[SPIR-V] Preserve offset alignment in pointer cast legalization#209251
Open
aobolensk wants to merge 1 commit into
Open
[SPIR-V] Preserve offset alignment in pointer cast legalization#209251aobolensk wants to merge 1 commit into
aobolensk wants to merge 1 commit into
Conversation
Splitting a load/store into per-element accesses reused the original alignment for every element, which could wrongly strengthen or discard alignment Compute each split access alignment via commonAlignment with its DataLayout derived byte offset matching SPIRVLegalizerInfo.cpp
|
@llvm/pr-subscribers-backend-spir-v Author: Arseniy Obolenskiy (aobolensk) ChangesSplitting a load/store into per-element accesses reused the original alignment for every element, which could wrongly strengthen or discard alignment Compute each split access alignment via commonAlignment with its DataLayout derived byte offset matching SPIRVLegalizerInfo.cpp Full diff: https://github.com/llvm/llvm-project/pull/209251.diff 2 Files Affected:
diff --git a/llvm/lib/Target/SPIRV/SPIRVLegalizePointerCast.cpp b/llvm/lib/Target/SPIRV/SPIRVLegalizePointerCast.cpp
index 5832d88af750c..38960b23b26db 100644
--- a/llvm/lib/Target/SPIRV/SPIRVLegalizePointerCast.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVLegalizePointerCast.cpp
@@ -99,8 +99,10 @@ class SPIRVLegalizePointerCastImpl {
// type.
// Returns the loaded value.
Value *loadVectorFromVector(IRBuilder<> &B, FixedVectorType *SourceType,
- FixedVectorType *TargetType, Value *Source) {
+ FixedVectorType *TargetType, Value *Source,
+ Align OriginalAlign) {
LoadInst *NewLoad = B.CreateLoad(SourceType, Source);
+ NewLoad->setAlignment(OriginalAlign);
buildAssignType(B, SourceType, NewLoad);
Value *AssignValue = NewLoad;
if (TargetType->getElementType() != SourceType->getElementType()) {
@@ -226,11 +228,11 @@ class SPIRVLegalizePointerCastImpl {
return LI;
}
if (SVT && DVT)
- return loadVectorFromVector(B, SVT, DVT, GEP);
+ return loadVectorFromVector(B, SVT, DVT, GEP, BadLoad->getAlign());
if (SAT && DVT && SAT->getElementType() == DVT->getElementType())
- return loadVectorFromArray(B, DVT, GEP);
+ return loadVectorFromArray(B, DVT, GEP, BadLoad->getAlign());
if (MAT && DVT && MAT->getElementType() == DVT->getElementType())
- return loadVectorFromMatrixArray(B, DVT, GEP, MAT);
+ return loadVectorFromMatrixArray(B, DVT, GEP, MAT, BadLoad->getAlign());
llvm_unreachable("Failed to load from aggregate.");
}
@@ -266,10 +268,12 @@ class SPIRVLegalizePointerCastImpl {
// Loads elements from a matrix with an array of vector memory layout and
// constructs a vector.
Value *loadVectorFromMatrixArray(IRBuilder<> &B, FixedVectorType *TargetType,
- Value *Source,
- FixedVectorType *ArrElemVecTy) {
+ Value *Source, FixedVectorType *ArrElemVecTy,
+ Align OriginalAlign) {
Type *TargetElemTy = TargetType->getElementType();
unsigned ScalarsPerArrayElement = ArrElemVecTy->getNumElements();
+ const DataLayout &DL = B.GetInsertBlock()->getModule()->getDataLayout();
+ uint64_t ArrElemVecSize = DL.getTypeAllocSize(ArrElemVecTy);
// Load each element of the array.
SmallVector<Value *, 4> LoadedElements;
std::array<Type *, 2> Types = {Source->getType(), Source->getType()};
@@ -282,7 +286,9 @@ class SPIRVLegalizePointerCastImpl {
ConstantInt::get(B.getInt32Ty(), ArrayIndex)};
auto *ElementPtr = B.CreateIntrinsic(Intrinsic::spv_gep, {Types}, {Args});
GR->buildAssignPtr(B, ArrElemVecTy, ElementPtr);
- Value *LoadVec = B.CreateLoad(ArrElemVecTy, ElementPtr);
+ LoadInst *LoadVec = B.CreateLoad(ArrElemVecTy, ElementPtr);
+ LoadVec->setAlignment(
+ commonAlignment(OriginalAlign, ArrayIndex * ArrElemVecSize));
buildAssignType(B, ArrElemVecTy, LoadVec);
LoadedElements.push_back(makeExtractElement(B, TargetElemTy, LoadVec,
ElementIndexInArrayElem));
@@ -291,10 +297,12 @@ class SPIRVLegalizePointerCastImpl {
}
// Loads elements from an array and constructs a vector.
Value *loadVectorFromArray(IRBuilder<> &B, FixedVectorType *TargetType,
- Value *Source) {
+ Value *Source, Align OriginalAlign) {
// Load each element of the array.
SmallVector<Value *, 4> LoadedElements;
std::array<Type *, 2> Types = {Source->getType(), Source->getType()};
+ const DataLayout &DL = B.GetInsertBlock()->getModule()->getDataLayout();
+ uint64_t ElemSize = DL.getTypeAllocSize(TargetType->getElementType());
for (unsigned I = 0, E = TargetType->getNumElements(); I < E; ++I) {
// Create a GEP to access the i-th element of the array.
std::array<Value *, 4> Args = {B.getInt1(/*Inbounds=*/false), Source,
@@ -304,7 +312,8 @@ class SPIRVLegalizePointerCastImpl {
GR->buildAssignPtr(B, TargetType->getElementType(), ElementPtr);
// Load the value from the element pointer.
- Value *Load = B.CreateLoad(TargetType->getElementType(), ElementPtr);
+ LoadInst *Load = B.CreateLoad(TargetType->getElementType(), ElementPtr);
+ Load->setAlignment(commonAlignment(OriginalAlign, I * ElemSize));
buildAssignType(B, TargetType->getElementType(), Load);
LoadedElements.push_back(Load);
}
@@ -326,6 +335,8 @@ class SPIRVLegalizePointerCastImpl {
std::array<Type *, 2> Types = {DstArrayPtr->getType(),
DstArrayPtr->getType()};
+ const DataLayout &DL = B.GetInsertBlock()->getModule()->getDataLayout();
+ uint64_t ArrElemVecSize = DL.getTypeAllocSize(ArrElemVecTy);
for (unsigned I = 0; I < SrcNumElements; I += ScalarsPerArrayElement) {
unsigned ArrayIndex = I / ScalarsPerArrayElement;
@@ -344,7 +355,7 @@ class SPIRVLegalizePointerCastImpl {
// Build a vector from the extracted elements and store it.
Value *Vec = buildVectorFromLoadedElements(B, ArrElemVecTy, Elements);
StoreInst *SI = B.CreateStore(Vec, ElementPtr);
- SI->setAlignment(Alignment);
+ SI->setAlignment(commonAlignment(Alignment, ArrayIndex * ArrElemVecSize));
}
}
@@ -360,6 +371,8 @@ class SPIRVLegalizePointerCastImpl {
"Element types of array and vector must be the same.");
std::array<Type *, 2> Types = {DstArrayPtr->getType(),
DstArrayPtr->getType()};
+ const DataLayout &DL = B.GetInsertBlock()->getModule()->getDataLayout();
+ uint64_t ElemSize = DL.getTypeAllocSize(ElemTy);
for (unsigned I = 0, E = VecTy->getNumElements(); I < E; ++I) {
// Create a GEP to access the i-th element of the array.
@@ -373,7 +386,7 @@ class SPIRVLegalizePointerCastImpl {
Value *Element =
E == 1 ? SrcVector : makeExtractElement(B, ElemTy, SrcVector, I);
StoreInst *SI = B.CreateStore(Element, ElementPtr);
- SI->setAlignment(Alignment);
+ SI->setAlignment(commonAlignment(Alignment, I * ElemSize));
}
}
diff --git a/llvm/test/CodeGen/SPIRV/passes/SPIRVLegalizePointerCast.ll b/llvm/test/CodeGen/SPIRV/passes/SPIRVLegalizePointerCast.ll
index 7d2fb71275290..e4ad402c1ec95 100644
--- a/llvm/test/CodeGen/SPIRV/passes/SPIRVLegalizePointerCast.ll
+++ b/llvm/test/CodeGen/SPIRV/passes/SPIRVLegalizePointerCast.ll
@@ -6,6 +6,8 @@
@M = internal addrspace(10) global [4 x <2 x float>] zeroinitializer, align 4
@OUT = internal addrspace(10) global float zeroinitializer, align 4
+@Arr = internal addrspace(10) global [8 x float] zeroinitializer, align 16
+@OUTV = internal addrspace(10) global <4 x float> zeroinitializer, align 4
; Loading a <5 x float> through a [4 x <2 x float>] forces emit-intrinsics to
; insert spv.ptrcast; legalize-pointer-cast lowers it to typed <2 x float>
@@ -25,4 +27,36 @@ entry:
ret void
}
+; Loading a <4 x float> from an [8 x float] with a base alignment of 16 must
+; not strengthen or discard alignment on the split per-element loads: the
+; alignment of each load is the common alignment of the base align and its
+; byte offset (16, 4, 8, 4 for offsets 0, 4, 8, 12).
+
+define spir_func void @loadAlign() #0 {
+; CHECK-LABEL: define spir_func void @loadAlign(
+; CHECK: load float, ptr addrspace(10) %{{.*}}, align 16
+; CHECK: load float, ptr addrspace(10) %{{.*}}, align 4
+; CHECK: load float, ptr addrspace(10) %{{.*}}, align 8
+; CHECK: load float, ptr addrspace(10) %{{.*}}, align 4
+entry:
+ %v = load <4 x float>, ptr addrspace(10) @Arr, align 16
+ store <4 x float> %v, ptr addrspace(10) @OUTV, align 4
+ ret void
+}
+
+; Storing a <4 x float> into an [8 x float] with a base alignment of 16 must
+; apply the same commonAlignment rule to the split per-element stores.
+
+define spir_func void @storeAlign() #0 {
+; CHECK-LABEL: define spir_func void @storeAlign(
+; CHECK: store float %{{.*}}, ptr addrspace(10) %{{.*}}, align 16
+; CHECK: store float %{{.*}}, ptr addrspace(10) %{{.*}}, align 4
+; CHECK: store float %{{.*}}, ptr addrspace(10) %{{.*}}, align 8
+; CHECK: store float %{{.*}}, ptr addrspace(10) %{{.*}}, align 4
+entry:
+ %v = load <4 x float>, ptr addrspace(10) @OUTV, align 4
+ store <4 x float> %v, ptr addrspace(10) @Arr, align 16
+ ret void
+}
+
attributes #0 = { "hlsl.numthreads"="1,1,1" "hlsl.shader"="compute" }
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Splitting a load/store into per-element accesses reused the original alignment for every element, which could wrongly strengthen or discard alignment
Compute each split access alignment via commonAlignment with its DataLayout derived byte offset matching SPIRVLegalizerInfo.cpp