Skip to content

[LV] Create in-loop sub reductions #147026

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 10 commits into
base: main
Choose a base branch
from
Open

[LV] Create in-loop sub reductions #147026

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
b0d9d6c
[LV] Create in-loop sub reductions
SamTebbs33 Jul 1, 2025
247b6b5
Use empty flags for negation
SamTebbs33 Jul 16, 2025
c8969e1
Improve test comment
SamTebbs33 Jul 16, 2025
995d4da
Add RecurKind::Sub
SamTebbs33 Jul 16, 2025
7911926
Address review
SamTebbs33 Jul 18, 2025
9038996
Fix inloop reductions
SamTebbs33 Jul 21, 2025
5fac73d
Remove CurrentLinkI opcode check
SamTebbs33 Jul 21, 2025
64aa25f
Fix interleaving case
SamTebbs33 Jul 22, 2025
0b93c24
format
SamTebbs33 Jul 22, 2025
6adb449
Improve codegen
SamTebbs33 Jul 29, 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
1 change: 1 addition & 0 deletions llvm/include/llvm/Analysis/IVDescriptors.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,6 +35,7 @@ enum class RecurKind {
// clang-format off
None, ///< Not a recurrence.
Add, ///< Sum of integers.
Sub, ///< Subtraction of integers
Mul, ///< Product of integers.
Or, ///< Bitwise or logical OR of integers.
And, ///< Bitwise or logical AND of integers.
Expand Down
18 changes: 16 additions & 2 deletions llvm/lib/Analysis/IVDescriptors.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,7 @@ bool RecurrenceDescriptor::isIntegerRecurrenceKind(RecurKind Kind) {
switch (Kind) {
default:
break;
case RecurKind::Sub:
case RecurKind::Add:
case RecurKind::Mul:
case RecurKind::Or:
Expand Down Expand Up @@ -897,8 +898,9 @@ RecurrenceDescriptor::InstDesc RecurrenceDescriptor::isRecurrenceInstr(
case Instruction::PHI:
return InstDesc(I, Prev.getRecKind(), Prev.getExactFPMathInst());
case Instruction::Sub:
return InstDesc(Kind == RecurKind::Sub, I);
case Instruction::Add:
return InstDesc(Kind == RecurKind::Add, I);
return InstDesc(Kind == RecurKind::Add || Kind == RecurKind::Sub, I);
case Instruction::Mul:
return InstDesc(Kind == RecurKind::Mul, I);
case Instruction::And:
Expand All @@ -917,7 +919,8 @@ RecurrenceDescriptor::InstDesc RecurrenceDescriptor::isRecurrenceInstr(
I->hasAllowReassoc() ? nullptr : I);
case Instruction::Select:
if (Kind == RecurKind::FAdd || Kind == RecurKind::FMul ||
Kind == RecurKind::Add || Kind == RecurKind::Mul)
Kind == RecurKind::Add || Kind == RecurKind::Mul ||
Kind == RecurKind::Sub)
return isConditionalRdxPattern(I);
if (isFindIVRecurrenceKind(Kind) && SE)
return isFindIVPattern(Kind, L, OrigPhi, I, *SE);
Expand Down Expand Up @@ -1003,6 +1006,11 @@ bool RecurrenceDescriptor::isReductionPHI(PHINode *Phi, Loop *TheLoop,
LLVM_DEBUG(dbgs() << "Found an ADD reduction PHI." << *Phi << "\n");
return true;
}
if (AddReductionVar(Phi, RecurKind::Sub, TheLoop, FMF, RedDes, DB, AC, DT,
SE)) {
LLVM_DEBUG(dbgs() << "Found a SUB reduction PHI." << *Phi << "\n");
return true;
}
if (AddReductionVar(Phi, RecurKind::Mul, TheLoop, FMF, RedDes, DB, AC, DT,
SE)) {
LLVM_DEBUG(dbgs() << "Found a MUL reduction PHI." << *Phi << "\n");
Expand Down Expand Up @@ -1201,6 +1209,8 @@ bool RecurrenceDescriptor::isFixedOrderRecurrence(PHINode *Phi, Loop *TheLoop,

unsigned RecurrenceDescriptor::getOpcode(RecurKind Kind) {
switch (Kind) {
case RecurKind::Sub:
Copy link
Collaborator

Choose a reason for hiding this comment

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

RecurKind::Sub should just return Instruction::Sub as its opcode. I can understand there are places where you still need to use an add instead of a sub, for example in the outer-reduction when it needs to combine each UF's reduction result into a single one (because each of those is negative, they need to be added instead of subtracted), but that would be a special case. That would also simplify the generated code to use sub rather than using an add and having to explicitly negate the inputs.

Copy link
Collaborator Author

Choose a reason for hiding this comment

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

That's much better, thanks! Done.

return Instruction::Sub;
case RecurKind::Add:
return Instruction::Add;
case RecurKind::Mul:
Expand Down Expand Up @@ -1288,6 +1298,10 @@ RecurrenceDescriptor::getReductionOpChain(PHINode *Phi, Loop *L) const {
if (isFMulAddIntrinsic(Cur))
return true;

// Recognize a sub reduction. It gets canonicalized to add(sub (0, ...)).
if (Cur->getOpcode() == Instruction::Sub && getOpcode() == Instruction::Add)
return true;

return Cur->getOpcode() == getOpcode();
};

Expand Down
1 change: 1 addition & 0 deletions llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5092,6 +5092,7 @@ bool AArch64TTIImpl::isLegalToVectorizeReduction(
return false;

switch (RdxDesc.getRecurrenceKind()) {
case RecurKind::Sub:
case RecurKind::Add:
case RecurKind::FAdd:
case RecurKind::And:
Expand Down
2 changes: 2 additions & 0 deletions llvm/lib/Transforms/Utils/LoopUtils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -914,6 +914,7 @@ constexpr Intrinsic::ID llvm::getReductionIntrinsicID(RecurKind RK) {
switch (RK) {
default:
llvm_unreachable("Unexpected recurrence kind");
case RecurKind::Sub:
case RecurKind::Add:
return Intrinsic::vector_reduce_add;
case RecurKind::Mul:
Expand Down Expand Up @@ -1301,6 +1302,7 @@ Value *llvm::createSimpleReduction(IRBuilderBase &Builder, Value *Src,
Builder.getFastMathFlags());
};
switch (RdxKind) {
case RecurKind::Sub:
case RecurKind::Add:
case RecurKind::Mul:
case RecurKind::And:
Expand Down
24 changes: 17 additions & 7 deletions llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -810,10 +810,15 @@ Value *VPInstruction::generate(VPTransformState &State) {
Value *RdxPart = RdxParts[Part];
if (RecurrenceDescriptor::isMinMaxRecurrenceKind(RK))
ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart);
else
ReducedPartRdx = Builder.CreateBinOp(
(Instruction::BinaryOps)RecurrenceDescriptor::getOpcode(RK),
RdxPart, ReducedPartRdx, "bin.rdx");
else {
// Interleaved/unrolled sub reductions should still be added together
Instruction::BinaryOps Opcode =
RK == RecurKind::Sub
? Instruction::Add
: (Instruction::BinaryOps)RecurrenceDescriptor::getOpcode(RK);
ReducedPartRdx =
Builder.CreateBinOp(Opcode, RdxPart, ReducedPartRdx, "bin.rdx");
}
}
}

Expand Down Expand Up @@ -2524,10 +2529,15 @@ void VPReductionRecipe::execute(VPTransformState &State) {
NewRed = createSimpleReduction(State.Builder, NewVecOp, Kind);
if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind))
NextInChain = createMinMaxOp(State.Builder, Kind, NewRed, PrevInChain);
else
else {
// Sub reductions aren't commutative so the operands need to be swapped.
bool IsSub = Kind == RecurKind::Sub;
Value *LHS = IsSub ? PrevInChain : NewRed;
Value *RHS = IsSub ? NewRed : PrevInChain;
NextInChain = State.Builder.CreateBinOp(
(Instruction::BinaryOps)RecurrenceDescriptor::getOpcode(Kind), NewRed,
PrevInChain);
(Instruction::BinaryOps)RecurrenceDescriptor::getOpcode(Kind), LHS,
RHS);
}
}
State.set(this, NextInChain, /*IsScalar*/ true);
}
Expand Down
2 changes: 1 addition & 1 deletion llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1745,7 +1745,7 @@ void VPlanTransforms::clearReductionWrapFlags(VPlan &Plan) {
if (!PhiR)
continue;
RecurKind RK = PhiR->getRecurrenceKind();
if (RK != RecurKind::Add && RK != RecurKind::Mul)
if (RK != RecurKind::Add && RK != RecurKind::Mul && RK != RecurKind::Sub)
continue;

for (VPUser *U : collectUsersRecursively(PhiR))
Expand Down
Loading