775
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/promote-alloca-pointer-array.ll
@@ -0,0 +1,28 @@
+; 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
+}