[VPlan] Split out VPBlendRecipe simplifications from simplifyRecipes. NFC #134073
+80
−68
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
… NFC This is split off from llvm#133977 VPBlendRecipe normalisation is sensitive to the number of users a mask has, so should probably be run after the masks are simplified as much as possible. Note this could be run after removeDeadRecipes but this causes test diffs, some regressions, so this is left to a later patch.
|
@llvm/pr-subscribers-llvm-transforms @llvm/pr-subscribers-vectorizers Author: Luke Lau (lukel97) ChangesThis is split off from #133977 VPBlendRecipe normalisation is sensitive to the number of users a mask has, so should probably be run after the masks are simplified as much as possible. Note this could be run after removeDeadRecipes but this causes test diffs, some regressions, so this is left to a later patch. Full diff: https://github.com/llvm/llvm-project/pull/134073.diff 2 Files Affected:
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
index 0c37db7f9d3a3..137ed330c31b2 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
@@ -926,74 +926,6 @@ static void recursivelyDeleteDeadRecipes(VPValue *V) {
static void simplifyRecipe(VPRecipeBase &R, VPTypeAnalysis &TypeInfo) {
using namespace llvm::VPlanPatternMatch;
- if (auto *Blend = dyn_cast<VPBlendRecipe>(&R)) {
- // Try to remove redundant blend recipes.
- SmallPtrSet<VPValue *, 4> UniqueValues;
- if (Blend->isNormalized() || !match(Blend->getMask(0), m_False()))
- UniqueValues.insert(Blend->getIncomingValue(0));
- for (unsigned I = 1; I != Blend->getNumIncomingValues(); ++I)
- if (!match(Blend->getMask(I), m_False()))
- UniqueValues.insert(Blend->getIncomingValue(I));
-
- if (UniqueValues.size() == 1) {
- Blend->replaceAllUsesWith(*UniqueValues.begin());
- Blend->eraseFromParent();
- return;
- }
-
- if (Blend->isNormalized())
- return;
-
- // Normalize the blend so its first incoming value is used as the initial
- // value with the others blended into it.
-
- unsigned StartIndex = 0;
- for (unsigned I = 0; I != Blend->getNumIncomingValues(); ++I) {
- // If a value's mask is used only by the blend then is can be deadcoded.
- // TODO: Find the most expensive mask that can be deadcoded, or a mask
- // that's used by multiple blends where it can be removed from them all.
- VPValue *Mask = Blend->getMask(I);
- if (Mask->getNumUsers() == 1 && !match(Mask, m_False())) {
- StartIndex = I;
- break;
- }
- }
-
- SmallVector<VPValue *, 4> OperandsWithMask;
- OperandsWithMask.push_back(Blend->getIncomingValue(StartIndex));
-
- for (unsigned I = 0; I != Blend->getNumIncomingValues(); ++I) {
- if (I == StartIndex)
- continue;
- OperandsWithMask.push_back(Blend->getIncomingValue(I));
- OperandsWithMask.push_back(Blend->getMask(I));
- }
-
- auto *NewBlend = new VPBlendRecipe(
- cast<PHINode>(Blend->getUnderlyingValue()), OperandsWithMask);
- NewBlend->insertBefore(&R);
-
- VPValue *DeadMask = Blend->getMask(StartIndex);
- Blend->replaceAllUsesWith(NewBlend);
- Blend->eraseFromParent();
- recursivelyDeleteDeadRecipes(DeadMask);
-
- /// Simplify BLEND %a, %b, Not(%mask) -> BLEND %b, %a, %mask.
- VPValue *NewMask;
- if (NewBlend->getNumOperands() == 3 &&
- match(NewBlend->getMask(1), m_Not(m_VPValue(NewMask)))) {
- VPValue *Inc0 = NewBlend->getOperand(0);
- VPValue *Inc1 = NewBlend->getOperand(1);
- VPValue *OldMask = NewBlend->getOperand(2);
- NewBlend->setOperand(0, Inc1);
- NewBlend->setOperand(1, Inc0);
- NewBlend->setOperand(2, NewMask);
- if (OldMask->getNumUsers() == 0)
- cast<VPInstruction>(OldMask)->eraseFromParent();
- }
- return;
- }
-
// VPScalarIVSteps can only be simplified after unrolling. VPScalarIVSteps for
// part 0 can be replaced by their start value, if only the first lane is
// demanded.
@@ -1092,6 +1024,85 @@ void VPlanTransforms::simplifyRecipes(VPlan &Plan, Type &CanonicalIVTy) {
}
}
+void VPlanTransforms::simplifyBlends(VPlan &Plan) {
+ using namespace llvm::VPlanPatternMatch;
+ ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT(
+ Plan.getEntry());
+ SetVector<VPRecipeBase *> Worklist;
+ for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {
+ for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {
+ auto *Blend = dyn_cast<VPBlendRecipe>(&R);
+ if (!Blend)
+ continue;
+
+ // Try to remove redundant blend recipes.
+ SmallPtrSet<VPValue *, 4> UniqueValues;
+ if (Blend->isNormalized() || !match(Blend->getMask(0), m_False()))
+ UniqueValues.insert(Blend->getIncomingValue(0));
+ for (unsigned I = 1; I != Blend->getNumIncomingValues(); ++I)
+ if (!match(Blend->getMask(I), m_False()))
+ UniqueValues.insert(Blend->getIncomingValue(I));
+
+ if (UniqueValues.size() == 1) {
+ Blend->replaceAllUsesWith(*UniqueValues.begin());
+ Blend->eraseFromParent();
+ continue;
+ }
+
+ if (Blend->isNormalized())
+ continue;
+
+ // Normalize the blend so its first incoming value is used as the initial
+ // value with the others blended into it.
+
+ unsigned StartIndex = 0;
+ for (unsigned I = 0; I != Blend->getNumIncomingValues(); ++I) {
+ // If a value's mask is used only by the blend then is can be deadcoded.
+ // TODO: Find the most expensive mask that can be deadcoded, or a mask
+ // that's used by multiple blends where it can be removed from them all.
+ VPValue *Mask = Blend->getMask(I);
+ if (Mask->getNumUsers() == 1 && !match(Mask, m_False())) {
+ StartIndex = I;
+ break;
+ }
+ }
+
+ SmallVector<VPValue *, 4> OperandsWithMask;
+ OperandsWithMask.push_back(Blend->getIncomingValue(StartIndex));
+
+ for (unsigned I = 0; I != Blend->getNumIncomingValues(); ++I) {
+ if (I == StartIndex)
+ continue;
+ OperandsWithMask.push_back(Blend->getIncomingValue(I));
+ OperandsWithMask.push_back(Blend->getMask(I));
+ }
+
+ auto *NewBlend = new VPBlendRecipe(
+ cast<PHINode>(Blend->getUnderlyingValue()), OperandsWithMask);
+ NewBlend->insertBefore(&R);
+
+ VPValue *DeadMask = Blend->getMask(StartIndex);
+ Blend->replaceAllUsesWith(NewBlend);
+ Blend->eraseFromParent();
+ recursivelyDeleteDeadRecipes(DeadMask);
+
+ /// Simplify BLEND %a, %b, Not(%mask) -> BLEND %b, %a, %mask.
+ VPValue *NewMask;
+ if (NewBlend->getNumOperands() == 3 &&
+ match(NewBlend->getMask(1), m_Not(m_VPValue(NewMask)))) {
+ VPValue *Inc0 = NewBlend->getOperand(0);
+ VPValue *Inc1 = NewBlend->getOperand(1);
+ VPValue *OldMask = NewBlend->getOperand(2);
+ NewBlend->setOperand(0, Inc1);
+ NewBlend->setOperand(1, Inc0);
+ NewBlend->setOperand(2, NewMask);
+ if (OldMask->getNumUsers() == 0)
+ cast<VPInstruction>(OldMask)->eraseFromParent();
+ }
+ }
+ }
+}
+
/// Optimize the width of vector induction variables in \p Plan based on a known
/// constant Trip Count, \p BestVF and \p BestUF.
static bool optimizeVectorInductionWidthForTCAndVFUF(VPlan &Plan,
@@ -1687,6 +1698,7 @@ void VPlanTransforms::optimize(VPlan &Plan) {
runPass(removeRedundantInductionCasts, Plan);
runPass(simplifyRecipes, Plan, *Plan.getCanonicalIV()->getScalarType());
+ runPass(simplifyBlends, Plan);
runPass(removeDeadRecipes, Plan);
runPass(legalizeAndOptimizeInductions, Plan);
runPass(removeRedundantExpandSCEVRecipes, Plan);
diff --git a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
index c23ff38265670..b008459ebad3f 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
+++ b/llvm/lib/Transforms/Vectorize/VPlanTransforms.h
@@ -183,6 +183,10 @@ struct VPlanTransforms {
/// CanonicalIVTy as type for all un-typed live-ins in VPTypeAnalysis.
static void simplifyRecipes(VPlan &Plan, Type &CanonicalIVTy);
+ /// Normalize and simplify VPBlendRecipes. Should be run after simplifyRecipes
+ /// to make sure the masks are simplified.
+ static void simplifyBlends(VPlan &Plan);
+
/// If there's a single exit block, optimize its phi recipes that use exiting
/// IV values by feeding them precomputed end values instead, possibly taken
/// one step backwards.
|
artagnon
reviewed
Apr 2, 2025
artagnon
reviewed
Apr 2, 2025
fhahn
reviewed
Apr 2, 2025
| using namespace llvm::VPlanPatternMatch; | ||
| ReversePostOrderTraversal<VPBlockDeepTraversalWrapper<VPBlockBase *>> RPOT( | ||
| Plan.getEntry()); | ||
| for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) { |
There was a problem hiding this comment.
Would it be sufficient to do a shallow traversal of the vector loop region?
There was a problem hiding this comment.
Looks like it, done in 4776c9c
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/66/builds/12220 Here is the relevant piece of the build log for the reference |
qiaojbao
pushed a commit
to GPUOpen-Drivers/llvm-project
that referenced
this pull request
Apr 29, 2025
Local branch origin/amd-gfx 943bd63 Merged main:146ad71bc71a into origin/amd-gfx:b969d463888c Remote branch main be6ccc9 [VPlan] Split out VPBlendRecipe simplifications from simplifyRecipes. NFC (llvm#134073)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is split off from #133977
VPBlendRecipe normalisation is sensitive to the number of users a mask has, so should probably be run after the masks are simplified as much as possible.
Note this could be run after removeDeadRecipes but this causes test diffs, some regressions, so this is left to a later patch.