Skip to content

[WIP][LV][VPlan] For reverse iterative continuous load and store access, use unit load/store #130032

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 1 commit into
base: main
Choose a base branch
from
Open

[WIP][LV][VPlan] For reverse iterative continuous load and store access, use unit load/store #130032

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
9 changes: 8 additions & 1 deletion llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7487,6 +7487,13 @@ static bool planContainsAdditionalSimplifications(VPlan &Plan,
// comparing against the legacy cost isn't desirable.
if (isa<VPPartialReductionRecipe>(&R))
return true;

// The VPlan-based cost model may calculate the cost of strided load/store
// which can't be modeled in the legacy cost model.
if (isa<VPWidenLoadEVLRecipe, VPWidenStoreEVLRecipe>(&R))
if (cast<VPWidenMemoryRecipe>(&R)->isReverse())
return true;

if (Instruction *UI = GetInstructionForCost(&R))
SeenInstrs.insert(UI);
}
Expand Down Expand Up @@ -8344,7 +8351,7 @@ VPRecipeBuilder::tryToWidenMemory(Instruction *I, ArrayRef<VPValue *> Operands,
auto *GEP = dyn_cast<GetElementPtrInst>(
Ptr->getUnderlyingValue()->stripPointerCasts());
VPSingleDefRecipe *VectorPtr;
if (Reverse) {
if (Reverse && !CM.foldTailWithEVL()) {
// When folding the tail, we may compute an address that we don't in the
// original scalar loop and it may not be inbounds. Drop Inbounds in that
// case.
Expand Down
86 changes: 44 additions & 42 deletions llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2640,17 +2640,6 @@ void VPWidenLoadRecipe::print(raw_ostream &O, const Twine &Indent,
}
#endif

/// Use all-true mask for reverse rather than actual mask, as it avoids a
/// dependence w/o affecting the result.
static Instruction *createReverseEVL(IRBuilderBase &Builder, Value *Operand,
Value *EVL, const Twine &Name) {
VectorType *ValTy = cast<VectorType>(Operand->getType());
Value *AllTrueMask =
Builder.CreateVectorSplat(ValTy->getElementCount(), Builder.getTrue());
return Builder.CreateIntrinsic(ValTy, Intrinsic::experimental_vp_reverse,
{Operand, AllTrueMask, EVL}, nullptr, Name);
}

void VPWidenLoadEVLRecipe::execute(VPTransformState &State) {
auto *LI = cast<LoadInst>(&Ingredient);

Expand All @@ -2665,19 +2654,26 @@ void VPWidenLoadEVLRecipe::execute(VPTransformState &State) {
Value *EVL = State.get(getEVL(), VPLane(0));
Value *Addr = State.get(getAddr(), !CreateGather);
Value *Mask = nullptr;
if (VPValue *VPMask = getMask()) {
if (VPValue *VPMask = getMask())
Mask = State.get(VPMask);
if (isReverse())
Mask = createReverseEVL(Builder, Mask, EVL, "vp.reverse.mask");
} else {
else
Mask = Builder.CreateVectorSplat(State.VF, Builder.getTrue());
}

if (CreateGather) {
NewLI =
Builder.CreateIntrinsic(DataTy, Intrinsic::vp_gather, {Addr, Mask, EVL},
nullptr, "wide.masked.gather");
} else {
if (isReverse()) {
auto *EltTy = DataTy->getElementType();
// if (EltTy->getScalarSizeInBits() !=
// EVL->getType()->getScalarSizeInBits())
// EVL = ConstantInt::getSigned(EVL->getType(),
// static_cast<int64_t>(EltTy->getScalarSizeInBits()) / 8);
auto *GEP = dyn_cast<GetElementPtrInst>(Addr->stripPointerCasts());
Value *Offset = Builder.CreateSub(State.Builder.getInt32(1), EVL);
Addr = Builder.CreateGEP(EltTy, Addr, Offset, "", GEP->isInBounds());
}
VectorBuilder VBuilder(Builder);
VBuilder.setEVL(EVL).setMask(Mask);
NewLI = cast<CallInst>(VBuilder.createVectorInstruction(
Expand All @@ -2686,10 +2682,7 @@ void VPWidenLoadEVLRecipe::execute(VPTransformState &State) {
NewLI->addParamAttr(
0, Attribute::getWithAlignment(NewLI->getContext(), Alignment));
State.addMetadata(NewLI, LI);
Instruction *Res = NewLI;
if (isReverse())
Res = createReverseEVL(Builder, Res, EVL, "vp.reverse");
State.set(this, Res);
State.set(this, NewLI);
}

InstructionCost VPWidenLoadEVLRecipe::computeCost(ElementCount VF,
Expand All @@ -2707,14 +2700,13 @@ InstructionCost VPWidenLoadEVLRecipe::computeCost(ElementCount VF,
getLoadStoreAlignment(const_cast<Instruction *>(&Ingredient));
unsigned AS =
getLoadStoreAddressSpace(const_cast<Instruction *>(&Ingredient));
InstructionCost Cost = Ctx.TTI.getMaskedMemoryOpCost(
Ingredient.getOpcode(), Ty, Alignment, AS, Ctx.CostKind);
if (!Reverse)
return Cost;
// if (!Reverse)
return Ctx.TTI.getMaskedMemoryOpCost(Ingredient.getOpcode(), Ty, Alignment,
AS, Ctx.CostKind);

return Cost + Ctx.TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
cast<VectorType>(Ty), {}, Ctx.CostKind,
0);
// return Ctx.TTI.getStridedMemoryOpCost(Ingredient.getOpcode(), Ty,
// getAddr()->getUnderlyingValue(), false,
// Alignment, Ctx.CostKind);
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
Expand Down Expand Up @@ -2775,7 +2767,8 @@ void VPWidenStoreRecipe::print(raw_ostream &O, const Twine &Indent,

void VPWidenStoreEVLRecipe::execute(VPTransformState &State) {
auto *SI = cast<StoreInst>(&Ingredient);

Type *ScalarDataTy = getLoadStoreType(&Ingredient);
auto *DataTy = VectorType::get(ScalarDataTy, State.VF);
VPValue *StoredValue = getStoredValue();
bool CreateScatter = !isConsecutive();
const Align Alignment = getLoadStoreAlignment(&Ingredient);
Expand All @@ -2786,22 +2779,32 @@ void VPWidenStoreEVLRecipe::execute(VPTransformState &State) {
CallInst *NewSI = nullptr;
Value *StoredVal = State.get(StoredValue);
Value *EVL = State.get(getEVL(), VPLane(0));
if (isReverse())
StoredVal = createReverseEVL(Builder, StoredVal, EVL, "vp.reverse");
Value *Mask = nullptr;
if (VPValue *VPMask = getMask()) {
if (VPValue *VPMask = getMask())
Mask = State.get(VPMask);
if (isReverse())
Mask = createReverseEVL(Builder, Mask, EVL, "vp.reverse.mask");
} else {
else
Mask = Builder.CreateVectorSplat(State.VF, Builder.getTrue());
}

Value *Addr = State.get(getAddr(), !CreateScatter);
if (CreateScatter) {
NewSI = Builder.CreateIntrinsic(Type::getVoidTy(EVL->getContext()),
Intrinsic::vp_scatter,
{StoredVal, Addr, Mask, EVL});
} else {
if (isReverse()) {
auto *EltTy = DataTy->getElementType();
// FIXME: we may need not deal with the size, the InstCombine will deal
// with the Offset Type if (EltTy->getScalarSizeInBits() !=
// EVL->getType()->getScalarSizeInBits())
// EVL = ConstantInt::getSigned(EVL->getType(),
// static_cast<int64_t>(EltTy->getScalarSizeInBits()) / 8);
auto *GEP = dyn_cast<GetElementPtrInst>(Addr->stripPointerCasts());
// Value *Offset =
// Builder.CreateSub(State.Builder.getIntN(EVL->getType()->getScalarSizeInBits(),
// 1), EVL);
Value *Offset = Builder.CreateSub(State.Builder.getInt32(1), EVL);
Addr = Builder.CreateGEP(EltTy, Addr, Offset, "", GEP->isInBounds());
}
VectorBuilder VBuilder(Builder);
VBuilder.setEVL(EVL).setMask(Mask);
NewSI = cast<CallInst>(VBuilder.createVectorInstruction(
Expand All @@ -2828,14 +2831,13 @@ InstructionCost VPWidenStoreEVLRecipe::computeCost(ElementCount VF,
getLoadStoreAlignment(const_cast<Instruction *>(&Ingredient));
unsigned AS =
getLoadStoreAddressSpace(const_cast<Instruction *>(&Ingredient));
InstructionCost Cost = Ctx.TTI.getMaskedMemoryOpCost(
Ingredient.getOpcode(), Ty, Alignment, AS, Ctx.CostKind);
if (!Reverse)
return Cost;
// if (!Reverse)
return Ctx.TTI.getMaskedMemoryOpCost(Ingredient.getOpcode(), Ty, Alignment,
AS, Ctx.CostKind);

return Cost + Ctx.TTI.getShuffleCost(TargetTransformInfo::SK_Reverse,
cast<VectorType>(Ty), {}, Ctx.CostKind,
0);
// return Ctx.TTI.getStridedMemoryOpCost(Ingredient.getOpcode(), Ty,
// getAddr()->getUnderlyingValue(), false,
// Alignment, Ctx.CostKind);
}

#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
Expand Down
Loading