[AMDGPU] Fix PromoteAlloca size check of alloca for store #72528
Conversation
When storing a subvector too many element were written when the size of the alloca is smaller than the size of the vector store. This patch checks for the minimum of the alloca vector and the store vector to determine the number of elements to store.
|
@llvm/pr-subscribers-backend-amdgpu Author: None (bcahoon) ChangesWhen storing a subvector too many element were written when the Full diff: https://github.com/llvm/llvm-project/pull/72528.diff 2 Files Affected:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
index 29591ddd669c95e..9293f8954cfe2b5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUPromoteAlloca.cpp
@@ -469,6 +469,7 @@ static Value *promoteAllocaUserToVector(
assert(AccessSize.isKnownMultipleOf(DL.getTypeStoreSize(VecEltTy)));
const unsigned NumWrittenElts =
AccessSize / DL.getTypeStoreSize(VecEltTy);
+ const unsigned NumVecElts = VectorTy->getNumElements();
auto *SubVecTy = FixedVectorType::get(VecEltTy, NumWrittenElts);
assert(DL.getTypeStoreSize(SubVecTy) == DL.getTypeStoreSize(AccessTy));
@@ -480,7 +481,8 @@ static Value *promoteAllocaUserToVector(
Val = Builder.CreateBitOrPointerCast(Val, SubVecTy);
Value *CurVec = GetOrLoadCurrentVectorValue();
- for (unsigned K = 0; K < NumWrittenElts; ++K) {
+ for (unsigned K = 0, NumElts = std::min(NumWrittenElts, NumVecElts);
+ K < NumElts; ++K) {
Value *CurIdx =
Builder.CreateAdd(Index, ConstantInt::get(Index->getType(), K));
CurVec = Builder.CreateInsertElement(
diff --git a/llvm/test/CodeGen/AMDGPU/promote-alloca-subvecs.ll b/llvm/test/CodeGen/AMDGPU/promote-alloca-subvecs.ll
index 765aa250a48f4f0..d8e228e27a8360c 100644
--- a/llvm/test/CodeGen/AMDGPU/promote-alloca-subvecs.ll
+++ b/llvm/test/CodeGen/AMDGPU/promote-alloca-subvecs.ll
@@ -458,3 +458,23 @@ finally:
%load = load <4 x i16>, ptr addrspace(5) %ptr.2, align 2
ret <4 x i16> %load
}
+
+
+; Check the case when the alloca is smaller than the vector size.
+define void @math_kernel3(<4 x i32> %store) {
+; CHECK-LABEL: define void @math_kernel3
+; CHECK-SAME: (<4 x i32> [[STORE:%.*]]) {
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[TMP0:%.*]] = extractelement <4 x i32> [[STORE]], i64 0
+; CHECK-NEXT: [[TMP1:%.*]] = insertelement <3 x i32> undef, i32 [[TMP0]], i32 0
+; CHECK-NEXT: [[TMP2:%.*]] = extractelement <4 x i32> [[STORE]], i64 1
+; CHECK-NEXT: [[TMP3:%.*]] = insertelement <3 x i32> [[TMP1]], i32 [[TMP2]], i32 1
+; CHECK-NEXT: [[TMP4:%.*]] = extractelement <4 x i32> [[STORE]], i64 2
+; CHECK-NEXT: [[TMP5:%.*]] = insertelement <3 x i32> [[TMP3]], i32 [[TMP4]], i32 2
+; CHECK-NEXT: ret void
+;
+entry:
+ %res = alloca <3 x i32>, align 16, addrspace(5)
+ store <4 x i32> %store, ptr addrspace(5) %res, align 16
+ ret void
+}
|
|
Hi @Pierre-vh, I decided the change the test to better show the bug. The following is the output prior to the patch. The set of extract/insert element for the 2nd store start with the last value generated by the first set of extract/insert element instruction. The initial value becomes poison because the last insertelement uses an invalid index (outside the range). Let me know if that does/doesn't help. %0 = extractelement <4 x i32> %store1, i64 0 |
|
Good catch! I thought in such cases the insertelement would just be eliminated (as it stores to nothing and it's a no-side-effect operation) but indeed semantics dictate it returns poison. https://llvm.org/docs/LangRef.html#insertelement-instruction |
When storing a subvector, too many element were written when the size of the alloca is smaller than the size of the vector store. This patch checks for the minimum of the alloca vector and the store vector to determine the number of elements to store.
When storing a subvector too many element were written when the
size of the alloca is smaller than the size of the vector store.
This patch checks for the minimum of the alloca vector and the
store vector to determine the number of elements to store.