mirrors/llvm-project
0
0
Fork 0
mirror of https://github.com/llvm/llvm-project.git synced 2025-04-18 19:16:43 +00:00
Code Issues Projects Releases Wiki Activity

[VPlan] Split out VPBlendRecipe simplifications from simplifyRecipes. NFC (#134073)

Browse Source 
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.
This commit is contained in:
Luke Lau 2025-04-07 09:55:52 +01:00 committed by GitHub
parent f20cb3f8d2
commit be6ccc98f3
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 80 additions and 68 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

148
llvm/lib/Transforms/Vectorize/VPlanTransforms.cpp
View File

@ -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) {
}
}
/// Normalize and simplify VPBlendRecipes. Should be run after simplifyRecipes
/// to make sure the masks are simplified.
static void simplifyBlends(VPlan &Plan) {
using namespace llvm::VPlanPatternMatch;
for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(
vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {
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,
@ -1733,6 +1744,7 @@ 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);