[VPlan] Manage noalias/alias_scope metadata in VPlan. (NFC) #136450
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
|
@llvm/pr-subscribers-vectorizers @llvm/pr-subscribers-llvm-transforms Author: Florian Hahn (fhahn) ChangesUse VPIRMetadata added in #135272 Note that this means we have to build the no-alias metadata up-front Patch is 53.26 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/136450.diff 10 Files Affected:
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h b/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
index f639f0adb9c43..d46c14f53b24f 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorizationPlanner.h
@@ -36,6 +36,7 @@ class LoopVectorizationLegality;
class LoopVectorizationCostModel;
class PredicatedScalarEvolution;
class LoopVectorizeHints;
+class LoopVersioning;
class OptimizationRemarkEmitter;
class TargetTransformInfo;
class TargetLibraryInfo;
@@ -515,7 +516,7 @@ class LoopVectorizationPlanner {
/// returned VPlan is valid for. If no VPlan can be built for the input range,
/// set the largest included VF to the maximum VF for which no plan could be
/// built.
- VPlanPtr tryToBuildVPlanWithVPRecipes(VFRange &Range);
+ VPlanPtr tryToBuildVPlanWithVPRecipes(VFRange &Range, LoopVersioning *LVer);
/// Build VPlans for power-of-2 VF's between \p MinVF and \p MaxVF inclusive,
/// according to the information gathered by Legal when it checked if it is
diff --git a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
index 7a5f618d09e95..f965d53f91bfb 100644
--- a/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -2371,7 +2371,7 @@ void InnerLoopVectorizer::scalarizeInstruction(const Instruction *Instr,
InputLane = VPLane::getFirstLane();
Cloned->setOperand(I.index(), State.get(Operand, InputLane));
}
- State.addNewMetadata(Cloned, Instr);
+ RepRecipe->applyMetadata(Cloned);
// Place the cloned scalar in the new loop.
State.Builder.Insert(Cloned);
@@ -7989,24 +7989,6 @@ DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(
if (VectorizingEpilogue)
VPlanTransforms::removeDeadRecipes(BestVPlan);
- // Only use noalias metadata when using memory checks guaranteeing no overlap
- // across all iterations.
- const LoopAccessInfo *LAI = ILV.Legal->getLAI();
- std::unique_ptr<LoopVersioning> LVer = nullptr;
- if (LAI && !LAI->getRuntimePointerChecking()->getChecks().empty() &&
- !LAI->getRuntimePointerChecking()->getDiffChecks()) {
-
- // We currently don't use LoopVersioning for the actual loop cloning but we
- // still use it to add the noalias metadata.
- // TODO: Find a better way to re-use LoopVersioning functionality to add
- // metadata.
- LVer = std::make_unique<LoopVersioning>(
- *LAI, LAI->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, DT,
- PSE.getSE());
- State.LVer = &*LVer;
- State.LVer->prepareNoAliasMetadata();
- }
-
ILV.printDebugTracesAtStart();
//===------------------------------------------------===//
@@ -8597,13 +8579,14 @@ VPRecipeBuilder::tryToWidenMemory(Instruction *I, ArrayRef<VPValue *> Operands,
Builder.insert(VectorPtr);
Ptr = VectorPtr;
}
+ auto Metadata = getMetadataToPropagate(I);
if (LoadInst *Load = dyn_cast<LoadInst>(I))
return new VPWidenLoadRecipe(*Load, Ptr, Mask, Consecutive, Reverse,
- I->getDebugLoc());
+ Metadata, I->getDebugLoc());
StoreInst *Store = cast<StoreInst>(I);
return new VPWidenStoreRecipe(*Store, Ptr, Operands[0], Mask, Consecutive,
- Reverse, I->getDebugLoc());
+ Reverse, Metadata, I->getDebugLoc());
}
/// Creates a VPWidenIntOrFpInductionRecpipe for \p Phi. If needed, it will also
@@ -8745,6 +8728,7 @@ VPSingleDefRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI,
Range);
if (ShouldUseVectorIntrinsic)
return new VPWidenIntrinsicRecipe(*CI, ID, Ops, CI->getType(),
+ getMetadataToPropagate(CI),
CI->getDebugLoc());
Function *Variant = nullptr;
@@ -8798,7 +8782,8 @@ VPSingleDefRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI,
}
Ops.push_back(Operands.back());
- return new VPWidenCallRecipe(CI, Variant, Ops, CI->getDebugLoc());
+ return new VPWidenCallRecipe(CI, Variant, Ops, getMetadataToPropagate(CI),
+ CI->getDebugLoc());
}
return nullptr;
@@ -8836,7 +8821,8 @@ VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I,
Plan.getOrAddLiveIn(ConstantInt::get(I->getType(), 1u, false));
auto *SafeRHS = Builder.createSelect(Mask, Ops[1], One, I->getDebugLoc());
Ops[1] = SafeRHS;
- return new VPWidenRecipe(*I, make_range(Ops.begin(), Ops.end()));
+ return new VPWidenRecipe(*I, make_range(Ops.begin(), Ops.end()),
+ getMetadataToPropagate(I));
}
[[fallthrough]];
}
@@ -8882,7 +8868,8 @@ VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I,
// For other binops, the legacy cost model only checks the second operand.
NewOps[1] = GetConstantViaSCEV(NewOps[1]);
}
- return new VPWidenRecipe(*I, make_range(NewOps.begin(), NewOps.end()));
+ return new VPWidenRecipe(*I, make_range(NewOps.begin(), NewOps.end()),
+ getMetadataToPropagate(I));
}
case Instruction::ExtractValue: {
SmallVector<VPValue *> NewOps(Operands);
@@ -8891,7 +8878,8 @@ VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I,
assert(EVI->getNumIndices() == 1 && "Expected one extractvalue index");
unsigned Idx = EVI->getIndices()[0];
NewOps.push_back(Plan.getOrAddLiveIn(ConstantInt::get(I32Ty, Idx, false)));
- return new VPWidenRecipe(*I, make_range(NewOps.begin(), NewOps.end()));
+ return new VPWidenRecipe(*I, make_range(NewOps.begin(), NewOps.end()),
+ getMetadataToPropagate(I));
}
};
}
@@ -8978,8 +8966,9 @@ VPRecipeBuilder::handleReplication(Instruction *I, ArrayRef<VPValue *> Operands,
assert((Range.Start.isScalar() || !IsUniform || !IsPredicated ||
(Range.Start.isScalable() && isa<IntrinsicInst>(I))) &&
"Should not predicate a uniform recipe");
- auto *Recipe = new VPReplicateRecipe(
- I, make_range(Operands.begin(), Operands.end()), IsUniform, BlockInMask);
+ auto *Recipe =
+ new VPReplicateRecipe(I, make_range(Operands.begin(), Operands.end()),
+ IsUniform, BlockInMask, getMetadataToPropagate(I));
return Recipe;
}
@@ -9096,6 +9085,20 @@ bool VPRecipeBuilder::getScaledReductions(
return false;
}
+SmallVector<std::pair<unsigned, MDNode *>>
+VPRecipeBuilder::getMetadataToPropagate(Instruction *I) const {
+ SmallVector<std::pair<unsigned, MDNode *>> Metadata;
+ ::getMetadataToPropagate(I, Metadata);
+ if (LVer && isa<LoadInst, StoreInst>(I)) {
+ const auto &[AliasScopeMD, NoAliasMD] = LVer->getNoAliasMetadataFor(I);
+ if (AliasScopeMD)
+ Metadata.emplace_back(LLVMContext::MD_alias_scope, AliasScopeMD);
+ if (NoAliasMD)
+ Metadata.emplace_back(LLVMContext::MD_noalias, NoAliasMD);
+ }
+ return Metadata;
+}
+
VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(
Instruction *Instr, ArrayRef<VPValue *> Operands, VFRange &Range) {
// First, check for specific widening recipes that deal with inductions, Phi
@@ -9168,13 +9171,14 @@ VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(
make_range(Operands.begin(), Operands.end()));
if (auto *SI = dyn_cast<SelectInst>(Instr)) {
- return new VPWidenSelectRecipe(
- *SI, make_range(Operands.begin(), Operands.end()));
+ return new VPWidenSelectRecipe(*SI,
+ make_range(Operands.begin(), Operands.end()),
+ getMetadataToPropagate(SI));
}
if (auto *CI = dyn_cast<CastInst>(Instr)) {
return new VPWidenCastRecipe(CI->getOpcode(), Operands[0], CI->getType(),
- *CI);
+ *CI, getMetadataToPropagate(CI));
}
return tryToWiden(Instr, Operands);
@@ -9200,7 +9204,8 @@ VPRecipeBuilder::tryToCreatePartialReduction(Instruction *Reduction,
SmallVector<VPValue *, 2> Ops;
Ops.push_back(Plan.getOrAddLiveIn(Zero));
Ops.push_back(BinOp);
- BinOp = new VPWidenRecipe(*Reduction, make_range(Ops.begin(), Ops.end()));
+ BinOp = new VPWidenRecipe(*Reduction, make_range(Ops.begin(), Ops.end()),
+ getMetadataToPropagate(Reduction));
Builder.insert(BinOp->getDefiningRecipe());
ReductionOpcode = Instruction::Add;
}
@@ -9223,10 +9228,22 @@ void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF,
ElementCount MaxVF) {
assert(OrigLoop->isInnermost() && "Inner loop expected.");
+ // Only use noalias metadata when using memory checks guaranteeing no overlap
+ // across all iterations.
+ const LoopAccessInfo *LAI = Legal->getLAI();
+ std::unique_ptr<LoopVersioning> LVer = nullptr;
+ if (LAI && !LAI->getRuntimePointerChecking()->getChecks().empty() &&
+ !LAI->getRuntimePointerChecking()->getDiffChecks()) {
+ LVer = std::make_unique<LoopVersioning>(
+ *LAI, LAI->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, DT,
+ PSE.getSE());
+ LVer->prepareNoAliasMetadata();
+ }
+
auto MaxVFTimes2 = MaxVF * 2;
for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFTimes2);) {
VFRange SubRange = {VF, MaxVFTimes2};
- if (auto Plan = tryToBuildVPlanWithVPRecipes(SubRange)) {
+ if (auto Plan = tryToBuildVPlanWithVPRecipes(SubRange, LVer.get())) {
bool HasScalarVF = Plan->hasScalarVFOnly();
// Now optimize the initial VPlan.
if (!HasScalarVF)
@@ -9534,7 +9551,8 @@ static void addExitUsersForFirstOrderRecurrences(
}
VPlanPtr
-LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VFRange &Range) {
+LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VFRange &Range,
+ LoopVersioning *LVer) {
using namespace llvm::VPlanPatternMatch;
SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups;
@@ -9580,7 +9598,7 @@ LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VFRange &Range) {
addCanonicalIVRecipes(*Plan, Legal->getWidestInductionType(), HasNUW, DL);
VPRecipeBuilder RecipeBuilder(*Plan, OrigLoop, TLI, &TTI, Legal, CM, PSE,
- Builder);
+ Builder, LVer);
// ---------------------------------------------------------------------------
// Pre-construction: record ingredients whose recipes we'll need to further
@@ -9694,8 +9712,9 @@ LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(VFRange &Range) {
Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) {
// Only create recipe for the final invariant store of the reduction.
if (Legal->isInvariantStoreOfReduction(SI)) {
- auto *Recipe =
- new VPReplicateRecipe(SI, R.operands(), true /* IsUniform */);
+ auto *Recipe = new VPReplicateRecipe(
+ SI, R.operands(), true /* IsUniform */, /*Mask*/ nullptr,
+ RecipeBuilder.getMetadataToPropagate(SI));
Recipe->insertBefore(*MiddleVPBB, MBIP);
}
R.eraseFromParent();
@@ -9881,7 +9900,7 @@ VPlanPtr LoopVectorizationPlanner::tryToBuildVPlan(VFRange &Range) {
// Collect mapping of IR header phis to header phi recipes, to be used in
// addScalarResumePhis.
VPRecipeBuilder RecipeBuilder(*Plan, OrigLoop, TLI, &TTI, Legal, CM, PSE,
- Builder);
+ Builder, nullptr);
for (auto &R : Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) {
if (isa<VPCanonicalIVPHIRecipe>(&R))
continue;
diff --git a/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h b/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
index fd0064a34c4c9..357a68972cb43 100644
--- a/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
+++ b/llvm/lib/Transforms/Vectorize/VPRecipeBuilder.h
@@ -90,6 +90,10 @@ class VPRecipeBuilder {
/// A mapping of partial reduction exit instructions to their scaling factor.
DenseMap<const Instruction *, unsigned> ScaledReductionMap;
+ /// Loop versioning instance for getting noalias metadata guaranteed by
+ /// runtime checks.
+ LoopVersioning *LVer;
+
/// Check if \p I can be widened at the start of \p Range and possibly
/// decrease the range such that the returned value holds for the entire \p
/// Range. The function should not be called for memory instructions or calls.
@@ -155,9 +159,10 @@ class VPRecipeBuilder {
const TargetTransformInfo *TTI,
LoopVectorizationLegality *Legal,
LoopVectorizationCostModel &CM,
- PredicatedScalarEvolution &PSE, VPBuilder &Builder)
+ PredicatedScalarEvolution &PSE, VPBuilder &Builder,
+ LoopVersioning *LVer)
: Plan(Plan), OrigLoop(OrigLoop), TLI(TLI), TTI(TTI), Legal(Legal),
- CM(CM), PSE(PSE), Builder(Builder) {}
+ CM(CM), PSE(PSE), Builder(Builder), LVer(LVer) {}
std::optional<unsigned> getScalingForReduction(const Instruction *ExitInst) {
auto It = ScaledReductionMap.find(ExitInst);
@@ -233,6 +238,11 @@ class VPRecipeBuilder {
}
return Plan.getOrAddLiveIn(V);
}
+
+ /// Returns the metatadata that can be preserved from the original instruction
+ /// \p I, including noalias metadata guaranteed by runtime checks.
+ SmallVector<std::pair<unsigned, MDNode *>>
+ getMetadataToPropagate(Instruction *I) const;
};
} // end namespace llvm
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.cpp b/llvm/lib/Transforms/Vectorize/VPlan.cpp
index fa2d95a44609a..7ca5008c607cf 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlan.cpp
@@ -220,8 +220,8 @@ VPTransformState::VPTransformState(const TargetTransformInfo *TTI,
InnerLoopVectorizer *ILV, VPlan *Plan,
Loop *CurrentParentLoop, Type *CanonicalIVTy)
: TTI(TTI), VF(VF), CFG(DT), LI(LI), Builder(Builder), ILV(ILV), Plan(Plan),
- CurrentParentLoop(CurrentParentLoop), LVer(nullptr),
- TypeAnalysis(CanonicalIVTy), VPDT(*Plan) {}
+ CurrentParentLoop(CurrentParentLoop), TypeAnalysis(CanonicalIVTy),
+ VPDT(*Plan) {}
Value *VPTransformState::get(const VPValue *Def, const VPLane &Lane) {
if (Def->isLiveIn())
@@ -355,25 +355,6 @@ BasicBlock *VPTransformState::CFGState::getPreheaderBBFor(VPRecipeBase *R) {
return VPBB2IRBB[LoopRegion->getPreheaderVPBB()];
}
-void VPTransformState::addNewMetadata(Instruction *To,
- const Instruction *Orig) {
- // If the loop was versioned with memchecks, add the corresponding no-alias
- // metadata.
- if (LVer && isa<LoadInst, StoreInst>(Orig))
- LVer->annotateInstWithNoAlias(To, Orig);
-}
-
-void VPTransformState::addMetadata(Value *To, Instruction *From) {
- // No source instruction to transfer metadata from?
- if (!From)
- return;
-
- if (Instruction *ToI = dyn_cast<Instruction>(To)) {
- propagateMetadata(ToI, From);
- addNewMetadata(ToI, From);
- }
-}
-
void VPTransformState::setDebugLocFrom(DebugLoc DL) {
const DILocation *DIL = DL;
// When a FSDiscriminator is enabled, we don't need to add the multiply
diff --git a/llvm/lib/Transforms/Vectorize/VPlan.h b/llvm/lib/Transforms/Vectorize/VPlan.h
index 7084676af6d5b..585fa3cd12ef6 100644
--- a/llvm/lib/Transforms/Vectorize/VPlan.h
+++ b/llvm/lib/Transforms/Vectorize/VPlan.h
@@ -1190,28 +1190,49 @@ struct VPIRPhi : public VPIRInstruction {
#endif
};
+using MDArrayRef = ArrayRef<std::pair<unsigned, MDNode *>>;
+
+/// Helper to manage IR metadata for recipes. It filters out metadata that
+/// cannot be proagated.
+class VPIRMetadata {
+ SmallVector<std::pair<unsigned, MDNode *>> Metadata;
+
+protected:
+ VPIRMetadata(MDArrayRef Metadata) : Metadata(Metadata) {}
+
+public:
+ /// Add all metadata to \p V if it is an instruction.
+ void applyMetadata(Value *V) const;
+
+ /// Return the IR metadata.
+ MDArrayRef getMetadata() const { return Metadata; }
+};
+
/// VPWidenRecipe is a recipe for producing a widened instruction using the
/// opcode and operands of the recipe. This recipe covers most of the
/// traditional vectorization cases where each recipe transforms into a
/// vectorized version of itself.
-class VPWidenRecipe : public VPRecipeWithIRFlags {
+class VPWidenRecipe : public VPRecipeWithIRFlags, public VPIRMetadata {
unsigned Opcode;
protected:
template <typename IterT>
VPWidenRecipe(unsigned VPDefOpcode, Instruction &I,
- iterator_range<IterT> Operands)
- : VPRecipeWithIRFlags(VPDefOpcode, Operands, I), Opcode(I.getOpcode()) {}
+ iterator_range<IterT> Operands, MDArrayRef Metadata)
+ : VPRecipeWithIRFlags(VPDefOpcode, Operands, I), VPIRMetadata(Metadata),
+ Opcode(I.getOpcode()) {}
public:
template <typename IterT>
- VPWidenRecipe(Instruction &I, iterator_range<IterT> Operands)
- : VPWidenRecipe(VPDef::VPWidenSC, I, Operands) {}
+ VPWidenRecipe(Instruction &I, iterator_range<IterT> Operands,
+ MDArrayRef Metadata)
+ : VPWidenRecipe(VPDef::VPWidenSC, I, Operands, Metadata) {}
~VPWidenRecipe() override = default;
VPWidenRecipe *clone() override {
- auto *R = new VPWidenRecipe(*getUnderlyingInstr(), operands());
+ auto *R =
+ new VPWidenRecipe(*getUnderlyingInstr(), operands(), getMetadata());
R->transferFlags(*this);
return R;
}
@@ -1236,7 +1257,7 @@ class VPWidenRecipe : public VPRecipeWithIRFlags {
};
/// VPWidenCastRecipe is a recipe to create vector cast instructions.
-class VPWidenCastRecipe : public VPRecipeWithIRFlags {
+class VPWidenCastRecipe : public VPRecipeWithIRFlags, public VPIRMetadata {
/// Cast instruction opcode.
Instruction::CastOps Opcode;
@@ -1245,23 +1266,23 @@ class VPWidenCastRecipe : public VPRecipeWithIRFlags {
public:
VPWidenCastRecipe(Instruction::CastOps Opcode, VPValue *Op, Type *ResultTy,
- CastInst &UI)
- : VPRecipeWithIRFlags(VPDef::VPWidenCastSC, Op, UI), Opcode(Opcode),
- ResultTy(ResultTy) {
+ CastInst &UI, MDArrayRef Metadata)
+ : VPRecipeWithIRFlags(VPDef::VPWidenCastSC, Op, UI),
+ VPIRMetadata(Metadata), Opcode(Opcode), ResultTy(ResultTy) {
assert(UI.getOpcode() == Opcode &&
"opcode of underlying cast doesn't match");
}
VPWidenCastRecipe(Instruction::CastOps Opcode, VPValue *Op, Type *ResultTy)
- : VPRecipeWithIRFlags(VPDef::VPWidenCastSC, Op), Opcode(Opcode),
- ResultTy(ResultTy) {}
+ : VPRecipeWithIRFlags(VPDef::VPWidenCastSC, Op), VPIRMetadata({}),
+ Opcode(Opcode), ResultTy(ResultTy) {}
~VPWidenCastRecipe() override = default;
VPWidenCastRecipe *clone() override {
if (auto *UV = getUnderlyingValue())
return new VPWidenCastRecipe(Opcode, getOperand(0), ResultTy,
- *cast<CastInst>(UV));
+ *cast<CastInst>(UV), getMetadata());
return new VPWidenCastRecipe(Opcode, getOperand(0), ResultTy);
}
@@ -1288,7 +1309,7 @@ class VPWidenCastRecipe : public VPRecipeWithIRFlags {
};
/// A recipe for widening vector intrinsics.
-class VPWidenIntrinsicRecipe : public VPRecipeWithIRFlags {
+class VPWidenIntrinsicRecipe : public VPRecipeWithIRFlags, public VPIRMetadata {
/// ID of the vector intrinsic to widen.
Intrinsic::ID VectorIntrinsicID;
@@ -1307,18 +1328,19 @@ class VPWidenIntrinsicRecipe : public VPRecipeWithIRFlags {
public:
VPWidenIntrinsicRecipe(CallInst &CI, Intrinsic::ID VectorIntrinsicID,
ArrayRef<VPValue *> CallArguments, Type *Ty,
- DebugLoc DL = {})
+ MDArrayRef Metadata, DebugLoc DL = {})
: VPRecipeWithIRFlags(VPDef::VPWidenIntrinsicSC, CallArguments, CI),
- VectorIntrinsicID(VectorIntrinsicID), ResultTy(Ty),
- MayReadFromMemory(CI.mayReadFromMemory()),
+ VPIRMetadata(Metadata), VectorIntrinsicID(VectorIntrinsicID),
+ ResultTy(Ty), MayRe...
[truncated]
|
Use VPIRMetadata added in llvm#135272 to also manage no-alias metadata added by versioning. Note that this means we have to build the no-alias metadata up-front once. If it is not used, it will be discarded automatically.
fhahn
force-pushed
the
perf/vplan-noalias-md
branch
from
2fd7844 to
55f1a6b
Compare
|
Could the no-alias metadata added by loop versioning be managed by a dedicated VPlan-to-VPlan pass, continuing to capture only metadata that propagates from underlying instructions during initial VPlan/recipe construction? May require introducing API to update recipe metadata, but hopefully simple as it merely augments additional nodes? |
We could, but we at the moment this would still mean we will be tied to the underlying instruction, which is then used to look up the metadata, which is why added it on construction of the recipes. We could add VPIRMetadata to VPInstruction, and add the metadata before creating widen recipes (at that point we must have the underlying instruction available, although it might warrant even more a separate class for such VPInstructions at the early stage, raises the question what it should be called and its relationship with VPIRInstruction). Or change the lookup of the metadata to be based on SCEV, which would require building SCEVs from various VP recipes. Both may also be good to do as follow-ups |
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.
Use VPIRMetadata added in #135272
to also manage no-alias metadata added by versioning.
Note that this means we have to build the no-alias metadata up-front
once. If it is not used, it will be discarded automatically.
Compile-time impact is neutral:
https://llvm-compile-time-tracker.com/compare.php?from=38bf1af41c5425a552a53feb13c71d82873f1c18&to=2fd7844cfdf5ec0f1c2ce0b9b3ae0763245b6922&stat=instructions:u