Skip to content

[SLP]Support vectorization of previously vectorized scalars in split nodes #134286

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Apr 10, 2025
Merged

[SLP]Support vectorization of previously vectorized scalars in split nodes #134286

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
3787680
[𝘀𝗽𝗿] initial version
alexey-bataev Apr 3, 2025
f64d965
Rebase
alexey-bataev Apr 9, 2025
ee0bcbe
Rebase, address comment
alexey-bataev Apr 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
44 changes: 20 additions & 24 deletions llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -895,6 +895,13 @@ class InstructionsState {
is_contained(AddSub, getAltOpcode());
}

/// Checks if main/alt instructions are cmp operations.
bool isCmpOp() const {
return (getOpcode() == Instruction::ICmp ||
getOpcode() == Instruction::FCmp) &&
getAltOpcode() == getOpcode();
}

/// Checks if the current state is valid, i.e. has non-null MainOp
bool valid() const { return MainOp && AltOp; }

Expand Down Expand Up @@ -9277,22 +9284,23 @@ bool BoUpSLP::canBuildSplitNode(ArrayRef<Value *> VL,
// as alternate ops.
if (NumParts >= VL.size())
return false;
constexpr TTI::TargetCostKind Kind = TTI::TCK_RecipThroughput;
InstructionCost InsertCost = ::getShuffleCost(
*TTI, TTI::SK_InsertSubvector, VecTy, {}, Kind, Op1.size(), Op2VecTy);
FixedVectorType *SubVecTy =
getWidenedType(ScalarTy, std::max(Op1.size(), Op2.size()));
InstructionCost NewShuffleCost =
::getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc, SubVecTy, Mask, Kind);
if (!LocalState.isCmpOp() && NumParts <= 1 &&
(Mask.empty() || InsertCost >= NewShuffleCost))
return false;
if ((LocalState.getMainOp()->isBinaryOp() &&
LocalState.getAltOp()->isBinaryOp() &&
(LocalState.isShiftOp() || LocalState.isBitwiseLogicOp() ||
LocalState.isAddSubLikeOp() || LocalState.isMulDivLikeOp())) ||
(LocalState.getMainOp()->isCast() && LocalState.getAltOp()->isCast()) ||
(LocalState.getMainOp()->isUnaryOp() &&
LocalState.getAltOp()->isUnaryOp())) {
constexpr TTI::TargetCostKind Kind = TTI::TCK_RecipThroughput;
InstructionCost InsertCost = ::getShuffleCost(
*TTI, TTI::SK_InsertSubvector, VecTy, {}, Kind, Op1.size(), Op2VecTy);
FixedVectorType *SubVecTy =
getWidenedType(ScalarTy, std::max(Op1.size(), Op2.size()));
InstructionCost NewShuffleCost =
::getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc, SubVecTy, Mask, Kind);
if (NumParts <= 1 && (Mask.empty() || InsertCost >= NewShuffleCost))
return false;
InstructionCost OriginalVecOpsCost =
TTI->getArithmeticInstrCost(Opcode0, VecTy, Kind) +
TTI->getArithmeticInstrCost(Opcode1, VecTy, Kind);
Expand Down Expand Up @@ -9429,18 +9437,6 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
if (!canBuildSplitNode(VL, LocalState, Op1, Op2, ReorderIndices))
return false;

// Any value is used in split node already - just gather.
if (any_of(VL, [&](Value *V) {
return ScalarsInSplitNodes.contains(V) || isVectorized(V);
})) {
if (TryToFindDuplicates(S)) {
auto Invalid = ScheduleBundle::invalid();
newTreeEntry(VL, Invalid /*not vectorized*/, S, UserTreeIdx,
ReuseShuffleIndices);
}
return true;
}

SmallVector<Value *> NewVL(VL.size());
copy(Op1, NewVL.begin());
copy(Op2, std::next(NewVL.begin(), Op1.size()));
Expand Down Expand Up @@ -9616,9 +9612,9 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
::getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc, VecTy, {}, Kind) +
::getScalarizationOverhead(*TTI, ScalarTy, VecTy, Extracted,
/*Insert=*/false, /*Extract=*/true, Kind);
InstructionCost ScalarizeCostEstimate =
::getScalarizationOverhead(*TTI, ScalarTy, VecTy, Vectorized,
/*Insert=*/true, /*Extract=*/false, Kind);
InstructionCost ScalarizeCostEstimate = ::getScalarizationOverhead(
*TTI, ScalarTy, VecTy, Vectorized,
/*Insert=*/true, /*Extract=*/false, Kind, /*ForPoisonSrc=*/false);
PreferScalarize = VectorizeCostEstimate > ScalarizeCostEstimate;
}
if (PreferScalarize) {
Expand Down
20 changes: 12 additions & 8 deletions llvm/test/Transforms/SLPVectorizer/X86/splat-score-adjustment.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,18 +7,22 @@ define i32 @a() {
; CHECK-NEXT: br label %[[BB1:.*]]
; CHECK: [[BB1]]:
; CHECK-NEXT: [[TMP2:%.*]] = phi <4 x i8> [ zeroinitializer, [[TMP0:%.*]] ], [ [[TMP6:%.*]], %[[BB1]] ]
; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <4 x i8> [[TMP2]], <4 x i8> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 0, i32 1, i32 2, i32 3>
; CHECK-NEXT: [[TMP3:%.*]] = phi <2 x i8> [ zeroinitializer, [[TMP0]] ], [ [[TMP17:%.*]], %[[BB1]] ]
; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <4 x i8> [[TMP2]], <4 x i8> poison, <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
; CHECK-NEXT: [[TMP6]] = load <4 x i8>, ptr null, align 4
; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <4 x i8> [[TMP2]], <4 x i8> poison, <4 x i32> <i32 2, i32 3, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <4 x i8> [[TMP7]], <4 x i8> [[TMP6]], <4 x i32> <i32 0, i32 1, i32 4, i32 5>
; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <4 x i8> [[TMP6]], <4 x i8> poison, <4 x i32> <i32 poison, i32 poison, i32 0, i32 1>
; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x i8> [[TMP3]], <2 x i8> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <4 x i8> [[TMP12]], <4 x i8> [[TMP7]], <4 x i32> <i32 4, i32 5, i32 2, i32 3>
; CHECK-NEXT: [[TMP9:%.*]] = xor <4 x i8> [[TMP6]], [[TMP8]]
; CHECK-NEXT: [[TMP10:%.*]] = shufflevector <4 x i8> [[TMP9]], <4 x i8> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP18:%.*]] = call <8 x i8> @llvm.vector.insert.v8i8.v4i8(<8 x i8> [[TMP10]], <4 x i8> [[TMP6]], i64 4)
; CHECK-NEXT: [[TMP21:%.*]] = shufflevector <8 x i8> [[TMP5]], <8 x i8> [[TMP18]], <8 x i32> <i32 1, i32 2, i32 3, i32 12, i32 3, i32 12, i32 13, i32 14>
; CHECK-NEXT: [[TMP10:%.*]] = shufflevector <4 x i8> [[TMP6]], <4 x i8> poison, <8 x i32> <i32 poison, i32 0, i32 poison, i32 1, i32 poison, i32 2, i32 poison, i32 3>
; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <4 x i8> [[TMP9]], <4 x i8> poison, <8 x i32> <i32 0, i32 poison, i32 1, i32 poison, i32 2, i32 poison, i32 3, i32 poison>
; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <8 x i8> [[TMP10]], <8 x i8> [[TMP11]], <8 x i32> <i32 8, i32 1, i32 10, i32 3, i32 12, i32 5, i32 14, i32 7>
; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <4 x i8> [[TMP2]], <4 x i8> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP21:%.*]] = shufflevector <8 x i8> [[TMP13]], <8 x i8> [[TMP18]], <8 x i32> <i32 1, i32 3, i32 2, i32 9, i32 3, i32 11, i32 9, i32 13>
; CHECK-NEXT: [[TMP22:%.*]] = xor <8 x i8> [[TMP18]], [[TMP21]]
; CHECK-NEXT: [[TMP23:%.*]] = xor <8 x i8> [[TMP22]], [[TMP5]]
; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <8 x i8> [[TMP23]], <8 x i8> poison, <8 x i32> <i32 0, i32 4, i32 1, i32 5, i32 2, i32 6, i32 3, i32 7>
; CHECK-NEXT: store <8 x i8> [[TMP13]], ptr null, align 4
; CHECK-NEXT: store <8 x i8> [[TMP23]], ptr null, align 4
; CHECK-NEXT: [[TMP17]] = shufflevector <4 x i8> [[TMP6]], <4 x i8> poison, <2 x i32> <i32 2, i32 3>
; CHECK-NEXT: br label %[[BB1]]
;
br label %1
Expand Down
21 changes: 10 additions & 11 deletions llvm/test/Transforms/SLPVectorizer/X86/split-node-last-inst-vectorized.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,20 +6,19 @@ define void @test(ptr %0, <8 x i8> %1) {
; CHECK-SAME: ptr [[TMP0:%.*]], <8 x i8> [[TMP1:%.*]]) {
; CHECK-NEXT: [[TMP3:%.*]] = load i8, ptr [[TMP0]], align 2
; CHECK-NEXT: [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i64 13436
; CHECK-NEXT: [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i64 13444
; CHECK-NEXT: [[TMP6:%.*]] = load i8, ptr [[TMP8]], align 4
; CHECK-NEXT: [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i64 13544
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i64 13536
; CHECK-NEXT: [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i64 13437
; CHECK-NEXT: [[TMP7:%.*]] = load <8 x i8>, ptr [[TMP4]], align 4
; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> poison, <8 x i32> <i32 0, i32 0, i32 2, i32 3, i32 4, i32 5, i32 0, i32 7>
; CHECK-NEXT: [[TMP9:%.*]] = insertelement <8 x i8> [[TMP7]], i8 [[TMP3]], i32 1
; CHECK-NEXT: [[TMP12:%.*]] = insertelement <8 x i8> poison, i8 [[TMP6]], i32 0
; CHECK-NEXT: [[TMP9:%.*]] = insertelement <8 x i8> [[TMP12]], i8 [[TMP3]], i32 1
; CHECK-NEXT: [[TMP10:%.*]] = shufflevector <8 x i8> [[TMP9]], <8 x i8> poison, <8 x i32> <i32 0, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1, i32 1>
; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <8 x i8> [[TMP8]], <8 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP12:%.*]] = call <16 x i8> @llvm.vector.insert.v16i8.v8i8(<16 x i8> [[TMP11]], <8 x i8> [[TMP10]], i64 8)
; CHECK-NEXT: [[TMP13:%.*]] = load <8 x i8>, ptr [[TMP6]], align 1
; CHECK-NEXT: [[TMP14:%.*]] = shufflevector <8 x i8> [[TMP13]], <8 x i8> poison, <8 x i32> <i32 7, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6>
; CHECK-NEXT: [[TMP15:%.*]] = call <16 x i8> @llvm.vector.insert.v16i8.v8i8(<16 x i8> poison, <8 x i8> [[TMP7]], i64 0)
; CHECK-NEXT: [[TMP16:%.*]] = call <16 x i8> @llvm.vector.insert.v16i8.v8i8(<16 x i8> [[TMP15]], <8 x i8> [[TMP14]], i64 8)
; CHECK-NEXT: [[TMP17:%.*]] = or <16 x i8> [[TMP16]], [[TMP12]]
; CHECK-NEXT: store <16 x i8> [[TMP17]], ptr [[TMP5]], align 4
; CHECK-NEXT: [[TMP13:%.*]] = or <8 x i8> [[TMP10]], [[TMP7]]
; CHECK-NEXT: store <8 x i8> [[TMP13]], ptr [[TMP11]], align 4
; CHECK-NEXT: [[TMP14:%.*]] = shufflevector <8 x i8> [[TMP1]], <8 x i8> poison, <8 x i32> <i32 0, i32 0, i32 2, i32 3, i32 4, i32 5, i32 0, i32 7>
; CHECK-NEXT: [[TMP15:%.*]] = or <8 x i8> [[TMP7]], [[TMP14]]
; CHECK-NEXT: store <8 x i8> [[TMP15]], ptr [[TMP5]], align 4
; CHECK-NEXT: ret void
;
%3 = load i8, ptr %0, align 2
Expand Down
5 changes: 2 additions & 3 deletions llvm/test/Transforms/SLPVectorizer/X86/split-node-no-reorder-copy.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,9 +16,8 @@ define i1 @test(ptr %0, ptr %1, <2 x float> %2, <2 x float> %3, <2 x float> %4)
; CHECK-NEXT: [[TMP15:%.*]] = insertelement <8 x float> [[TMP14]], float [[TMP9]], i32 7
; CHECK-NEXT: [[TMP16:%.*]] = shufflevector <8 x float> [[TMP13]], <8 x float> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP17:%.*]] = call <16 x float> @llvm.vector.insert.v16f32.v8f32(<16 x float> [[TMP16]], <8 x float> [[TMP15]], i64 8)
; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <8 x float> [[TMP14]], <8 x float> [[TMP12]], <16 x i32> <i32 8, i32 9, i32 9, i32 9, i32 9, i32 9, i32 14, i32 14, i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 poison>
; CHECK-NEXT: [[TMP19:%.*]] = insertelement <16 x float> [[TMP18]], float [[TMP9]], i32 15
; CHECK-NEXT: [[TMP20:%.*]] = fmul <16 x float> [[TMP17]], [[TMP19]]
; CHECK-NEXT: [[TMP18:%.*]] = call <16 x float> @llvm.vector.insert.v16f32.v8f32(<16 x float> [[TMP16]], <8 x float> [[TMP15]], i64 8)
; CHECK-NEXT: [[TMP20:%.*]] = fmul <16 x float> [[TMP18]], [[TMP17]]
; CHECK-NEXT: [[TMP21:%.*]] = call reassoc nsz float @llvm.vector.reduce.fadd.v16f32(float 0.000000e+00, <16 x float> [[TMP20]])
; CHECK-NEXT: [[TMP22:%.*]] = call float @foo(float [[TMP21]])
; CHECK-NEXT: ret i1 false
Expand Down
Loading