Skip to content

[VectorCombine] foldShuffleOfBinops - fold shuffle(binop(shuffle(x),shuffle(z)),binop(shuffle(y),shuffle(w)) -> binop(shuffle(x,z),shuffle(y,w)) #120984

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
Jan 3, 2025
Merged

[VectorCombine] foldShuffleOfBinops - fold shuffle(binop(shuffle(x),shuffle(z)),binop(shuffle(y),shuffle(w)) -> binop(shuffle(x,z),shuffle(y,w)) #120984

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
6375772
[VectorCombine] foldShuffleOfBinops - fold shuffle(binop(shuffle(x),s…
RKSimon Dec 23, 2024
1bb33e0
Move Type equivalence earlier (cheap checks first)
RKSimon Dec 28, 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
34 changes: 32 additions & 2 deletions llvm/lib/Transforms/Vectorize/VectorCombine.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1743,6 +1743,36 @@ bool VectorCombine::foldShuffleOfBinops(Instruction &I) {
TTI.getShuffleCost(TargetTransformInfo::SK_PermuteTwoSrc, BinResTy,
OldMask, CostKind, 0, nullptr, {LHS, RHS}, &I);

// Handle shuffle(binop(shuffle(x),y),binop(z,shuffle(w))) style patterns
// where one use shuffles have gotten split across the binop/cmp. These
// often allow a major reduction in total cost that wouldn't happen as
// individual folds.
auto MergeInner = [&](Value *&Op, int Offset, MutableArrayRef<int> Mask,
TTI::TargetCostKind CostKind) -> bool {
Value *InnerOp;
ArrayRef<int> InnerMask;
if (match(Op, m_OneUse(m_Shuffle(m_Value(InnerOp), m_Undef(),
m_Mask(InnerMask)))) &&
InnerOp->getType() == Op->getType() &&
all_of(InnerMask,
[NumSrcElts](int M) { return M < (int)NumSrcElts; })) {
for (int &M : Mask)
if (Offset <= M && M < (int)(Offset + NumSrcElts)) {
M = InnerMask[M - Offset];
M = 0 <= M ? M + Offset : M;
Copy link
Member

Choose a reason for hiding this comment

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

Can M be less than 0 here at all? If has a check that M>=Offset and Offset is either 0 or NumSrcElts

Copy link
Collaborator Author

Choose a reason for hiding this comment

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

Yes, we've replaced the original M with the InnerMask value to fold the shuffles together - and InnerMask could contain PoisonMaskElem

}
OldCost += TTI.getInstructionCost(cast<Instruction>(Op), CostKind);
Op = InnerOp;
return true;
}
return false;
};
bool ReducedInstCount = false;
ReducedInstCount |= MergeInner(X, 0, NewMask0, CostKind);
ReducedInstCount |= MergeInner(Y, 0, NewMask1, CostKind);
ReducedInstCount |= MergeInner(Z, NumSrcElts, NewMask0, CostKind);
ReducedInstCount |= MergeInner(W, NumSrcElts, NewMask1, CostKind);

InstructionCost NewCost =
TTI.getShuffleCost(SK0, BinOpTy, NewMask0, CostKind, 0, nullptr, {X, Z}) +
TTI.getShuffleCost(SK1, BinOpTy, NewMask1, CostKind, 0, nullptr, {Y, W});
Expand All @@ -1763,8 +1793,8 @@ bool VectorCombine::foldShuffleOfBinops(Instruction &I) {

// If either shuffle will constant fold away, then fold for the same cost as
// we will reduce the instruction count.
bool ReducedInstCount = (isa<Constant>(X) && isa<Constant>(Z)) ||
(isa<Constant>(Y) && isa<Constant>(W));
ReducedInstCount |= (isa<Constant>(X) && isa<Constant>(Z)) ||
(isa<Constant>(Y) && isa<Constant>(W));
if (ReducedInstCount ? (NewCost > OldCost) : (NewCost >= OldCost))
return false;

Expand Down
Loading
Loading