Skip to content

[SLP]Allow matching and shuffling of extractelement vector operands with different VF. #97414

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 2 commits into from
Jul 8, 2024
Merged

[SLP]Allow matching and shuffling of extractelement vector operands with different VF. #97414

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4ad68bf
[𝘀𝗽𝗿] initial version
alexey-bataev Jul 2, 2024
48c73a9
Rebase
alexey-bataev Jul 3, 2024
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
99 changes: 46 additions & 53 deletions llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -510,11 +510,17 @@ isFixedVectorShuffle(ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask) {
const auto *It = find_if(VL, IsaPred<ExtractElementInst>);
if (It == VL.end())
return std::nullopt;
auto *EI0 = cast<ExtractElementInst>(*It);
if (isa<ScalableVectorType>(EI0->getVectorOperandType()))
return std::nullopt;
unsigned Size =
cast<FixedVectorType>(EI0->getVectorOperandType())->getNumElements();
std::accumulate(VL.begin(), VL.end(), 0u, [](unsigned S, Value *V) {
auto *EI = dyn_cast<ExtractElementInst>(V);
if (!EI)
return S;
auto *VTy = dyn_cast<FixedVectorType>(EI->getVectorOperandType());
if (!VTy)
return S;
return std::max(S, VTy->getNumElements());
});

Value *Vec1 = nullptr;
Value *Vec2 = nullptr;
bool HasNonUndefVec = any_of(VL, [](Value *V) {
Expand Down Expand Up @@ -544,8 +550,6 @@ isFixedVectorShuffle(ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask) {
if (isa<UndefValue>(Vec)) {
Mask[I] = I;
} else {
if (cast<FixedVectorType>(Vec->getType())->getNumElements() != Size)
return std::nullopt;
if (isa<UndefValue>(EI->getIndexOperand()))
continue;
auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand());
Expand Down Expand Up @@ -10660,36 +10664,20 @@ BoUpSLP::tryToGatherSingleRegisterExtractElements(
VectorOpToIdx[EI->getVectorOperand()].push_back(I);
}
// Sort the vector operands by the maximum number of uses in extractelements.
MapVector<unsigned, SmallVector<Value *>> VFToVector;
for (const auto &Data : VectorOpToIdx)
VFToVector[cast<FixedVectorType>(Data.first->getType())->getNumElements()]
.push_back(Data.first);
for (auto &Data : VFToVector) {
stable_sort(Data.second, [&VectorOpToIdx](Value *V1, Value *V2) {
return VectorOpToIdx.find(V1)->second.size() >
VectorOpToIdx.find(V2)->second.size();
});
}
// Find the best pair of the vectors with the same number of elements or a
// single vector.
SmallVector<std::pair<Value *, SmallVector<int>>> Vectors =
VectorOpToIdx.takeVector();
stable_sort(Vectors, [](const auto &P1, const auto &P2) {
return P1.second.size() > P2.second.size();
});
// Find the best pair of the vectors or a single vector.
const int UndefSz = UndefVectorExtracts.size();
unsigned SingleMax = 0;
Value *SingleVec = nullptr;
unsigned PairMax = 0;
std::pair<Value *, Value *> PairVec(nullptr, nullptr);
for (auto &Data : VFToVector) {
Value *V1 = Data.second.front();
if (SingleMax < VectorOpToIdx[V1].size() + UndefSz) {
SingleMax = VectorOpToIdx[V1].size() + UndefSz;
SingleVec = V1;
}
Value *V2 = nullptr;
if (Data.second.size() > 1)
V2 = *std::next(Data.second.begin());
if (V2 && PairMax < VectorOpToIdx[V1].size() + VectorOpToIdx[V2].size() +
UndefSz) {
PairMax = VectorOpToIdx[V1].size() + VectorOpToIdx[V2].size() + UndefSz;
PairVec = std::make_pair(V1, V2);
if (!Vectors.empty()) {
SingleMax = Vectors.front().second.size() + UndefSz;
if (Vectors.size() > 1) {
auto *ItNext = std::next(Vectors.begin());
PairMax = SingleMax + ItNext->second.size();
}
}
if (SingleMax == 0 && PairMax == 0 && UndefSz == 0)
Expand All @@ -10700,11 +10688,11 @@ BoUpSLP::tryToGatherSingleRegisterExtractElements(
SmallVector<Value *> GatheredExtracts(
VL.size(), PoisonValue::get(VL.front()->getType()));
if (SingleMax >= PairMax && SingleMax) {
for (int Idx : VectorOpToIdx[SingleVec])
for (int Idx : Vectors.front().second)
std::swap(GatheredExtracts[Idx], VL[Idx]);
} else {
for (Value *V : {PairVec.first, PairVec.second})
for (int Idx : VectorOpToIdx[V])
} else if (!Vectors.empty()) {
for (unsigned Idx : {0, 1})
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

why can we guarantee that Vectors.size() == 2 here (but not earlier)?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Because we checked SingleMax != 0 in the first condition, which is a guarantee Vectors is not empty. Here I can change the condition just to check that PairMax != 0 instead.

for (int Idx : Vectors[Idx].second)
std::swap(GatheredExtracts[Idx], VL[Idx]);
}
// Add extracts from undefs too.
Expand Down Expand Up @@ -11773,25 +11761,29 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
MutableArrayRef<int> SubMask = Mask.slice(Part * SliceSize, Limit);
constexpr int MaxBases = 2;
SmallVector<Value *, MaxBases> Bases(MaxBases);
#ifndef NDEBUG
int PrevSize = 0;
#endif // NDEBUG
for (const auto [I, V]: enumerate(VL)) {
if (SubMask[I] == PoisonMaskElem)
auto VLMask = zip(VL, SubMask);
const unsigned VF = std::accumulate(
VLMask.begin(), VLMask.end(), 0U, [&](unsigned S, const auto &D) {
if (std::get<1>(D) == PoisonMaskElem)
return S;
Value *VecOp =
cast<ExtractElementInst>(std::get<0>(D))->getVectorOperand();
if (const TreeEntry *TE = R.getTreeEntry(VecOp))
VecOp = TE->VectorizedValue;
assert(VecOp && "Expected vectorized value.");
const unsigned Size =
cast<FixedVectorType>(VecOp->getType())->getNumElements();
return std::max(S, Size);
});
for (const auto [V, I] : VLMask) {
if (I == PoisonMaskElem)
continue;
Value *VecOp = cast<ExtractElementInst>(V)->getVectorOperand();
if (const TreeEntry *TE = R.getTreeEntry(VecOp))
VecOp = TE->VectorizedValue;
assert(VecOp && "Expected vectorized value.");
const int Size =
cast<FixedVectorType>(VecOp->getType())->getNumElements();
#ifndef NDEBUG
assert((PrevSize == Size || PrevSize == 0) &&
"Expected vectors of the same size.");
PrevSize = Size;
#endif // NDEBUG
VecOp = castToScalarTyElem(VecOp);
Bases[SubMask[I] < Size ? 0 : 1] = VecOp;
Bases[I / VF] = VecOp;
}
if (!Bases.front())
continue;
Expand All @@ -11817,16 +11809,17 @@ class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
"Expected first part or all previous parts masked.");
copy(SubMask, std::next(VecMask.begin(), Part * SliceSize));
} else {
unsigned VF = cast<FixedVectorType>(Vec->getType())->getNumElements();
unsigned NewVF =
cast<FixedVectorType>(Vec->getType())->getNumElements();
if (Vec->getType() != SubVec->getType()) {
unsigned SubVecVF =
cast<FixedVectorType>(SubVec->getType())->getNumElements();
VF = std::max(VF, SubVecVF);
NewVF = std::max(NewVF, SubVecVF);
}
// Adjust SubMask.
for (int &Idx : SubMask)
if (Idx != PoisonMaskElem)
Idx += VF;
Idx += NewVF;
copy(SubMask, std::next(VecMask.begin(), Part * SliceSize));
Vec = createShuffle(Vec, SubVec, VecMask);
TransformToIdentity(VecMask);
Expand Down
10 changes: 4 additions & 6 deletions llvm/test/Transforms/SLPVectorizer/X86/extractelement-single-use-many-nodes.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,11 +9,10 @@ define void @foo(double %i) {
; CHECK-NEXT: [[TMP1:%.*]] = fsub <4 x double> zeroinitializer, [[TMP0]]
; CHECK-NEXT: [[TMP3:%.*]] = insertelement <2 x double> poison, double [[I]], i32 0
; CHECK-NEXT: [[TMP4:%.*]] = fsub <2 x double> zeroinitializer, [[TMP3]]
; CHECK-NEXT: [[TMP5:%.*]] = extractelement <2 x double> [[TMP4]], i32 1
; CHECK-NEXT: [[TMP9:%.*]] = shufflevector <4 x double> [[TMP1]], <4 x double> poison, <8 x i32> <i32 0, i32 poison, i32 poison, i32 1, i32 poison, i32 0, i32 poison, i32 1>
; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <8 x double> [[TMP9]], <8 x double> <double poison, double 0.000000e+00, double poison, double poison, double 0.000000e+00, double poison, double poison, double poison>, <8 x i32> <i32 0, i32 9, i32 poison, i32 3, i32 12, i32 5, i32 poison, i32 7>
; CHECK-NEXT: [[TMP7:%.*]] = insertelement <8 x double> [[TMP6]], double [[TMP5]], i32 2
; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <8 x double> [[TMP7]], <8 x double> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 2, i32 7>
; CHECK-NEXT: [[TMP22:%.*]] = shufflevector <2 x double> [[TMP4]], <2 x double> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <4 x double> [[TMP1]], <4 x double> [[TMP22]], <4 x i32> <i32 0, i32 0, i32 5, i32 1>
; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <4 x double> [[TMP5]], <4 x double> [[TMP5]], <8 x i32> <i32 0, i32 poison, i32 2, i32 3, i32 poison, i32 5, i32 6, i32 7>
; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <8 x double> [[TMP6]], <8 x double> <double poison, double 0.000000e+00, double poison, double poison, double 0.000000e+00, double poison, double poison, double poison>, <8 x i32> <i32 0, i32 9, i32 2, i32 3, i32 12, i32 5, i32 6, i32 7>
; CHECK-NEXT: [[TMP12:%.*]] = fmul <8 x double> <double 0.000000e+00, double poison, double 0.000000e+00, double 0.000000e+00, double poison, double 0.000000e+00, double 0.000000e+00, double 0.000000e+00>, [[TMP8]]
; CHECK-NEXT: [[TMP13:%.*]] = fadd <8 x double> zeroinitializer, [[TMP12]]
; CHECK-NEXT: [[TMP14:%.*]] = fadd <8 x double> [[TMP13]], zeroinitializer
Expand All @@ -27,7 +26,6 @@ define void @foo(double %i) {
; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[TMP18]], i32 1
; CHECK-NEXT: [[I118:%.*]] = fadd double [[TMP19]], [[TMP20]]
; CHECK-NEXT: [[TMP21:%.*]] = fmul <4 x double> zeroinitializer, [[TMP1]]
; CHECK-NEXT: [[TMP22:%.*]] = shufflevector <2 x double> [[TMP4]], <2 x double> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP23:%.*]] = shufflevector <4 x double> <double 0.000000e+00, double 0.000000e+00, double 0.000000e+00, double poison>, <4 x double> [[TMP22]], <4 x i32> <i32 0, i32 1, i32 2, i32 5>
; CHECK-NEXT: [[TMP24:%.*]] = fadd <4 x double> [[TMP21]], [[TMP23]]
; CHECK-NEXT: [[TMP25:%.*]] = fadd <4 x double> [[TMP24]], zeroinitializer
Expand Down
1 change: 0 additions & 1 deletion llvm/test/Transforms/SLPVectorizer/X86/extractlements-gathered-first-node.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,7 +6,6 @@ define void @test() {
; CHECK-NEXT: bb:
; CHECK-NEXT: [[TMP0:%.*]] = extractelement <4 x i32> zeroinitializer, i32 0
; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x i32> zeroinitializer, i32 0
; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x i32> <i32 0, i32 undef>, i32 [[TMP1]], i32 1
; CHECK-NEXT: [[ICMP:%.*]] = icmp ult i32 [[TMP0]], [[TMP1]]
; CHECK-NEXT: ret void
;
Expand Down
Loading