[SLP]Improve minbitwidth analysis for operands of IToFP and ICmp instructions. #85966
Conversation
This file contains 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
Created using spr 1.3.5
|
@llvm/pr-subscribers-llvm-transforms Author: Alexey Bataev (alexey-bataev) Changesinstructions. Compiler can improve analysis for operands of UIToFP/SIToFP instructions Full diff: https://github.com/llvm/llvm-project/pull/85966.diff 3 Files Affected:
diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index a52064e5417b27..d47c395c012c7a 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -1088,7 +1088,7 @@ class BoUpSLP {
MinBWs.clear();
ReductionBitWidth = 0;
CastMaxMinBWSizes.reset();
- TruncNodes.clear();
+ ExtraBitWidthNodes.clear();
InstrElementSize.clear();
UserIgnoreList = nullptr;
PostponedGathers.clear();
@@ -3662,7 +3662,7 @@ class BoUpSLP {
std::optional<std::pair<unsigned, unsigned>> CastMaxMinBWSizes;
/// Indices of the vectorized trunc nodes.
- DenseSet<unsigned> TruncNodes;
+ DenseSet<unsigned> ExtraBitWidthNodes;
};
} // end namespace slpvectorizer
@@ -6595,7 +6595,13 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
PrevMaxBW),
std::min<unsigned>(DL->getTypeSizeInBits(VL0->getType()),
PrevMinBW));
- TruncNodes.insert(VectorizableTree.size());
+ ExtraBitWidthNodes.insert(VectorizableTree.size() + 1);
+ } else if (ShuffleOrOp == Instruction::SIToFP ||
+ ShuffleOrOp == Instruction::UIToFP) {
+ if (ComputeNumSignBits(VL0->getOperand(0), *DL, 0, AC, nullptr, DT) *
+ 2 >=
+ DL->getTypeSizeInBits(VL0->getOperand(0)->getType()))
+ ExtraBitWidthNodes.insert(VectorizableTree.size() + 1);
}
TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
ReuseShuffleIndicies);
@@ -6643,6 +6649,16 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
TE->setOperand(1, Right);
buildTree_rec(Left, Depth + 1, {TE, 0});
buildTree_rec(Right, Depth + 1, {TE, 1});
+ if (ShuffleOrOp == Instruction::ICmp) {
+ if (ComputeNumSignBits(VL0->getOperand(0), *DL, 0, AC, nullptr, DT) *
+ 2 >=
+ DL->getTypeSizeInBits(VL0->getOperand(0)->getType()))
+ ExtraBitWidthNodes.insert(getOperandEntry(TE, 0)->Idx);
+ if (ComputeNumSignBits(VL0->getOperand(1), *DL, 0, AC, nullptr, DT) *
+ 2 >=
+ DL->getTypeSizeInBits(VL0->getOperand(1)->getType()))
+ ExtraBitWidthNodes.insert(getOperandEntry(TE, 1)->Idx);
+ }
return;
}
case Instruction::Select:
@@ -12468,12 +12484,12 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E, bool PostponedPHIs) {
if (UseIntrinsic && isVectorIntrinsicWithOverloadTypeAtArg(ID, -1))
TysForDecl.push_back(
FixedVectorType::get(CI->getType(), E->Scalars.size()));
+ auto *CEI = cast<CallInst>(VL0);
for (unsigned I : seq<unsigned>(0, CI->arg_size())) {
ValueList OpVL;
// Some intrinsics have scalar arguments. This argument should not be
// vectorized.
if (UseIntrinsic && isVectorIntrinsicWithScalarOpAtArg(ID, I)) {
- CallInst *CEI = cast<CallInst>(VL0);
ScalarArg = CEI->getArgOperand(I);
OpVecs.push_back(CEI->getArgOperand(I));
if (isVectorIntrinsicWithOverloadTypeAtArg(ID, I))
@@ -12486,6 +12502,13 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E, bool PostponedPHIs) {
LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n");
return E->VectorizedValue;
}
+ ScalarArg = CEI->getArgOperand(I);
+ if (cast<VectorType>(OpVec->getType())->getElementType() !=
+ ScalarArg->getType()) {
+ auto *CastTy = FixedVectorType::get(ScalarArg->getType(),
+ VecTy->getNumElements());
+ OpVec = Builder.CreateIntCast(OpVec, CastTy, GetOperandSignedness(I));
+ }
LLVM_DEBUG(dbgs() << "SLP: OpVec[" << I << "]: " << *OpVec << "\n");
OpVecs.push_back(OpVec);
if (UseIntrinsic && isVectorIntrinsicWithOverloadTypeAtArg(ID, I))
@@ -14213,7 +14236,7 @@ void BoUpSLP::computeMinimumValueSizes() {
bool IsStoreOrInsertElt =
VectorizableTree.front()->getOpcode() == Instruction::Store ||
VectorizableTree.front()->getOpcode() == Instruction::InsertElement;
- if ((IsStoreOrInsertElt || UserIgnoreList) && TruncNodes.size() <= 1 &&
+ if ((IsStoreOrInsertElt || UserIgnoreList) && ExtraBitWidthNodes.size() <= 1 &&
(!CastMaxMinBWSizes || CastMaxMinBWSizes->second == 0 ||
CastMaxMinBWSizes->first / CastMaxMinBWSizes->second <= 2))
return;
@@ -14398,11 +14421,11 @@ void BoUpSLP::computeMinimumValueSizes() {
IsTopRoot = false;
IsProfitableToDemoteRoot = true;
- if (TruncNodes.empty()) {
+ if (ExtraBitWidthNodes.empty()) {
NodeIdx = VectorizableTree.size();
} else {
- NodeIdx = *TruncNodes.begin() + 1;
- TruncNodes.erase(TruncNodes.begin());
+ NodeIdx = *ExtraBitWidthNodes.begin();
+ ExtraBitWidthNodes.erase(ExtraBitWidthNodes.begin());
IsTruncRoot = true;
}
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-icmp-to-trunc.ll b/llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-icmp-to-trunc.ll
index fc28d7ab4ee746..c640b1ed63cc03 100644
--- a/llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-icmp-to-trunc.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-icmp-to-trunc.ll
@@ -19,8 +19,8 @@ define i1 @test(ptr noalias %0, i64 %1, ptr noalias %p, ptr %p1) {
; CHECK-NEXT: [[TMP9:%.*]] = icmp eq <2 x i24> [[TMP8]], <i24 24, i24 24>
; CHECK-NEXT: [[TMP10:%.*]] = select <2 x i1> [[TMP9]], <2 x i24> <i24 23, i24 23>, <2 x i24> [[TMP8]]
; CHECK-NEXT: [[TMP23:%.*]] = trunc <2 x i24> [[TMP10]] to <2 x i8>
-; CHECK-NEXT: [[TMP11:%.*]] = zext <2 x i8> [[TMP23]] to <2 x i32>
-; CHECK-NEXT: [[TMP12:%.*]] = and <2 x i32> [[TMP11]], <i32 254, i32 254>
+; CHECK-NEXT: [[TMP26:%.*]] = and <2 x i8> [[TMP23]], <i8 -2, i8 -2>
+; CHECK-NEXT: [[TMP12:%.*]] = zext <2 x i8> [[TMP26]] to <2 x i32>
; CHECK-NEXT: [[TMP13:%.*]] = icmp eq <2 x i32> [[TMP12]], <i32 4, i32 4>
; CHECK-NEXT: [[TMP25:%.*]] = select <2 x i1> [[TMP13]], <2 x i8> <i8 2, i8 2>, <2 x i8> [[TMP23]]
; CHECK-NEXT: [[TMP14:%.*]] = zext <2 x i8> [[TMP25]] to <2 x i32>
diff --git a/llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-node-with-multi-users.ll b/llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-node-with-multi-users.ll
index 136ab64007732f..668d3c3c8c82c5 100644
--- a/llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-node-with-multi-users.ll
+++ b/llvm/test/Transforms/SLPVectorizer/X86/minbitwidth-node-with-multi-users.ll
@@ -10,12 +10,14 @@ define void @test() {
; CHECK-NEXT: [[TMP3:%.*]] = select i1 false, i32 0, i32 0
; CHECK-NEXT: [[TMP4:%.*]] = insertelement <4 x i8> <i8 poison, i8 0, i8 poison, i8 poison>, i8 [[TMP1]], i32 0
; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <4 x i8> [[TMP4]], <4 x i8> poison, <4 x i32> <i32 0, i32 0, i32 0, i32 1>
-; CHECK-NEXT: [[TMP6:%.*]] = sext <4 x i8> [[TMP5]] to <4 x i32>
+; CHECK-NEXT: [[TMP15:%.*]] = trunc <4 x i8> [[TMP5]] to <4 x i1>
; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <4 x i8> [[TMP4]], <4 x i8> poison, <4 x i32> zeroinitializer
; CHECK-NEXT: [[TMP8:%.*]] = or <4 x i8> [[TMP7]], zeroinitializer
-; CHECK-NEXT: [[TMP9:%.*]] = sext <4 x i8> [[TMP8]] to <4 x i32>
-; CHECK-NEXT: [[TMP10:%.*]] = or <4 x i32> zeroinitializer, [[TMP6]]
-; CHECK-NEXT: [[TMP11:%.*]] = icmp eq <4 x i32> [[TMP9]], [[TMP10]]
+; CHECK-NEXT: [[TMP9:%.*]] = trunc <4 x i8> [[TMP8]] to <4 x i1>
+; CHECK-NEXT: [[TMP10:%.*]] = or <4 x i1> zeroinitializer, [[TMP15]]
+; CHECK-NEXT: [[TMP11:%.*]] = icmp eq <4 x i1> [[TMP9]], [[TMP10]]
+; CHECK-NEXT: [[TMP16:%.*]] = shufflevector <4 x i1> [[TMP15]], <4 x i1> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
+; CHECK-NEXT: [[TMP6:%.*]] = zext <4 x i1> [[TMP16]] to <4 x i32>
; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <4 x i32> [[TMP6]], <4 x i32> <i32 0, i32 0, i32 poison, i32 0>, <4 x i32> <i32 4, i32 5, i32 2, i32 7>
; CHECK-NEXT: [[TMP13:%.*]] = select <4 x i1> [[TMP11]], <4 x i32> [[TMP12]], <4 x i32> zeroinitializer
; CHECK-NEXT: [[TMP14:%.*]] = call i32 @llvm.vector.reduce.and.v4i32(<4 x i32> [[TMP13]])
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Created using spr 1.3.5
alexey-bataev
changed the title
|
✅ With the latest revision this PR passed the Python code formatter. |
|
Ping! |
RKSimon
reviewed
Mar 27, 2024
Created using spr 1.3.5
|
Ping! |
alexey-bataev
deleted the
users/alexey-bataev/spr/slpimprove-minbitwidth-analysis-for-operands-of-itofp-and-icmp
branch
alexey-bataev
added a commit
that referenced
this pull request
Apr 3, 2024
…ructions. Compiler can improve analysis for operands of UIToFP/SIToFP instructions and operands of ICmp instruction. Reviewers: RKSimon Reviewed By: RKSimon Pull Request: #85966
|
I've bisected a miscompile back to this patch. For input %in1=0x3 and %in2=ffff the input function returns 0, but after slp-vectorizer with this patch is returns 2. |
Thanks for the reproducer, will double check |
Must be fixed in 102a811 |
Yep, thanks |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Compiler can improve analysis for operands of UIToFP/SIToFP instructions
and operands of ICmp instruction.