Skip to content

[SLPVectorizer] Widen constant strided loads. #162324

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

Open
wants to merge 6 commits into
base: main
Choose a base branch
from
Open

[SLPVectorizer] Widen constant strided loads. #162324

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
6ec2679
[SLPVectorizer] Widen constant strided loads.
Sep 23, 2025
7639394
updated failing tests.
Sep 24, 2025
e8895fb
rename VecSz to Sz to make diff look better.
Oct 2, 2025
eece1bf
update test.
Oct 2, 2025
4c7d605
added / improved comments,
Oct 8, 2025
0cefc4a
more comments, simplifications, more `const`
Oct 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
185 changes: 155 additions & 30 deletions llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2242,8 +2242,47 @@ class BoUpSLP {
/// may not be necessary.
bool isLoadCombineCandidate(ArrayRef<Value *> Stores) const;
bool isStridedLoad(ArrayRef<Value *> PointerOps, Type *ScalarTy,
Align Alignment, const int64_t Diff, Value *Ptr0,
Value *PtrN, StridedPtrInfo &SPtrInfo) const;
Align Alignment, int64_t Diff, size_t Sz) const;

/// Return true if an array of scalar loads can be replaced with a strided
/// load (with constant stride).
///
/// It is possible that the load gets "widened". Suppose that originally each load loads `k` bytes and `PointerOps` can be arranged as follows (`%s` is constant):
/// %b + 0 * %s + 0
/// %b + 0 * %s + 1
/// %b + 0 * %s + 2
/// ...
/// %b + 0 * %s + (w - 1)
///
/// %b + 1 * %s + 0
/// %b + 1 * %s + 1
/// %b + 1 * %s + 2
/// ...
/// %b + 1 * %s + (w - 1)
/// ...
///
/// %b + (n - 1) * %s + 0
/// %b + (n - 1) * %s + 1
/// %b + (n - 1) * %s + 2
/// ...
/// %b + (n - 1) * %s + (w - 1)
///
/// In this case we will generate a strided load of type `<n x (k * w)>`.
///
/// \param PointerOps list of pointer arguments of loads.
/// \param ElemTy original scalar type of loads.
/// \param Alignment alignment of the first load.
/// \param SortedIndices is the order of PointerOps as returned by `sortPtrAccesses`
/// \param Diff Pointer difference between the lowest and the highes pointer in `PointerOps` as returned by `getPointersDiff`.
/// \param Ptr0 first pointer in `PointersOps`.
/// \param PtrN last pointer in `PointersOps`.
/// \param SPtrInfo If the function return `true`, it also sets all the fields
/// of `SPtrInfo` necessary to generate the strided load later.
bool analyzeConstantStrideCandidate(const ArrayRef<Value *> PointerOps,
Type *ElemTy, Align Alignment,
const SmallVectorImpl<unsigned> &SortedIndices,
const int64_t Diff, Value *Ptr0, Value *PtrN,
StridedPtrInfo &SPtrInfo) const;

/// Return true if an array of scalar loads can be replaced with a strided
/// load (with run-time stride).
Expand Down Expand Up @@ -6844,12 +6883,7 @@ isMaskedLoadCompress(ArrayRef<Value *> VL, ArrayRef<Value *> PointerOps,
/// current graph (for masked gathers extra extractelement instructions
/// might be required).
bool BoUpSLP::isStridedLoad(ArrayRef<Value *> PointerOps, Type *ScalarTy,
Align Alignment, const int64_t Diff, Value *Ptr0,
Value *PtrN, StridedPtrInfo &SPtrInfo) const {
const size_t Sz = PointerOps.size();
if (Diff % (Sz - 1) != 0)
return false;

Align Alignment, int64_t Diff, size_t Sz) const {
// Try to generate strided load node.
auto IsAnyPointerUsedOutGraph = any_of(PointerOps, [&](Value *V) {
return isa<Instruction>(V) && any_of(V->users(), [&](User *U) {
Expand All @@ -6870,29 +6904,110 @@ bool BoUpSLP::isStridedLoad(ArrayRef<Value *> PointerOps, Type *ScalarTy,
return false;
if (!TTI->isLegalStridedLoadStore(VecTy, Alignment))
return false;
}
return true;
}

// Iterate through all pointers and check if all distances are
// unique multiple of Dist.
SmallSet<int64_t, 4> Dists;
for (Value *Ptr : PointerOps) {
int64_t Dist = 0;
if (Ptr == PtrN)
Dist = Diff;
else if (Ptr != Ptr0)
Dist = *getPointersDiff(ScalarTy, Ptr0, ScalarTy, Ptr, *DL, *SE);
// If the strides are not the same or repeated, we can't
// vectorize.
if (((Dist / Stride) * Stride) != Dist || !Dists.insert(Dist).second)
break;
bool BoUpSLP::analyzeConstantStrideCandidate(
const ArrayRef<Value *> PointerOps, Type *ElemTy, Align CommonAlignment,
const SmallVectorImpl<unsigned> &SortedIndices, const int64_t Diff, Value *Ptr0,
Value *PtrN, StridedPtrInfo &SPtrInfo) const {
const unsigned Sz = PointerOps.size();
SmallVector<int64_t> SortedOffsetsFromBase(Sz);
// Go through `PointerOps` in sorted order and record offsets from `Ptr0`.
for (unsigned I : seq<unsigned>(Sz)) {
Value *Ptr =
SortedIndices.empty() ? PointerOps[I] : PointerOps[SortedIndices[I]];
SortedOffsetsFromBase[I] =
*getPointersDiff(ElemTy, Ptr0, ElemTy, Ptr, *DL, *SE);
}
assert(SortedOffsetsFromBase.size() > 1 &&
"Trying to generate strided load for less than 2 loads");
// The code below checks that `SortedOffsetsFromBase` looks as follows:
// ```
// [
// (e_{0, 0}, e_{0, 1}, ..., e_{0, GroupSize - 1}), // first group
// (e_{1, 0}, e_{1, 1}, ..., e_{1, GroupSize - 1}), // secon group
// ...
// (e_{NumGroups - 1, 0}, e_{NumGroups - 1, 1}, ..., e_{NumGroups - 1,
// GroupSize - 1}), // last group
// ]
// ```
// The distance between consecutive elements within each group should all be
// the same `StrideWithinGroup`. The distance between the first elements of
// consecutive groups should all be the same `StrideBetweenGroups`.

int64_t StrideWithinGroup =
SortedOffsetsFromBase[1] - SortedOffsetsFromBase[0];
// Determine size of the first group. Later we will check that all other
// groups have the same size.
unsigned GroupSize = 1;
for (; GroupSize != SortedOffsetsFromBase.size(); ++GroupSize) {
if (SortedOffsetsFromBase[GroupSize] -
SortedOffsetsFromBase[GroupSize - 1] !=
StrideWithinGroup)
break;
}
unsigned VecSz = Sz;
Type *ScalarTy = ElemTy;
int64_t StrideIntVal = StrideWithinGroup;
FixedVectorType *StridedLoadTy = getWidenedType(ScalarTy, VecSz);

// Quick detour: at this point we can say what the type of strided load would
// be if all the checks pass. Check if this type is legal for the target.
if (Sz != GroupSize) {
if (Sz % GroupSize != 0)
return false;
VecSz = Sz / GroupSize;

if (StrideWithinGroup != 1)
return false;
unsigned VecSz = Sz / GroupSize;
ScalarTy = Type::getIntNTy(SE->getContext(),
DL->getTypeSizeInBits(ElemTy).getFixedValue() *
GroupSize);
StridedLoadTy = getWidenedType(ScalarTy, VecSz);
if (!TTI->isTypeLegal(StridedLoadTy) ||
!TTI->isLegalStridedLoadStore(StridedLoadTy, CommonAlignment))
return false;

// Continue with checking the "shape" of `SortedOffsetsFromBase`.
// Check that the strides between groups are all the same.
unsigned CurrentGroupStartIdx = GroupSize;
int64_t StrideBetweenGroups =
SortedOffsetsFromBase[GroupSize] - SortedOffsetsFromBase[0];
StrideIntVal = StrideBetweenGroups;
for (; CurrentGroupStartIdx < Sz; CurrentGroupStartIdx += GroupSize) {
if (SortedOffsetsFromBase[CurrentGroupStartIdx] -
SortedOffsetsFromBase[CurrentGroupStartIdx - GroupSize] !=
StrideBetweenGroups)
return false;
}
if (Dists.size() == Sz) {
Type *StrideTy = DL->getIndexType(Ptr0->getType());
SPtrInfo.StrideVal = ConstantInt::get(StrideTy, Stride);
SPtrInfo.Ty = getWidenedType(ScalarTy, Sz);
return true;

auto CheckGroup = [&](const unsigned StartIdx, const unsigned GroupSize0,
const int64_t StrideWithinGroup) -> bool {
unsigned GroupEndIdx = StartIdx + 1;
for (; GroupEndIdx != Sz; ++GroupEndIdx) {
if (SortedOffsetsFromBase[GroupEndIdx] -
SortedOffsetsFromBase[GroupEndIdx - 1] !=
StrideWithinGroup)
break;
}
return GroupEndIdx - StartIdx == GroupSize0;
};
for (unsigned I = 0; I < Sz; I += GroupSize) {
if (!CheckGroup(I, GroupSize, StrideWithinGroup))
return false;
}
}
return false;

if (!isStridedLoad(PointerOps, ScalarTy, CommonAlignment, Diff, VecSz))
return false;

Type *StrideTy = DL->getIndexType(Ptr0->getType());
SPtrInfo.StrideVal = ConstantInt::get(StrideTy, StrideIntVal);
SPtrInfo.Ty = StridedLoadTy;
return true;
}

bool BoUpSLP::analyzeRtStrideCandidate(ArrayRef<Value *> PointerOps,
Expand Down Expand Up @@ -6990,8 +7105,8 @@ BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
Align Alignment =
cast<LoadInst>(Order.empty() ? VL.front() : VL[Order.front()])
->getAlign();
if (isStridedLoad(PointerOps, ScalarTy, Alignment, *Diff, Ptr0, PtrN,
SPtrInfo))
if (analyzeConstantStrideCandidate(PointerOps, ScalarTy, Alignment, Order,
*Diff, Ptr0, PtrN, SPtrInfo))
return LoadsState::StridedVectorize;
}
if (!TTI->isLegalMaskedGather(VecTy, CommonAlignment) ||
Expand Down Expand Up @@ -14902,11 +15017,19 @@ BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals,
}
break;
case TreeEntry::StridedVectorize: {
const StridedPtrInfo &SPtrInfo = TreeEntryToStridedPtrInfoMap.at(E);
FixedVectorType *StridedLoadTy = SPtrInfo.Ty;
assert(StridedLoadTy && "Missing StridedPoinerInfo for tree entry.");
Align CommonAlignment =
computeCommonAlignment<LoadInst>(UniqueValues.getArrayRef());
VecLdCost = TTI->getStridedMemoryOpCost(
Instruction::Load, VecTy, LI0->getPointerOperand(),
Instruction::Load, StridedLoadTy, LI0->getPointerOperand(),
/*VariableMask=*/false, CommonAlignment, CostKind);
if (StridedLoadTy != VecTy)
VecLdCost +=
TTI->getCastInstrCost(Instruction::BitCast, VecTy, StridedLoadTy,
getCastContextHint(*E), CostKind);

break;
}
case TreeEntry::CompressVectorize: {
Expand Down Expand Up @@ -19670,6 +19793,8 @@ Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
? NewLI
: ::propagateMetadata(NewLI, E->Scalars);

if (StridedLoadTy != VecTy)
V = Builder.CreateBitOrPointerCast(V, VecTy);
V = FinalShuffle(V, E);
E->VectorizedValue = V;
++NumVectorInstructions;
Expand Down
20 changes: 4 additions & 16 deletions llvm/test/Transforms/SLPVectorizer/RISCV/basic-strided-loads.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5

; RUN: opt -mtriple=riscv64 -mattr=+m,+v -passes=slp-vectorizer -S < %s | FileCheck %s
; RUN: opt -mtriple=riscv64 -mattr=+m,+v,+unaligned-vector-mem -passes=slp-vectorizer -S < %s | FileCheck %s

define void @const_stride_1_no_reordering(ptr %pl, ptr %ps) {
; CHECK-LABEL: define void @const_stride_1_no_reordering(
Expand Down Expand Up @@ -621,22 +621,10 @@ define void @constant_stride_widen_no_reordering(ptr %pl, i64 %stride, ptr %ps)
; CHECK-LABEL: define void @constant_stride_widen_no_reordering(
; CHECK-SAME: ptr [[PL:%.*]], i64 [[STRIDE:%.*]], ptr [[PS:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: [[GEP_L0:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 0
; CHECK-NEXT: [[GEP_L4:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 100
; CHECK-NEXT: [[GEP_L8:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 200
; CHECK-NEXT: [[GEP_L12:%.*]] = getelementptr inbounds i8, ptr [[PL]], i64 300
; CHECK-NEXT: [[GEP_S0:%.*]] = getelementptr inbounds i8, ptr [[PS]], i64 0
; CHECK-NEXT: [[TMP1:%.*]] = load <4 x i8>, ptr [[GEP_L0]], align 1
; CHECK-NEXT: [[TMP2:%.*]] = load <4 x i8>, ptr [[GEP_L4]], align 1
; CHECK-NEXT: [[TMP3:%.*]] = load <4 x i8>, ptr [[GEP_L8]], align 1
; CHECK-NEXT: [[TMP4:%.*]] = load <4 x i8>, ptr [[GEP_L12]], align 1
; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <4 x i8> [[TMP1]], <4 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <4 x i8> [[TMP2]], <4 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <4 x i8> [[TMP1]], <4 x i8> [[TMP2]], <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: [[TMP11:%.*]] = shufflevector <4 x i8> [[TMP3]], <4 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP9:%.*]] = shufflevector <16 x i8> [[TMP7]], <16 x i8> [[TMP11]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 16, i32 17, i32 18, i32 19, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP10:%.*]] = shufflevector <4 x i8> [[TMP4]], <4 x i8> poison, <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP8:%.*]] = shufflevector <16 x i8> [[TMP9]], <16 x i8> [[TMP10]], <16 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 16, i32 17, i32 18, i32 19>
; CHECK-NEXT: store <16 x i8> [[TMP8]], ptr [[GEP_S0]], align 1
; CHECK-NEXT: [[TMP1:%.*]] = call <4 x i32> @llvm.experimental.vp.strided.load.v4i32.p0.i64(ptr align 1 [[GEP_L0]], i64 100, <4 x i1> splat (i1 true), i32 4)
; CHECK-NEXT: [[TMP11:%.*]] = bitcast <4 x i32> [[TMP1]] to <16 x i8>
; CHECK-NEXT: store <16 x i8> [[TMP11]], ptr [[GEP_S0]], align 1
; CHECK-NEXT: ret void
;
%gep_l0 = getelementptr inbounds i8, ptr %pl, i64 0
Expand Down
15 changes: 8 additions & 7 deletions llvm/test/Transforms/SLPVectorizer/RISCV/gather-insert-point-restore.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,18 +5,19 @@ define i16 @test(ptr %i) {
; CHECK-LABEL: define i16 @test(
; CHECK-SAME: ptr [[I:%.*]]) #[[ATTR0:[0-9]+]] {
; CHECK-NEXT: [[ENTRY:.*:]]
; CHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x ptr> poison, ptr [[I]], i32 0
; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <2 x ptr> [[TMP0]], <2 x ptr> poison, <2 x i32> zeroinitializer
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr i8, <2 x ptr> [[TMP1]], <2 x i64> <i64 132860, i64 137774>
; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x ptr> [[TMP2]], <2 x ptr> poison, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
; CHECK-NEXT: [[GEP_US154:%.*]] = getelementptr i8, ptr [[I]], i64 132860
; CHECK-NEXT: [[GEP_US154_2:%.*]] = getelementptr i8, ptr [[I]], i64 142688
; CHECK-NEXT: br label %[[FOR_COND5_US:.*]]
; CHECK: [[FOR_COND5_US]]:
; CHECK-NEXT: [[TMP0:%.*]] = call <2 x i16> @llvm.experimental.vp.strided.load.v2i16.p0.i64(ptr align 2 [[GEP_US154]], i64 4914, <2 x i1> splat (i1 true), i32 2)
; CHECK-NEXT: [[TMP4:%.*]] = call <4 x i16> @llvm.experimental.vp.strided.load.v4i16.p0.i64(ptr align 2 [[GEP_US154_2]], i64 4914, <4 x i1> splat (i1 true), i32 4)
; CHECK-NEXT: [[TMP5:%.*]] = call <4 x i16> @llvm.masked.gather.v4i16.v4p0(<4 x ptr> [[TMP3]], i32 2, <4 x i1> splat (i1 true), <4 x i16> poison)
; CHECK-NEXT: [[TMP2:%.*]] = call <2 x i16> @llvm.experimental.vp.strided.load.v2i16.p0.i64(ptr align 2 [[GEP_US154]], i64 4914, <2 x i1> splat (i1 true), i32 2)
; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <4 x i16> [[TMP4]], <4 x i16> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP10:%.*]] = shufflevector <4 x i16> [[TMP5]], <4 x i16> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <4 x i16> [[TMP4]], <4 x i16> [[TMP5]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>
; CHECK-NEXT: [[TMP10:%.*]] = shufflevector <2 x i16> [[TMP0]], <2 x i16> poison, <8 x i32> <i32 0, i32 1, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x i16> [[TMP0]], <2 x i16> poison, <4 x i32> <i32 0, i32 1, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <4 x i16> [[TMP4]], <4 x i16> [[TMP5]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <2 x i16> [[TMP2]], <2 x i16> poison, <8 x i32> <i32 0, i32 1, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison, i32 poison>
; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <8 x i16> [[TMP11]], <8 x i16> [[TMP12]], <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 8, i32 9>
; CHECK-NEXT: [[TMP8:%.*]] = call i16 @llvm.vector.reduce.umax.v8i16(<8 x i16> [[TMP7]])
; CHECK-NEXT: [[TMP9:%.*]] = call i16 @llvm.umax.i16(i16 [[TMP8]], i16 0)
; CHECK-NEXT: ret i16 [[TMP9]]
Expand Down
Loading
Loading