Skip to content

[VPlan] Add exit phi operands during initial construction (NFC). #136455

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.

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 9 commits into from
Apr 23, 2025
6 changes: 1 addition & 5 deletions llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -9392,11 +9392,7 @@ collectUsersInExitBlocks(Loop *OrigLoop, VPRecipeBuilder &Builder,
continue;
Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: can R iterate over ExitVPBB->phis() above?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Updated in 8c83355, thanks

Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: is this early continue needed:

    // Nothing to do for unreachable exit blocks.
    if (ExitVPBB->getNumPredecessors() == 0)
      continue;

given that unreachable exit blocks have been emptied of their all their recipes - including phi ones?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Removed in e268f71, thanks

Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: loop-unswitch this condition

      if (ExitVPBB->getSinglePredecessor() != Plan.getMiddleBlock()) {
        assert(ExitIRI->getNumOperands() ==
                   ExitVPBB->getPredecessors().size() &&
               "early-exit must update exit values on construction");
        continue;

rather than check it for all phi's?
assert can be removed given that the number of operands of all phi recipes should be equal to the number of predecessors - check in verifier instead?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Will do, thanks

}

Copy link
Collaborator

Choose a reason for hiding this comment

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

Above "and add the exiting VPValue as operand" is now obsolete. Worth noting that users of multiple (early) exits are excluded?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done, aslo dropped .. modeled in VPlan, as all are now modeled in VPlan.

PHINode &ExitPhi = ExitIRI->getIRPhi();
BasicBlock *ExitingBB = OrigLoop->getLoopLatch();
Value *IncomingValue = ExitPhi.getIncomingValueForBlock(ExitingBB);
VPValue *V = Builder.getVPValueOrAddLiveIn(IncomingValue);
ExitIRI->addOperand(V);
VPValue *V = ExitIRI->getOperand(0);
Copy link
Collaborator

Choose a reason for hiding this comment

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

Worth asserting that ExitIRI has a single operand? Expected to match its single middle-block predecessor.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done, thanks

if (V->isLiveIn())
continue;
assert(V->getDefiningRecipe()->getParent()->getEnclosingLoopRegion() &&
Expand Down
19 changes: 19 additions & 0 deletions llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -352,6 +352,19 @@ std::unique_ptr<VPlan> PlainCFGBuilder::buildPlainCFG(
Plan->getEntry()->setOneSuccessor(getOrCreateVPBB(TheLoop->getHeader()));
Plan->getEntry()->setPlan(&*Plan);

// Add incoming operands for the VPIRInstructions wrapping the exit phis.
Copy link
Collaborator

Choose a reason for hiding this comment

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

based on their IR predecessors, as they still have no VPlan predecessors?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Yes, this will be the next change ;)

Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
// Add incoming operands for the VPIRInstructions wrapping the exit phis.
// Fix VPlan loop-closed-ssa exit phi's by add incoming operands to the VPIRInstructions wrapping them.

more consistent with above comment.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done thanks

for (auto *EB : Plan->getExitBlocks()) {
for (VPRecipeBase &R : *EB) {
Copy link
Collaborator

Choose a reason for hiding this comment

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

nit: can R iterate over EB->phis()?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done thanks

auto *PhiR = dyn_cast<VPIRPhi>(&R);
if (!PhiR)
break;
PHINode &Phi = PhiR->getIRPhi();
Copy link
Collaborator

Choose a reason for hiding this comment

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

assert PhiR is still w/o any operand?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done thanks

for (BasicBlock *Pred : predecessors(EB->getIRBasicBlock()))
PhiR->addOperand(
getOrCreateVPOperand(Phi.getIncomingValueForBlock(Pred)));
Comment on lines +366 to +368
Copy link
Collaborator

Choose a reason for hiding this comment

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

Order of operands set here corresponds to the order of predecessors of underlying IRBB, before EB VPBB has predecessors. This inconsistency requires attention later, when these predecessors are added, possibly in a different order. May be worth leaving a note.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done, thanks

}
}

for (const auto &[IRBB, VPB] : BB2VPBB)
VPB2IRBB[VPB] = IRBB;

Expand Down Expand Up @@ -464,6 +477,12 @@ void VPlanTransforms::createLoopRegions(VPlan &Plan, Type *InductionTy,
VPBlockUtils::connectBlocks(ScalarPH, Plan.getScalarHeader());
if (!RequiresScalarEpilogueCheck) {
VPBlockUtils::connectBlocks(MiddleVPBB, ScalarPH);
// The exit blocks are dead, remove any recipes to make sure no users remain
// that may pessimize transforms.
Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
// The exit blocks are dead, remove any recipes to make sure no users remain
// that may pessimize transforms.
// The exit blocks are unreachable, remove their recipes to make sure no users remain
// that may pessimize transforms.

Have dce take care of this?
The blocks themselves are collected lazily when the plan is destroyed, hence we don't simply erase the exit blocks here. But should ExitBlocks be emptied?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Doing this elsewhere is a bit tricky at the moment; the phis themselves cannot be removed, as they are part of the wrapped IR leave nodes.

We could empty them here, but some transforms may later have to add them back in the future, so might be better to keep as-is.

for (auto *EB : Plan.getExitBlocks()) {
for (VPRecipeBase &R : make_early_inc_range(*EB))
R.eraseFromParent();
}
return;
}

Copy link
Collaborator

Choose a reason for hiding this comment

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

The case

  // 2) If we require a scalar epilogue, there is no conditional branch as
  //    we unconditionally branch to the scalar preheader.  Do nothing.

is handle by early return above. Better place the explanation earlier, and replace "Do nothing" with "Empty the unreachable exit blocks of their recipes". The connection from scalar loop to exit blocks is (currently) outside of VPlan's scope.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done, thanks. Also changed 2->1 as this is the first handled case.

Expand Down
2 changes: 1 addition & 1 deletion llvm/lib/Transforms/Vectorize/VPlanRecipes.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -1139,7 +1139,7 @@ InstructionCost VPIRInstruction::computeCost(ElementCount VF,
void VPIRInstruction::extractLastLaneOfOperand(VPBuilder &Builder) {
assert(isa<PHINode>(getInstruction()) &&
Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: should this be implemented for VPIRPhi instead?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Will update separately, thanks

"can only add exiting operands to phi nodes");
Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent:

Suggested change
"can only add exiting operands to phi nodes");
"can only update exiting operands to phi nodes");

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Doen thanks.

assert(getNumOperands() == 1 && "must have a single operand");
assert(getNumOperands() > 0 && "must have at least one operand");
Copy link
Collaborator

Choose a reason for hiding this comment

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

And yet we continue to extract the first operand, only? extractLastLaneOf[First]Operand()

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Yep, updated the name, thanks

VPValue *Exiting = getOperand(0);
if (!Exiting->isLiveIn()) {
Comment on lines 1144 to 1145
Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
VPValue *Exiting = getOperand(0);
if (!Exiting->isLiveIn()) {
if (Exiting->isLiveIn())
return;
VPValue *Exiting = getOperand(0);

?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Will adjust separately, thanks

Copy link
Contributor Author

Choose a reason for hiding this comment

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

done in 71f2c1e

LLVMContext &Ctx = getInstruction().getContext();
Expand Down
28 changes: 17 additions & 11 deletions llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -2501,35 +2501,41 @@ void VPlanTransforms::handleUncountableEarlyExit(
if (!ExitIRI)
Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: have R iterate over VPEarlyExitBlock->phis()?

Independent: it's somewhat confusing having both VectorEarlyExitVPBB and VPEarlyExitBlock (connected to it). Perhaps the former is better called Middle2EarlyExitVPBB, along with Middle2EarlyExitB as its builder? Cf. https://llvm.org/docs/Vectorizers.html#early-exit-vectorization

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Will check and put up a PR to update here. (and. possibly docs)

break;

PHINode &ExitPhi = ExitIRI->getIRPhi();
VPValue *IncomingFromEarlyExit = RecipeBuilder.getVPValueOrAddLiveIn(
ExitPhi.getIncomingValueForBlock(UncountableExitingBlock));

unsigned EarlyExitIdx = 0;
Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
unsigned EarlyExitIdx = 0;
// By default, assume early exit operand is first, e.g., when the two exit blocks are distinct - VPEarlyExitBlock has a single predecessor.
unsigned EarlyExitIdx = 0;

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Added, thanks

Copy link
Collaborator

Choose a reason for hiding this comment

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

Hmm, perhaps its better instead to set EarlyExitIdx to be the last operand, consistently?

If VPEarlyExitBlock has two predecessors, they are already ordered such that early exit is second, by construction. But its underlying IRBB may have its predecessors ordered the other way around, and it is this order which corresponds to the order of operands of VPEarlyExitBlock's phi recipes. Therefore, if early exit is the first predecessor of the underlying IRBB, we swap the operands of phi recipes, bringing them to match VPEarlyExitBlock's predecessor order with early exit being last (second).

if (OrigLoop->getUniqueExitBlock()) {
Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: better ask instead if VPEarlyExitBlock has two predecessors? One (the last) was added above - VectorEarlyExitVPBB. This should match the case where OrigLoop has a unique exit block - which would be aka VPEarlyExitBlock, or rather its underlying EarlyExitIRBB.

Suggested change
if (OrigLoop->getUniqueExitBlock()) {
if (VPEarlyExitBlock->getNumPredecessors() == 2) {

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Will check separately, thanks

Copy link
Collaborator

Choose a reason for hiding this comment

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

Ok. It's more consistent with following explanation.

// After the transform, the first incoming value is coming from the
Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
// After the transform, the first incoming value is coming from the
// The incoming values currently correspond to the original IR predecessors.
// After the transform, the first incoming value should correspond to the

?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Added, thanks

// orignial loop latch, while the second operand is from the early exit.
Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
// orignial loop latch, while the second operand is from the early exit.
// original loop latch, while the second operand to the early exit.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Fixed thanks

// Sawp the phi operands, if the first predecessor in the original IR is
Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
// Sawp the phi operands, if the first predecessor in the original IR is
// Swap the phi operands if the first predecessor in the original IR is

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Fixed, thanks

// not the loop latch.
if (*pred_begin(VPEarlyExitBlock->getIRBasicBlock()) !=
OrigLoop->getLoopLatch())
Copy link
Collaborator

Choose a reason for hiding this comment

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

Is this clearer?

Suggested change
if (*pred_begin(VPEarlyExitBlock->getIRBasicBlock()) !=
OrigLoop->getLoopLatch())
if (*pred_begin(EarlyExitIRBB) == UncountableExitingBlock)

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done thanks

ExitIRI->swapOperands();

EarlyExitIdx = 1;
// If there's a unique exit block, VPEarlyExitBlock has 2 predecessors
Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
// If there's a unique exit block, VPEarlyExitBlock has 2 predecessors

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done thanks

// (MiddleVPBB and NewMiddle). Add the incoming value from MiddleVPBB
Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: worth indicating that early.exit and latch.exit may be the same block, in https://llvm.org/docs/Vectorizers.html#early-exit-vectorization

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Will do separately, thanks

// which is coming from the original latch.
Copy link
Collaborator

Choose a reason for hiding this comment

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

We're no longer adding this operand here, but only updating it with an extract.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done thanks

VPValue *IncomingFromLatch = RecipeBuilder.getVPValueOrAddLiveIn(
ExitPhi.getIncomingValueForBlock(OrigLoop->getLoopLatch()));
ExitIRI->addOperand(IncomingFromLatch);
ExitIRI->extractLastLaneOfOperand(MiddleBuilder);
Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: the extract from the latch created for (first) operand uses extractLastLaneOfOperand(), the extract from early exit replaced below explicitly. Can this be done more consistently?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Not sure, we could as follow-up add the extracts from last lane for all exit phis during construction, then have the extract from the early exit updated here?

}

VPValue *IncomingFromEarlyExit = ExitIRI->getOperand(EarlyExitIdx);
auto IsVector = [](ElementCount VF) { return VF.isVector(); };
// When the VFs are vectors, need to add `extract` to get the incoming value
// from early exit. When the range contains scalar VF, limit the range to
// scalar VF to prevent mis-compilation for the range containing both scalar
// and vector VFs.
if (!IncomingFromEarlyExit->isLiveIn() &&
LoopVectorizationPlanner::getDecisionAndClampRange(IsVector, Range)) {
Copy link
Collaborator

Choose a reason for hiding this comment

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

Independent: clamping range inside VPlanTransform? Limiting the range to scalar VF - is another VPlan constructed for the vector (sub)range?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Yes, this is done early on when we are clamping the loop range for other reasons as well.

Copy link
Collaborator

Choose a reason for hiding this comment

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

(Independent) Still puzzled about clamping the range at this stage, when all VPlans were already created following range clampings. Rather than say asserting that the range contains either scalar or vector VF's but not both, and introduce extracts if it's the latter. Extracts added above for latch exit need not check vector VF's?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

the transform isn't run when all VPlans have been created, but in tryToBuildVPlanWithVPRecipes where we the range is also clamped in other places ( https://github.com/llvm/llvm-project/blob/main/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp#L9758)

Copy link
Collaborator

Choose a reason for hiding this comment

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

Ahh, very well, thanks.

(Independent) May be good to indicate somehow transforms that may clamp the range - that belong to tryToBuildVPlanWithVPRecipes stage, and prevent them from operating afterwards, perhaps by disabling range clamping afterwards?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

The range should only be available in tryToBuildVPlanWithVPRecipes and not subsequent optimizations; we could seaprate things even clearer by first building all VPlans (i.e. only do tryToBuildVPlanWithVPRecipes) and then optimize all VPlans separatel

// Add the incoming value from the early exit.
VPValue *FirstActiveLane = EarlyExitB.createNaryOp(
VPInstruction::FirstActiveLane, {EarlyExitTakenCond}, nullptr,
"first.active.lane");
IncomingFromEarlyExit = EarlyExitB.createNaryOp(
Instruction::ExtractElement, {IncomingFromEarlyExit, FirstActiveLane},
nullptr, "early.exit.value");
ExitIRI->setOperand(
EarlyExitIdx,
EarlyExitB.createNaryOp(Instruction::ExtractElement,
{IncomingFromEarlyExit, FirstActiveLane},
nullptr, "early.exit.value"));
Copy link
Collaborator

Choose a reason for hiding this comment

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

Can continue to (re)set IncomingFromEarlyExit here, followed by resetting the operand (instead of adding it):
ExitIRI->setOperand(EarlyExitIdx, IncomingFromEarlyExit);
?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Done ,thanks

}
ExitIRI->addOperand(IncomingFromEarlyExit);
}
MiddleBuilder.createNaryOp(VPInstruction::BranchOnCond, {IsEarlyExitTaken});

Expand Down
Loading