[llvm] [VPlan] Add ComputeFindLastIVResult opcode (NFC). (PR #132689)
via llvm-commits
llvm-commits at lists.llvm.org
Mon Mar 24 01:01:36 PDT 2025
llvmbot wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-vectorizers
Author: Florian Hahn (fhahn)
<details>
<summary>Changes</summary>
This moves the logic for computing the FindLastIV reduction result to its own opcode. A follow-up patch will update the new opcode to also take the start value, to fix
https://github.com/llvm/llvm-project/issues/126836.
---
Full diff: https://github.com/llvm/llvm-project/pull/132689.diff
5 Files Affected:
- (modified) llvm/lib/Transforms/Vectorize/LoopVectorize.cpp (+12-5)
- (modified) llvm/lib/Transforms/Vectorize/VPlan.h (+1)
- (modified) llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp (+1)
- (modified) llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp (+28-8)
- (modified) llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp (+2)
``````````diff
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 92160a421e59c..1168211e3d87b 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -7612,7 +7612,8 @@ static void fixReductionScalarResumeWhenVectorizingEpilog(
BasicBlock *BypassBlock) {
auto *EpiRedResult = dyn_cast<VPInstruction>(R);
if (!EpiRedResult ||
- EpiRedResult->getOpcode() != VPInstruction::ComputeReductionResult)
+ (EpiRedResult->getOpcode() != VPInstruction::ComputeReductionResult &&
+ EpiRedResult->getOpcode() != VPInstruction::ComputeFindLastIVResult))
return;
auto *EpiRedHeaderPhi =
@@ -9817,8 +9818,10 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
Builder.createSelect(Cond, OrigExitingVPV, PhiR, {}, "", FMFs);
OrigExitingVPV->replaceUsesWithIf(NewExitingVPV, [](VPUser &U, unsigned) {
return isa<VPInstruction>(&U) &&
- cast<VPInstruction>(&U)->getOpcode() ==
- VPInstruction::ComputeReductionResult;
+ (cast<VPInstruction>(&U)->getOpcode() ==
+ VPInstruction::ComputeReductionResult ||
+ cast<VPInstruction>(&U)->getOpcode() ==
+ VPInstruction::ComputeFindLastIVResult);
});
if (CM.usePredicatedReductionSelect(
PhiR->getRecurrenceDescriptor().getOpcode(), PhiTy))
@@ -9863,8 +9866,12 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
// also modeled in VPlan.
VPBuilder::InsertPointGuard Guard(Builder);
Builder.setInsertPoint(MiddleVPBB, IP);
- auto *FinalReductionResult = Builder.createNaryOp(
- VPInstruction::ComputeReductionResult, {PhiR, NewExitingVPV}, ExitDL);
+ auto *FinalReductionResult =
+ Builder.createNaryOp(RecurrenceDescriptor::isFindLastIVRecurrenceKind(
+ RdxDesc.getRecurrenceKind())
+ ? VPInstruction::ComputeFindLastIVResult
+ : VPInstruction::ComputeReductionResult,
+ {PhiR, NewExitingVPV}, ExitDL);
// Update all users outside the vector region.
OrigExitingVPV->replaceUsesWithIf(
FinalReductionResult, [FinalReductionResult](VPUser &User, unsigned) {
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 3059b87ae63c8..64e7f2bddb668 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -866,6 +866,7 @@ class VPInstruction : public VPRecipeWithIRFlags,
BranchOnCount,
BranchOnCond,
Broadcast,
+ ComputeFindLastIVResult,
ComputeReductionResult,
// Takes the VPValue to extract from as first operand and the lane or part
// to extract as second operand, counting from the end starting with 1 for
diff --git a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
index 38bec733dbf73..d404ce46fae4a 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanAnalysis.cpp
@@ -66,6 +66,7 @@ Type *VPTypeAnalysis::inferScalarTypeForRecipe(const VPInstruction *R) {
inferScalarType(R->getOperand(1)) &&
"different types inferred for different operands");
return IntegerType::get(Ctx, 1);
+ case VPInstruction::ComputeFindLastIVResult:
case VPInstruction::ComputeReductionResult: {
auto *PhiR = cast<VPReductionPHIRecipe>(R->getOperand(0));
auto *OrigPhi = cast<PHINode>(PhiR->getUnderlyingValue());
diff --git a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
index c7190b3187d94..2f1182399ee4a 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
@@ -614,6 +614,27 @@ Value *VPInstruction::generate(VPTransformState &State) {
return Builder.CreateVectorSplat(
State.VF, State.get(getOperand(0), /*IsScalar*/ true), "broadcast");
}
+ case VPInstruction::ComputeFindLastIVResult: {
+ // The recipe's operands are the reduction phi, followed by one operand for
+ // each part of the reduction.
+ unsigned UF = getNumOperands() - 1;
+ Value *ReducedPartRdx = State.get(getOperand(1));
+ for (unsigned Part = 1; Part < UF; ++Part) {
+ ReducedPartRdx = createMinMaxOp(Builder, RecurKind::SMax, ReducedPartRdx,
+ State.get(getOperand(1 + Part)));
+ }
+
+ // FIXME: The cross-recipe dependency on VPReductionPHIRecipe is temporary
+ // and will be removed by breaking up the recipe further.
+ auto *PhiR = cast<VPReductionPHIRecipe>(getOperand(0));
+ // Get its reduction variable descriptor.
+ const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
+ RecurKind RK = RdxDesc.getRecurrenceKind();
+
+ assert(RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK));
+ assert(!PhiR->isInLoop());
+ return createFindLastIVReduction(Builder, ReducedPartRdx, RdxDesc);
+ }
case VPInstruction::ComputeReductionResult: {
// FIXME: The cross-recipe dependency on VPReductionPHIRecipe is temporary
// and will be removed by breaking up the recipe further.
@@ -623,6 +644,8 @@ Value *VPInstruction::generate(VPTransformState &State) {
const RecurrenceDescriptor &RdxDesc = PhiR->getRecurrenceDescriptor();
RecurKind RK = RdxDesc.getRecurrenceKind();
+ assert(!RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK) &&
+ "should be handled by ComputeFindLastIVResult");
Type *PhiTy = OrigPhi->getType();
// The recipe's operands are the reduction phi, followed by one operand for
@@ -658,9 +681,6 @@ Value *VPInstruction::generate(VPTransformState &State) {
if (Op != Instruction::ICmp && Op != Instruction::FCmp)
ReducedPartRdx = Builder.CreateBinOp(
(Instruction::BinaryOps)Op, RdxPart, ReducedPartRdx, "bin.rdx");
- else if (RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK))
- ReducedPartRdx =
- createMinMaxOp(Builder, RecurKind::SMax, ReducedPartRdx, RdxPart);
else
ReducedPartRdx = createMinMaxOp(Builder, RK, ReducedPartRdx, RdxPart);
}
@@ -669,8 +689,7 @@ Value *VPInstruction::generate(VPTransformState &State) {
// Create the reduction after the loop. Note that inloop reductions create
// the target reduction in the loop using a Reduction recipe.
if ((State.VF.isVector() ||
- RecurrenceDescriptor::isAnyOfRecurrenceKind(RK) ||
- RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK)) &&
+ RecurrenceDescriptor::isAnyOfRecurrenceKind(RK)) &&
!PhiR->isInLoop()) {
// TODO: Support in-order reductions based on the recurrence descriptor.
// All ops in the reduction inherit fast-math-flags from the recurrence
@@ -681,9 +700,6 @@ Value *VPInstruction::generate(VPTransformState &State) {
if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RK))
ReducedPartRdx =
createAnyOfReduction(Builder, ReducedPartRdx, RdxDesc, OrigPhi);
- else if (RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK))
- ReducedPartRdx =
- createFindLastIVReduction(Builder, ReducedPartRdx, RdxDesc);
else
ReducedPartRdx = createSimpleReduction(Builder, ReducedPartRdx, RK);
@@ -829,6 +845,7 @@ bool VPInstruction::isVectorToScalar() const {
return getOpcode() == VPInstruction::ExtractFromEnd ||
getOpcode() == Instruction::ExtractElement ||
getOpcode() == VPInstruction::FirstActiveLane ||
+ getOpcode() == VPInstruction::ComputeFindLastIVResult ||
getOpcode() == VPInstruction::ComputeReductionResult ||
getOpcode() == VPInstruction::AnyOf;
}
@@ -1011,6 +1028,9 @@ void VPInstruction::print(raw_ostream &O, const Twine &Indent,
case VPInstruction::ExtractFromEnd:
O << "extract-from-end";
break;
+ case VPInstruction::ComputeFindLastIVResult:
+ O << "compute-find-last-iv-result";
+ break;
case VPInstruction::ComputeReductionResult:
O << "compute-reduction-result";
break;
diff --git a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
index a36c2aeb3da5c..ad957f33ee699 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanUnroll.cpp
@@ -348,6 +348,8 @@ void UnrollState::unrollBlock(VPBlockBase *VPB) {
// the parts to compute the final reduction value.
VPValue *Op1;
if (match(&R, m_VPInstruction<VPInstruction::ComputeReductionResult>(
+ m_VPValue(), m_VPValue(Op1))) ||
+ match(&R, m_VPInstruction<VPInstruction::ComputeFindLastIVResult>(
m_VPValue(), m_VPValue(Op1)))) {
addUniformForAllParts(cast<VPInstruction>(&R));
for (unsigned Part = 1; Part != UF; ++Part)
``````````
</details>
https://github.com/llvm/llvm-project/pull/132689
More information about the llvm-commits
mailing list