Skip to content

[VPlan] Truncate/Extend ComputeReductionResult at construction (NFC). #141860

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

[VPlan] Truncate/Extend ComputeReductionResult at construction (NFC). #141860

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
52 changes: 29 additions & 23 deletions llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7226,7 +7226,10 @@ static void fixReductionScalarResumeWhenVectorizingEpilog(
// Get the VPInstruction computing the reduction result in the middle block.
// The first operand may not be from the middle block if it is not connected
// to the scalar preheader. In that case, there's nothing to fix.
auto *EpiRedResult = dyn_cast<VPInstruction>(EpiResumePhiR->getOperand(0));
VPValue *Incoming = EpiResumePhiR->getOperand(0);
match(Incoming, VPlanPatternMatch::m_ZExtOrSExt(
VPlanPatternMatch::m_VPValue(Incoming)));
auto *EpiRedResult = dyn_cast<VPInstruction>(Incoming);
if (!EpiRedResult ||
(EpiRedResult->getOpcode() != VPInstruction::ComputeAnyOfResult &&
EpiRedResult->getOpcode() != VPInstruction::ComputeReductionResult &&
Expand Down Expand Up @@ -9201,28 +9204,6 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
PhiR->setOperand(1, NewExitingVPV);
}

// If the vector reduction can be performed in a smaller type, we truncate
// then extend the loop exit value to enable InstCombine to evaluate the
// entire expression in the smaller type.
if (MinVF.isVector() && PhiTy != RdxDesc.getRecurrenceType() &&
!RecurrenceDescriptor::isAnyOfRecurrenceKind(
RdxDesc.getRecurrenceKind())) {
assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!");
Type *RdxTy = RdxDesc.getRecurrenceType();
auto *Trunc =
new VPWidenCastRecipe(Instruction::Trunc, NewExitingVPV, RdxTy);
auto *Extnd =
RdxDesc.isSigned()
? new VPWidenCastRecipe(Instruction::SExt, Trunc, PhiTy)
: new VPWidenCastRecipe(Instruction::ZExt, Trunc, PhiTy);

Trunc->insertAfter(NewExitingVPV->getDefiningRecipe());
Extnd->insertAfter(Trunc);
if (PhiR->getOperand(1) == NewExitingVPV)
PhiR->setOperand(1, Extnd->getVPSingleValue());
NewExitingVPV = Extnd;
}

// We want code in the middle block to appear to execute on the location of
// the scalar loop's latch terminator because: (a) it is all compiler
// generated, (b) these instructions are always executed after evaluating
Expand Down Expand Up @@ -9260,6 +9241,31 @@ void LoopVectorizationPlanner::adjustRecipesForReductions(
Builder.createNaryOp(VPInstruction::ComputeReductionResult,
{PhiR, NewExitingVPV}, Flags, ExitDL);
}
// If the vector reduction can be performed in a smaller type, we truncate
// then extend the loop exit value to enable InstCombine to evaluate the
// entire expression in the smaller type.
if (MinVF.isVector() && PhiTy != RdxDesc.getRecurrenceType() &&
!RecurrenceDescriptor::isAnyOfRecurrenceKind(
RdxDesc.getRecurrenceKind())) {
assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!");
Type *RdxTy = RdxDesc.getRecurrenceType();
auto *Trunc =
new VPWidenCastRecipe(Instruction::Trunc, NewExitingVPV, RdxTy);
Instruction::CastOps ExtendOpc =
RdxDesc.isSigned() ? Instruction::SExt : Instruction::ZExt;
auto *Extnd = new VPWidenCastRecipe(ExtendOpc, Trunc, PhiTy);
Trunc->insertAfter(NewExitingVPV->getDefiningRecipe());
Extnd->insertAfter(Trunc);
if (PhiR->getOperand(1) == NewExitingVPV)
PhiR->setOperand(1, Extnd->getVPSingleValue());

// Update ComputeReductionResult with the truncated exiting value and
// extend its result.
FinalReductionResult->setOperand(1, Trunc);
FinalReductionResult =
Builder.createScalarCast(ExtendOpc, FinalReductionResult, PhiTy, {});
}

// Update all users outside the vector region.
OrigExitingVPV->replaceUsesWithIf(
FinalReductionResult, [FinalReductionResult](VPUser &User, unsigned) {
Expand Down
18 changes: 1 addition & 17 deletions llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -668,7 +668,6 @@ Value *VPInstruction::generate(VPTransformState &State) {
assert(!RecurrenceDescriptor::isFindLastIVRecurrenceKind(RK) &&
"should be handled by ComputeFindLastIVResult");

Type *ResultTy = State.TypeAnalysis.inferScalarType(this);
// The recipe's operands are the reduction phi, followed by one operand for
// each part of the reduction.
unsigned UF = getNumOperands() - 1;
Expand All @@ -680,15 +679,6 @@ Value *VPInstruction::generate(VPTransformState &State) {
if (hasFastMathFlags())
Builder.setFastMathFlags(getFastMathFlags());

// If the vector reduction can be performed in a smaller type, we truncate
// then extend the loop exit value to enable InstCombine to evaluate the
// entire expression in the smaller type.
// TODO: Handle this in truncateToMinBW.
if (State.VF.isVector() && ResultTy != RdxDesc.getRecurrenceType()) {
Type *RdxVecTy = VectorType::get(RdxDesc.getRecurrenceType(), State.VF);
for (unsigned Part = 0; Part < UF; ++Part)
RdxParts[Part] = Builder.CreateTrunc(RdxParts[Part], RdxVecTy);
}
// Reduce all of the unrolled parts into a single vector.
Value *ReducedPartRdx = RdxParts[0];
if (PhiR->isOrdered()) {
Expand All @@ -713,13 +703,6 @@ Value *VPInstruction::generate(VPTransformState &State) {
// All ops in the reduction inherit fast-math-flags from the recurrence
// descriptor.
ReducedPartRdx = createSimpleReduction(Builder, ReducedPartRdx, RK);

// If the reduction can be performed in a smaller type, we need to extend
// the reduction to the wider type before we branch to the original loop.
if (ResultTy != RdxDesc.getRecurrenceType())
ReducedPartRdx = RdxDesc.isSigned()
? Builder.CreateSExt(ReducedPartRdx, ResultTy)
: Builder.CreateZExt(ReducedPartRdx, ResultTy);
}

return ReducedPartRdx;
Expand Down Expand Up @@ -1070,6 +1053,7 @@ void VPInstruction::print(raw_ostream &O, const Twine &Indent,
void VPInstructionWithType::execute(VPTransformState &State) {
State.setDebugLocFrom(getDebugLoc());
switch (getOpcode()) {
case Instruction::SExt:
case Instruction::ZExt:
case Instruction::Trunc: {
Value *Op = State.get(getOperand(0), VPLane(0));
Expand Down
3 changes: 1 addition & 2 deletions llvm/test/Transforms/LoopVectorize/X86/cost-model.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1167,8 +1167,7 @@ define i32 @narrowed_reduction(ptr %a, i1 %cmp) #0 {
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 16
; CHECK-NEXT: br i1 true, label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP28:![0-9]+]]
; CHECK: middle.block:
; CHECK-NEXT: [[TMP10:%.*]] = trunc <16 x i32> [[TMP7]] to <16 x i1>
; CHECK-NEXT: [[TMP20:%.*]] = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> [[TMP10]])
; CHECK-NEXT: [[TMP20:%.*]] = call i1 @llvm.vector.reduce.or.v16i1(<16 x i1> [[TMP5]])
; CHECK-NEXT: [[TMP21:%.*]] = zext i1 [[TMP20]] to i32
; CHECK-NEXT: br i1 true, label [[EXIT:%.*]], label [[VEC_EPILOG_PH]]
; CHECK: scalar.ph:
Expand Down
6 changes: 2 additions & 4 deletions llvm/test/Transforms/LoopVectorize/epilog-vectorization-reductions.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -208,8 +208,7 @@ define i16 @reduction_or_trunc(ptr noalias nocapture %ptr) {
; CHECK-NEXT: [[TMP8:%.*]] = icmp eq i32 [[INDEX_NEXT]], 256
; CHECK-NEXT: br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
; CHECK: middle.block:
; CHECK-NEXT: [[TMP9:%.*]] = trunc <4 x i32> [[TMP7]] to <4 x i16>
; CHECK-NEXT: [[TMP10:%.*]] = call i16 @llvm.vector.reduce.or.v4i16(<4 x i16> [[TMP9]])
; CHECK-NEXT: [[TMP10:%.*]] = call i16 @llvm.vector.reduce.or.v4i16(<4 x i16> [[TMP6]])
; CHECK-NEXT: [[TMP11:%.*]] = zext i16 [[TMP10]] to i32
; CHECK-NEXT: br i1 true, label [[FOR_END:%.*]], label [[VEC_EPILOG_ITER_CHECK:%.*]]
; CHECK: vec.epilog.iter.check:
Expand All @@ -234,8 +233,7 @@ define i16 @reduction_or_trunc(ptr noalias nocapture %ptr) {
; CHECK-NEXT: [[TMP21:%.*]] = icmp eq i32 [[INDEX_NEXT4]], 256
; CHECK-NEXT: br i1 [[TMP21]], label [[VEC_EPILOG_MIDDLE_BLOCK:%.*]], label [[VEC_EPILOG_VECTOR_BODY]], !llvm.loop [[LOOP9:![0-9]+]]
; CHECK: vec.epilog.middle.block:
; CHECK-NEXT: [[TMP22:%.*]] = trunc <4 x i32> [[TMP20]] to <4 x i16>
; CHECK-NEXT: [[TMP23:%.*]] = call i16 @llvm.vector.reduce.or.v4i16(<4 x i16> [[TMP22]])
; CHECK-NEXT: [[TMP23:%.*]] = call i16 @llvm.vector.reduce.or.v4i16(<4 x i16> [[TMP19]])
; CHECK-NEXT: [[TMP24:%.*]] = zext i16 [[TMP23]] to i32
; CHECK-NEXT: br i1 true, label [[FOR_END]], label [[VEC_EPILOG_SCALAR_PH]]
; CHECK: vec.epilog.scalar.ph:
Expand Down
9 changes: 3 additions & 6 deletions llvm/test/Transforms/LoopVectorize/reduction-small-size.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,8 +25,7 @@ define i8 @PR34687(i1 %c, i32 %x, i32 %n) {
; CHECK-NEXT: [[TMP5:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP5]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: middle.block:
; CHECK-NEXT: [[TMP6:%.*]] = trunc <4 x i32> [[TMP4]] to <4 x i8>
; CHECK-NEXT: [[TMP7:%.*]] = call i8 @llvm.vector.reduce.add.v4i8(<4 x i8> [[TMP6]])
; CHECK-NEXT: [[TMP7:%.*]] = call i8 @llvm.vector.reduce.add.v4i8(<4 x i8> [[TMP3]])
; CHECK-NEXT: [[TMP8:%.*]] = zext i8 [[TMP7]] to i32
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]]
; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
Expand Down Expand Up @@ -104,8 +103,7 @@ define i8 @PR34687_no_undef(i1 %c, i32 %x, i32 %n) {
; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
; CHECK: middle.block:
; CHECK-NEXT: [[TMP8:%.*]] = trunc <4 x i32> [[TMP6]] to <4 x i8>
; CHECK-NEXT: [[TMP9:%.*]] = call i8 @llvm.vector.reduce.add.v4i8(<4 x i8> [[TMP8]])
; CHECK-NEXT: [[TMP9:%.*]] = call i8 @llvm.vector.reduce.add.v4i8(<4 x i8> [[TMP5]])
; CHECK-NEXT: [[TMP10:%.*]] = zext i8 [[TMP9]] to i32
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[N]], [[N_VEC]]
; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
Expand Down Expand Up @@ -183,8 +181,7 @@ define i32 @PR35734(i32 %x, i32 %y) {
; CHECK-NEXT: [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
; CHECK: middle.block:
; CHECK-NEXT: [[TMP8:%.*]] = trunc <4 x i32> [[TMP6]] to <4 x i1>
; CHECK-NEXT: [[TMP9:%.*]] = call i1 @llvm.vector.reduce.add.v4i1(<4 x i1> [[TMP8]])
; CHECK-NEXT: [[TMP9:%.*]] = call i1 @llvm.vector.reduce.add.v4i1(<4 x i1> [[TMP5]])
; CHECK-NEXT: [[TMP10:%.*]] = sext i1 [[TMP9]] to i32
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP1]], [[N_VEC]]
; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
Expand Down
4 changes: 1 addition & 3 deletions llvm/test/Transforms/LoopVectorize/scalable-reduction-inloop.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,9 +43,7 @@ define i8 @reduction_add_trunc(ptr noalias nocapture %A) {
; CHECK-NEXT: [[TMP21:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP21]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: middle.block:
; CHECK-NEXT: [[TMP37:%.*]] = trunc <vscale x 8 x i32> [[TMP34]] to <vscale x 8 x i8>
; CHECK-NEXT: [[TMP38:%.*]] = trunc <vscale x 8 x i32> [[TMP36]] to <vscale x 8 x i8>
; CHECK-NEXT: [[BIN_RDX:%.*]] = add <vscale x 8 x i8> [[TMP38]], [[TMP37]]
; CHECK-NEXT: [[BIN_RDX:%.*]] = add <vscale x 8 x i8> [[TMP35]], [[TMP33]]
; CHECK-NEXT: [[TMP39:%.*]] = call i8 @llvm.vector.reduce.add.nxv8i8(<vscale x 8 x i8> [[BIN_RDX]])
; CHECK-NEXT: [[TMP40:%.*]] = zext i8 [[TMP39]] to i32
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i32 256, [[N_VEC]]
Expand Down
Loading