Skip to content

[SLP]Initial support for non-power-of-2 (but still whole register) number of elements in operands. #107273

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 7 commits into from
Sep 25, 2024
Merged

[SLP]Initial support for non-power-of-2 (but still whole register) number of elements in operands. #107273

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
f927502
[𝘀𝗽𝗿] initial version
alexey-bataev Sep 4, 2024
a79159b
Fix formatting
alexey-bataev Sep 4, 2024
0a78a26
Rebase, address comments
alexey-bataev Sep 6, 2024
b8b1527
Rebase, address comments
alexey-bataev Sep 6, 2024
7c68dbf
Rebase, restored some code, some optimizations
alexey-bataev Sep 9, 2024
affdc53
Rebase, address comments
alexey-bataev Sep 16, 2024
13c2844
Rebase
alexey-bataev Sep 25, 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
14 changes: 13 additions & 1 deletion llvm/include/llvm/CodeGen/BasicTTIImpl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2531,7 +2531,19 @@ class BasicTTIImplBase : public TargetTransformInfoImplCRTPBase<T> {

unsigned getNumberOfParts(Type *Tp) {
std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Tp);
return LT.first.isValid() ? *LT.first.getValue() : 0;
if (!LT.first.isValid())
return 0;
// Try to find actual number of parts for non-power-of-2 elements as
// ceil(num-of-elements/num-of-subtype-elements).
if (auto *FTp = dyn_cast<FixedVectorType>(Tp);
Tp && LT.second.isFixedLengthVector() &&
!has_single_bit(FTp->getNumElements())) {
if (auto *SubTp = dyn_cast_if_present<FixedVectorType>(
EVT(LT.second).getTypeForEVT(Tp->getContext()));
SubTp && SubTp->getElementType() == FTp->getElementType())
return divideCeil(FTp->getNumElements(), SubTp->getNumElements());
}
return *LT.first.getValue();
}

InstructionCost getAddressComputationCost(Type *Ty, ScalarEvolution *,
Expand Down
107 changes: 72 additions & 35 deletions llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -260,6 +260,20 @@ static FixedVectorType *getWidenedType(Type *ScalarTy, unsigned VF) {
VF * getNumElements(ScalarTy));
}

/// Returns the number of elements of the given type \p Ty, not less than \p Sz,
/// which forms type, which splits by \p TTI into whole vector types during
/// legalization.
static unsigned getFullVectorNumberOfElements(const TargetTransformInfo &TTI,
Type *Ty, unsigned Sz) {
if (!isValidElementType(Ty))
return bit_ceil(Sz);
// Find the number of elements, which forms full vectors.
const unsigned NumParts = TTI.getNumberOfParts(getWidenedType(Ty, Sz));
if (NumParts == 0 || NumParts >= Sz)
return bit_ceil(Sz);
return bit_ceil(divideCeil(Sz, NumParts)) * NumParts;
}

static void transformScalarShuffleIndiciesToVector(unsigned VecTyNumElements,
SmallVectorImpl<int> &Mask) {
// The ShuffleBuilder implementation use shufflevector to splat an "element".
Expand Down Expand Up @@ -394,7 +408,7 @@ static bool isVectorLikeInstWithConstOps(Value *V) {
/// total number of elements \p Size and number of registers (parts) \p
/// NumParts.
static unsigned getPartNumElems(unsigned Size, unsigned NumParts) {
return PowerOf2Ceil(divideCeil(Size, NumParts));
return std::min<unsigned>(Size, bit_ceil(divideCeil(Size, NumParts)));
}

/// Returns correct remaining number of elements, considering total amount \p
Expand Down Expand Up @@ -1223,6 +1237,22 @@ static bool doesNotNeedToSchedule(ArrayRef<Value *> VL) {
(all_of(VL, isUsedOutsideBlock) || all_of(VL, areAllOperandsNonInsts));
}

/// Returns true if widened type of \p Ty elements with size \p Sz represents
/// full vector type, i.e. adding extra element results in extra parts upon type
/// legalization.
static bool hasFullVectorsOnly(const TargetTransformInfo &TTI, Type *Ty,
unsigned Sz) {
if (Sz <= 1)
return false;
if (!isValidElementType(Ty) && !isa<FixedVectorType>(Ty))
return false;
if (has_single_bit(Sz))
Copy link
Collaborator

@preames preames Sep 13, 2024

Choose a reason for hiding this comment

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

I don't think there's any requirement that a power of two VL is a full register? Consider SEW=64, VL=2 on RISCV. With zvl256b this is at most half of a register. At zvl128b, it might be a full register (or might not.)

Copy link
Member Author

@alexey-bataev alexey-bataev Sep 16, 2024 •
edited
Loading

Choose a reason for hiding this comment

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

Cust currently SLP supports power-of-2 elements, no matter if it uses full registers or not. This check just checks for what we currently have in the compiler by default, it is not new requirement. I'll rename the function.

return true;
const unsigned NumParts = TTI.getNumberOfParts(getWidenedType(Ty, Sz));
return NumParts > 0 && NumParts < Sz && has_single_bit(Sz / NumParts) &&
Sz % NumParts == 0;
}

namespace slpvectorizer {

/// Bottom Up SLP Vectorizer.
Expand Down Expand Up @@ -2466,7 +2496,9 @@ class BoUpSLP {
}
// TODO: Check if we can remove a check for non-power-2 number of
// scalars after full support of non-power-2 vectorization.
return UniqueValues.size() != 2 && has_single_bit(UniqueValues.size());
return UniqueValues.size() != 2 &&
hasFullVectorsOnly(*R.TTI, (*UniqueValues.begin())->getType(),
Copy link
Collaborator

@preames preames Sep 13, 2024

Choose a reason for hiding this comment

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

Instead of white listing only the new class here, can we remove the restriction in a separate patch? This should be fully VL independent right?

Copy link
Member Author

@alexey-bataev alexey-bataev Sep 16, 2024

Choose a reason for hiding this comment

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

Currently no, it will regress some code. I 'll restore an original check here and will update it later

UniqueValues.size());
};

// If the initial strategy fails for any of the operand indexes, then we
Expand Down Expand Up @@ -3275,8 +3307,9 @@ class BoUpSLP {
SmallVectorImpl<Value *> *AltScalars = nullptr) const;

/// Return true if this is a non-power-of-2 node.
bool isNonPowOf2Vec() const {
Copy link
Collaborator

@preames preames Sep 13, 2024

Choose a reason for hiding this comment

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

Request: Please leave the original method, and add a new one called isAllowedNonPowOfTwo.

bool IsNonPowerOf2 = !has_single_bit(Scalars.size());
bool hasNonWholeRegisterElems(const TargetTransformInfo &TTI) const {
bool IsNonPowerOf2 = !hasFullVectorsOnly(
TTI, getValueType(Scalars.front()), Scalars.size());
assert((!IsNonPowerOf2 || ReuseShuffleIndices.empty()) &&
"Reshuffling not supported with non-power-of-2 vectors yet.");
return IsNonPowerOf2;
Expand Down Expand Up @@ -3390,7 +3423,7 @@ class BoUpSLP {
Last->State = EntryState;
// FIXME: Remove once support for ReuseShuffleIndices has been implemented
// for non-power-of-two vectors.
assert((has_single_bit(VL.size()) || ReuseShuffleIndices.empty()) &&
assert((((VL.size() & 1U) == 0) || ReuseShuffleIndices.empty()) &&
"Reshuffling scalars not yet supported for nodes with padding");
Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(),
ReuseShuffleIndices.end());
Expand Down Expand Up @@ -4781,12 +4814,13 @@ BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
});
});
const unsigned AbsoluteDiff = std::abs(*Diff);
if (IsPossibleStrided && (IsAnyPointerUsedOutGraph ||
((Sz > MinProfitableStridedLoads ||
(AbsoluteDiff <= MaxProfitableLoadStride * Sz &&
has_single_bit(AbsoluteDiff))) &&
AbsoluteDiff > Sz) ||
*Diff == -(static_cast<int>(Sz) - 1))) {
if (IsPossibleStrided &&
(IsAnyPointerUsedOutGraph ||
((Sz > MinProfitableStridedLoads ||
(AbsoluteDiff <= MaxProfitableLoadStride * Sz &&
hasFullVectorsOnly(*TTI, ScalarTy, AbsoluteDiff))) &&
Copy link
Collaborator

@preames preames Sep 6, 2024

Choose a reason for hiding this comment

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

I think you're confusing two things here. The has_single_bit check was on the size (and thus stride), not on the element type. I think we can probably just relax this check to allow any stride >= Sz.

Copy link
Member Author

@alexey-bataev alexey-bataev Sep 6, 2024

Choose a reason for hiding this comment

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

Mistakenly replaced, when searched for has_single_bit function use

AbsoluteDiff > Sz) ||
*Diff == -(static_cast<int>(Sz) - 1))) {
int Stride = *Diff / static_cast<int>(Sz - 1);
if (*Diff == Stride * static_cast<int>(Sz - 1)) {
Align Alignment =
Expand Down Expand Up @@ -5200,7 +5234,7 @@ BoUpSLP::getReorderingData(const TreeEntry &TE, bool TopToBottom) {
// node.
if (!TE.ReuseShuffleIndices.empty()) {
// FIXME: Support ReuseShuffleIndices for non-power-of-two vectors.
assert(!TE.isNonPowOf2Vec() &&
assert(!TE.hasNonWholeRegisterElems(*TTI) &&
Copy link
Collaborator

@preames preames Sep 13, 2024

Choose a reason for hiding this comment

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

Same as above.

Copy link
Member Author

@alexey-bataev alexey-bataev Sep 16, 2024

Choose a reason for hiding this comment

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

Again, the test checks this

"Reshuffling scalars not yet supported for nodes with padding");

if (isSplat(TE.Scalars))
Expand Down Expand Up @@ -5431,7 +5465,7 @@ BoUpSLP::getReorderingData(const TreeEntry &TE, bool TopToBottom) {

// FIXME: Remove the non-power-of-two check once findReusedOrderedScalars
// has been auditted for correctness with non-power-of-two vectors.
if (!TE.isNonPowOf2Vec())
if (!TE.hasNonWholeRegisterElems(*TTI))
Copy link
Collaborator

@preames preames Sep 13, 2024

Choose a reason for hiding this comment

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

If you've audited this for arbitrary VL, please just remove the check. (In it's own commit.)

Copy link
Member Author

@alexey-bataev alexey-bataev Sep 16, 2024

Choose a reason for hiding this comment

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

Reordering requires some extra testing, reverting it for now

if (std::optional<OrdersType> CurrentOrder = findReusedOrderedScalars(TE))
return CurrentOrder;
}
Expand Down Expand Up @@ -5585,7 +5619,7 @@ void BoUpSLP::reorderTopToBottom() {

// Reorder the graph nodes according to their vectorization factor.
for (unsigned VF = VectorizableTree.front()->getVectorFactor(); VF > 1;
VF = bit_ceil(VF) / 2) {
VF = (VF & ~1U) - 2) {
Copy link
Collaborator

@preames preames Sep 6, 2024

Choose a reason for hiding this comment

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

What is this trying to do? & -1 should just be VF shouldn't it?

Copy link
Member Author

@alexey-bataev alexey-bataev Sep 6, 2024

Choose a reason for hiding this comment

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

If VF is odd (for 3, 7, etc. number of elements), want to try next even. I reworked it to VF += 2 - (VF & 1) to jump over even VFs

auto It = VFToOrderedEntries.find(VF);
if (It == VFToOrderedEntries.end())
continue;
Expand Down Expand Up @@ -6574,7 +6608,7 @@ BoUpSLP::TreeEntry::EntryState BoUpSLP::getScalarsVectorizationState(
case Instruction::ExtractElement: {
bool Reuse = canReuseExtract(VL, VL0, CurrentOrder);
// FIXME: Vectorizing is not supported yet for non-power-of-2 ops.
if (!has_single_bit(VL.size()))
if (!hasFullVectorsOnly(*TTI, VL0->getType(), VL.size()))
Copy link
Collaborator

@preames preames Sep 13, 2024

Choose a reason for hiding this comment

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

Again, have you audited the exposed code?

Copy link
Member Author

@alexey-bataev alexey-bataev Sep 16, 2024

Choose a reason for hiding this comment

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

Reverting it for now

return TreeEntry::NeedToGather;
if (Reuse || !CurrentOrder.empty())
return TreeEntry::Vectorize;
Expand Down Expand Up @@ -6980,33 +7014,36 @@ void BoUpSLP::buildTree_rec(ArrayRef<Value *> VL, unsigned Depth,
UniqueValues.emplace_back(V);
}
size_t NumUniqueScalarValues = UniqueValues.size();
if (NumUniqueScalarValues == VL.size()) {
bool IsFullVectors = hasFullVectorsOnly(
*TTI, UniqueValues.front()->getType(), NumUniqueScalarValues);
if (NumUniqueScalarValues == VL.size() &&
(VectorizeNonPowerOf2 || IsFullVectors)) {
ReuseShuffleIndices.clear();
} else {
// FIXME: Reshuffing scalars is not supported yet for non-power-of-2 ops.
if ((UserTreeIdx.UserTE && UserTreeIdx.UserTE->isNonPowOf2Vec()) ||
!llvm::has_single_bit(VL.size())) {
if ((UserTreeIdx.UserTE &&
UserTreeIdx.UserTE->hasNonWholeRegisterElems(*TTI)) ||
!has_single_bit(VL.size())) {
LLVM_DEBUG(dbgs() << "SLP: Reshuffling scalars not yet supported "
"for nodes with padding.\n");
newTreeEntry(VL, std::nullopt /*not vectorized*/, S, UserTreeIdx);
return false;
}
LLVM_DEBUG(dbgs() << "SLP: Shuffle for reused scalars.\n");
if (NumUniqueScalarValues <= 1 ||
(UniquePositions.size() == 1 && all_of(UniqueValues,
[](Value *V) {
return isa<UndefValue>(V) ||
!isConstant(V);
})) ||
!llvm::has_single_bit<uint32_t>(NumUniqueScalarValues)) {
if (NumUniqueScalarValues <= 1 || !IsFullVectors ||
(UniquePositions.size() == 1 && all_of(UniqueValues, [](Value *V) {
return isa<UndefValue>(V) || !isConstant(V);
}))) {
if (DoNotFail && UniquePositions.size() > 1 &&
NumUniqueScalarValues > 1 && S.MainOp->isSafeToRemove() &&
all_of(UniqueValues, [=](Value *V) {
return isa<ExtractElementInst>(V) ||
areAllUsersVectorized(cast<Instruction>(V),
UserIgnoreList);
})) {
unsigned PWSz = PowerOf2Ceil(UniqueValues.size());
// Find the number of elements, which forms full vectors.
unsigned PWSz = getFullVectorNumberOfElements(
*TTI, UniqueValues.front()->getType(), UniqueValues.size());
if (PWSz == VL.size()) {
ReuseShuffleIndices.clear();
} else {
Expand Down Expand Up @@ -9101,9 +9138,6 @@ class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
return nullptr;
Value *VecBase = nullptr;
ArrayRef<Value *> VL = E->Scalars;
// If the resulting type is scalarized, do not adjust the cost.
if (NumParts == VL.size())
return nullptr;
// Check if it can be considered reused if same extractelements were
// vectorized already.
bool PrevNodeFound = any_of(
Expand Down Expand Up @@ -9756,7 +9790,7 @@ BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals,
InsertMask[Idx] = I + 1;
}
unsigned VecScalarsSz = PowerOf2Ceil(NumElts);
if (NumOfParts > 0)
if (NumOfParts > 0 && NumOfParts < NumElts)
VecScalarsSz = PowerOf2Ceil((NumElts + NumOfParts - 1) / NumOfParts);
unsigned VecSz = (1 + OffsetEnd / VecScalarsSz - OffsetBeg / VecScalarsSz) *
VecScalarsSz;
Expand Down Expand Up @@ -10985,7 +11019,9 @@ InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) {
// Keep original scalar if number of externally used instructions in
// the same entry is not power of 2. It may help to do some extra
// vectorization for now.
KeepScalar = ScalarUsesCount <= 1 || !has_single_bit(ScalarUsesCount);
KeepScalar =
ScalarUsesCount <= 1 ||
!hasFullVectorsOnly(*TTI, EU.Scalar->getType(), ScalarUsesCount);
}
if (KeepScalar) {
ExternalUsesAsOriginalScalar.insert(EU.Scalar);
Expand Down Expand Up @@ -11682,13 +11718,14 @@ BoUpSLP::isGatherShuffledEntry(
if (TE == VectorizableTree.front().get())
return {};
// FIXME: Gathering for non-power-of-2 nodes not implemented yet.
if (TE->isNonPowOf2Vec())
if (TE->hasNonWholeRegisterElems(*TTI))
return {};
Mask.assign(VL.size(), PoisonMaskElem);
assert(TE->UserTreeIndices.size() == 1 &&
"Expected only single user of the gather node.");
assert(VL.size() % NumParts == 0 &&
"Number of scalars must be divisible by NumParts.");
// Number of scalars must be divisible by NumParts.
if (VL.size() % NumParts != 0)
Copy link
Collaborator

@preames preames Sep 6, 2024

Choose a reason for hiding this comment

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

I'm reading over the existing NumParts code, and not sure I'm grasping what it's doing. Is the idea to cost manual splitting of the vector to a smaller legal vector type? If so, why is this not something that TTI should be doing internally?

Copy link
Member Author

@alexey-bataev alexey-bataev Sep 6, 2024

Choose a reason for hiding this comment

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

No, not vector type splitting, but finding subvectors, which can be shuffled

return {};
unsigned SliceSize = getPartNumElems(VL.size(), NumParts);
SmallVector<std::optional<TTI::ShuffleKind>> Res;
for (unsigned Part : seq<unsigned>(NumParts)) {
Expand Down Expand Up @@ -16999,7 +17036,7 @@ bool SLPVectorizerPass::tryToVectorizeList(ArrayRef<Value *> VL, BoUpSLP &R,
for (unsigned I = NextInst; I < MaxInst; ++I) {
unsigned ActualVF = std::min(MaxInst - I, VF);

if (!has_single_bit(ActualVF))
if (!hasFullVectorsOnly(*TTI, ScalarTy, ActualVF))
continue;

if (MaxVFOnly && ActualVF < MaxVF)
Expand Down
28 changes: 18 additions & 10 deletions llvm/test/Transforms/SLPVectorizer/reduction-whole-regs-loads.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,21 +1,29 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=riscv64-unknown-linux -mattr=+v -slp-threshold=-100 | FileCheck %s
; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=riscv64-unknown-linux -mattr=+v -slp-threshold=-100 | FileCheck %s --check-prefix=RISCV
; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=x86_64-unknown-linux -slp-threshold=-100 | FileCheck %s
; RUN: opt < %s -passes=slp-vectorizer -S -mtriple=aarch64-unknown-linux -slp-threshold=-100 | FileCheck %s
; REQUIRES: aarch64-registered-target, x86-registered-target, riscv-registered-target

define i64 @test(ptr %p) {
; RISCV-LABEL: @test(
; RISCV-NEXT: entry:
; RISCV-NEXT: [[ARRAYIDX_4:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i64 4
; RISCV-NEXT: [[TMP0:%.*]] = load <4 x i64>, ptr [[P]], align 4
; RISCV-NEXT: [[TMP1:%.*]] = load <2 x i64>, ptr [[ARRAYIDX_4]], align 4
; RISCV-NEXT: [[TMP2:%.*]] = shufflevector <4 x i64> [[TMP0]], <4 x i64> poison, <8 x i32> <i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 0, i32 0>
; RISCV-NEXT: [[TMP3:%.*]] = call <8 x i64> @llvm.vector.insert.v8i64.v4i64(<8 x i64> [[TMP2]], <4 x i64> [[TMP0]], i64 0)
; RISCV-NEXT: [[TMP4:%.*]] = call <8 x i64> @llvm.vector.insert.v8i64.v2i64(<8 x i64> [[TMP3]], <2 x i64> [[TMP1]], i64 4)
; RISCV-NEXT: [[TMP5:%.*]] = mul <8 x i64> [[TMP4]], <i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42>
; RISCV-NEXT: [[TMP6:%.*]] = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> [[TMP5]])
; RISCV-NEXT: ret i64 [[TMP6]]
;
; CHECK-LABEL: @test(
; CHECK-NEXT: entry:
; CHECK-NEXT: [[ARRAYIDX_4:%.*]] = getelementptr inbounds i64, ptr [[P:%.*]], i64 4
; CHECK-NEXT: [[TMP0:%.*]] = load <4 x i64>, ptr [[P]], align 4
; CHECK-NEXT: [[TMP1:%.*]] = load <2 x i64>, ptr [[ARRAYIDX_4]], align 4
; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <4 x i64> [[TMP0]], <4 x i64> poison, <8 x i32> <i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 0, i32 0>
; CHECK-NEXT: [[TMP3:%.*]] = call <8 x i64> @llvm.vector.insert.v8i64.v4i64(<8 x i64> [[TMP2]], <4 x i64> [[TMP0]], i64 0)
; CHECK-NEXT: [[TMP4:%.*]] = call <8 x i64> @llvm.vector.insert.v8i64.v2i64(<8 x i64> [[TMP3]], <2 x i64> [[TMP1]], i64 4)
; CHECK-NEXT: [[TMP5:%.*]] = mul <8 x i64> [[TMP4]], <i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42>
; CHECK-NEXT: [[TMP6:%.*]] = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> [[TMP5]])
; CHECK-NEXT: ret i64 [[TMP6]]
; CHECK-NEXT: [[TMP0:%.*]] = load <6 x i64>, ptr [[P:%.*]], align 4
; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <6 x i64> [[TMP0]], <6 x i64> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 0, i32 0>
; CHECK-NEXT: [[TMP2:%.*]] = mul <8 x i64> [[TMP1]], <i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42, i64 42>
; CHECK-NEXT: [[TMP3:%.*]] = call i64 @llvm.vector.reduce.add.v8i64(<8 x i64> [[TMP2]])
; CHECK-NEXT: ret i64 [[TMP3]]
;
entry:
%arrayidx.1 = getelementptr inbounds i64, ptr %p, i64 1
Expand Down