LLVM 20.0.0git
|
LoopVectorizationLegality checks if it is legal to vectorize a loop, and to what vectorization factor. More...
#include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h"
Public Types | |
using | ReductionList = MapVector< PHINode *, RecurrenceDescriptor > |
ReductionList contains the reduction descriptors for all of the reductions that were found in the loop. | |
using | InductionList = MapVector< PHINode *, InductionDescriptor > |
InductionList saves induction variables and maps them to the induction descriptor. | |
using | RecurrenceSet = SmallPtrSet< const PHINode *, 8 > |
RecurrenceSet contains the phi nodes that are recurrences other than inductions and reductions. | |
Public Member Functions | |
LoopVectorizationLegality (Loop *L, PredicatedScalarEvolution &PSE, DominatorTree *DT, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, Function *F, LoopAccessInfoManager &LAIs, LoopInfo *LI, OptimizationRemarkEmitter *ORE, LoopVectorizationRequirements *R, LoopVectorizeHints *H, DemandedBits *DB, AssumptionCache *AC, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI) | |
bool | canVectorize (bool UseVPlanNativePath) |
Returns true if it is legal to vectorize this loop. | |
bool | canVectorizeFPMath (bool EnableStrictReductions) |
Returns true if it is legal to vectorize the FP math operations in this loop. | |
bool | canFoldTailByMasking () const |
Return true if we can vectorize this loop while folding its tail by masking. | |
void | prepareToFoldTailByMasking () |
Mark all respective loads/stores for masking. | |
PHINode * | getPrimaryInduction () |
Returns the primary induction variable. | |
const ReductionList & | getReductionVars () const |
Returns the reduction variables found in the loop. | |
const InductionList & | getInductionVars () const |
Returns the induction variables found in the loop. | |
RecurrenceSet & | getFixedOrderRecurrences () |
Return the fixed-order recurrences found in the loop. | |
Type * | getWidestInductionType () |
Returns the widest induction type. | |
bool | isInvariantStoreOfReduction (StoreInst *SI) |
Returns True if given store is a final invariant store of one of the reductions found in the loop. | |
bool | isInvariantAddressOfReduction (Value *V) |
Returns True if given address is invariant and is used to store recurrent expression. | |
bool | isInductionPhi (const Value *V) const |
Returns True if V is a Phi node of an induction variable in this loop. | |
const InductionDescriptor * | getIntOrFpInductionDescriptor (PHINode *Phi) const |
Returns a pointer to the induction descriptor, if Phi is an integer or floating point induction. | |
const InductionDescriptor * | getPointerInductionDescriptor (PHINode *Phi) const |
Returns a pointer to the induction descriptor, if Phi is pointer induction. | |
bool | isCastedInductionVariable (const Value *V) const |
Returns True if V is a cast that is part of an induction def-use chain, and had been proven to be redundant under a runtime guard (in other words, the cast has the same SCEV expression as the induction phi). | |
bool | isInductionVariable (const Value *V) const |
Returns True if V can be considered as an induction variable in this loop. | |
bool | isReductionVariable (PHINode *PN) const |
Returns True if PN is a reduction variable in this loop. | |
bool | isFixedOrderRecurrence (const PHINode *Phi) const |
Returns True if Phi is a fixed-order recurrence in this loop. | |
bool | blockNeedsPredication (BasicBlock *BB) const |
Return true if the block BB needs to be predicated in order for the loop to be vectorized. | |
int | isConsecutivePtr (Type *AccessTy, Value *Ptr) const |
Check if this pointer is consecutive when vectorizing. | |
bool | isInvariant (Value *V) const |
Returns true if V is invariant across all loop iterations according to SCEV. | |
bool | isUniform (Value *V, ElementCount VF) const |
Returns true if value V is uniform across VF lanes, when VF is provided, and otherwise if V is invariant across all loop iterations. | |
bool | isUniformMemOp (Instruction &I, ElementCount VF) const |
A uniform memory op is a load or store which accesses the same memory location on all VF lanes, if VF is provided and otherwise if the memory location is invariant. | |
const RuntimePointerChecking * | getRuntimePointerChecking () const |
Returns the information that we collected about runtime memory check. | |
const LoopAccessInfo * | getLAI () const |
bool | isSafeForAnyVectorWidth () const |
uint64_t | getMaxSafeVectorWidthInBits () const |
bool | hasUncountableEarlyExit () const |
Returns true if the loop has an uncountable early exit, i.e. | |
BasicBlock * | getUncountableEarlyExitingBlock () const |
Returns the uncountable early exiting block. | |
BasicBlock * | getUncountableEarlyExitBlock () const |
Returns the destination of an uncountable early exiting block. | |
bool | isMaskRequired (const Instruction *I) const |
Returns true if vector representation of the instruction I requires mask. | |
bool | hasVectorCallVariants () const |
Returns true if there is at least one function call in the loop which has a vectorized variant available. | |
unsigned | getNumStores () const |
unsigned | getNumLoads () const |
std::optional< const HistogramInfo * > | getHistogramInfo (Instruction *I) const |
Returns a HistogramInfo* for the given instruction if it was determined to be part of a load -> update -> store sequence where multiple lanes may be working on the same memory address. | |
bool | hasHistograms () const |
Returns a list of all known histogram operations in the loop. | |
PredicatedScalarEvolution * | getPredicatedScalarEvolution () const |
Loop * | getLoop () const |
LoopInfo * | getLoopInfo () const |
AssumptionCache * | getAssumptionCache () const |
ScalarEvolution * | getScalarEvolution () const |
DominatorTree * | getDominatorTree () const |
const SmallVector< BasicBlock *, 4 > & | getCountableExitingBlocks () const |
Returns all exiting blocks with a countable exit, i.e. | |
const SmallVector< BasicBlock *, 4 > & | getUncountableExitingBlocks () const |
Returns all the exiting blocks with an uncountable exit. | |
SmallVector< BasicBlock *, 4 > | getUncountableExitBlocks () const |
Returns all the exit blocks from uncountable exiting blocks. | |
LoopVectorizationLegality checks if it is legal to vectorize a loop, and to what vectorization factor.
This class does not look at the profitability of vectorization, only the legality. This class has two main kinds of checks:
Definition at line 252 of file LoopVectorizationLegality.h.
InductionList saves induction variables and maps them to the induction descriptor.
Definition at line 270 of file LoopVectorizationLegality.h.
RecurrenceSet contains the phi nodes that are recurrences other than inductions and reductions.
Definition at line 274 of file LoopVectorizationLegality.h.
ReductionList contains the reduction descriptors for all of the reductions that were found in the loop.
Definition at line 266 of file LoopVectorizationLegality.h.
|
inline |
Definition at line 254 of file LoopVectorizationLegality.h.
bool llvm::LoopVectorizationLegality::blockNeedsPredication | ( | BasicBlock * | BB | ) | const |
Return true if the block BB needs to be predicated in order for the loop to be vectorized.
Definition at line 1372 of file LoopVectorizationLegality.cpp.
References assert(), llvm::LoopAccessInfo::blockNeedsPredication(), llvm::LoopBase< BlockT, LoopT >::getLoopLatch(), getUncountableEarlyExitingBlock(), hasUncountableEarlyExit(), llvm::is_contained(), and llvm::predecessors().
Referenced by llvm::LoopVectorizationCostModel::blockNeedsPredicationForAnyReason(), llvm::LoopVectorizationCostModel::expectedCost(), llvm::LoopVectorizationCostModel::isPredicatedInst(), and isUniformMemOp().
bool llvm::LoopVectorizationLegality::canFoldTailByMasking | ( | ) | const |
Return true if we can vectorize this loop while folding its tail by masking.
Definition at line 1838 of file LoopVectorizationLegality.cpp.
References llvm::LoopBase< BlockT, LoopT >::blocks(), llvm::LoopBase< BlockT, LoopT >::contains(), llvm::SmallPtrSetImpl< PtrType >::count(), llvm::dbgs(), getInductionVars(), getReductionVars(), llvm::SmallPtrSetImpl< PtrType >::insert(), LLVM_DEBUG, Reduction, and llvm::Value::users().
Referenced by llvm::LoopVectorizationCostModel::setTailFoldingStyles().
Returns true if it is legal to vectorize this loop.
This does not mean that it is profitable to vectorize this loop, only that it is legal to do so. Temporarily taking UseVPlanNativePath parameter. If true, take the new code path being implemented for outer loop vectorization (should be functional for inner loop vectorization) based on VPlan. If false, good old LV code.
Definition at line 1727 of file LoopVectorizationLegality.cpp.
References llvm::OptimizationRemarkEmitter::allowExtraAnalysis(), assert(), llvm::dbgs(), DEBUG_TYPE, llvm::LoopVectorizeHints::FK_Enabled, llvm::PredicatedScalarEvolution::getBackedgeTakenCount(), llvm::SCEVPredicate::getComplexity(), llvm::LoopBase< BlockT, LoopT >::getHeader(), llvm::Value::getName(), llvm::LoopBase< BlockT, LoopT >::getNumBlocks(), llvm::PredicatedScalarEvolution::getPredicate(), llvm::LoopAccessInfo::getRuntimePointerChecking(), llvm::LoopBase< BlockT, LoopT >::isInnermost(), LLVM_DEBUG, llvm::RuntimePointerChecking::Need, PragmaVectorizeSCEVCheckThreshold, llvm::reportVectorizationFailure(), and VectorizeSCEVCheckThreshold.
Referenced by llvm::LoopVectorizePass::processLoop().
Returns true if it is legal to vectorize the FP math operations in this loop.
Vectorizing is legal if we allow reordering of FP operations, or if we can use in-order reductions.
Definition at line 1281 of file LoopVectorizationLegality.cpp.
References llvm::all_of(), llvm::any_of(), llvm::LoopVectorizationRequirements::getExactFPInst(), llvm::InductionDescriptor::getExactFPMathInst(), getInductionVars(), getReductionVars(), llvm::RecurrenceDescriptor::hasExactFPMath(), llvm::RecurrenceDescriptor::isOrdered(), and Reduction.
Referenced by llvm::LoopVectorizePass::processLoop().
|
inline |
Definition at line 450 of file LoopVectorizationLegality.h.
Referenced by llvm::ARMTTIImpl::preferPredicateOverEpilogue().
|
inline |
Returns all exiting blocks with a countable exit, i.e.
the exit-not-taken count is known exactly at compile time.
Definition at line 458 of file LoopVectorizationLegality.h.
Referenced by llvm::LoopVectorizePass::processLoop().
|
inline |
Definition at line 454 of file LoopVectorizationLegality.h.
Referenced by llvm::ARMTTIImpl::preferPredicateOverEpilogue().
|
inline |
Return the fixed-order recurrences found in the loop.
Definition at line 308 of file LoopVectorizationLegality.h.
Referenced by llvm::AArch64TTIImpl::preferPredicateOverEpilogue().
|
inline |
Returns a HistogramInfo* for the given instruction if it was determined to be part of a load -> update -> store sequence where multiple lanes may be working on the same memory address.
Definition at line 431 of file LoopVectorizationLegality.h.
References I.
Referenced by llvm::LoopVectorizationCostModel::getInstructionCost(), and llvm::VPRecipeBuilder::tryToCreateWidenRecipe().
|
inline |
Returns the induction variables found in the loop.
Definition at line 305 of file LoopVectorizationLegality.h.
Referenced by canFoldTailByMasking(), canVectorizeFPMath(), llvm::LoopVectorizationCostModel::collectValuesToIgnore(), llvm::InnerLoopVectorizer::createInductionResumeVPValues(), llvm::LoopVectorizationPlanner::executePlan(), llvm::LoopVectorizationCostModel::expectedCost(), llvm::InnerLoopVectorizer::fixVectorizedLoop(), getIntOrFpInductionDescriptor(), and getPointerInductionDescriptor().
const InductionDescriptor * llvm::LoopVectorizationLegality::getIntOrFpInductionDescriptor | ( | PHINode * | Phi | ) | const |
Returns a pointer to the induction descriptor, if Phi
is an integer or floating point induction.
Definition at line 1337 of file LoopVectorizationLegality.cpp.
References llvm::MapVector< KeyT, ValueT, MapType, VectorType >::find(), getInductionVars(), llvm::InductionDescriptor::IK_FpInduction, llvm::InductionDescriptor::IK_IntInduction, and isInductionPhi().
|
inline |
Definition at line 382 of file LoopVectorizationLegality.h.
Referenced by llvm::LoopVectorizationPlanner::executePlan(), llvm::ARMTTIImpl::preferPredicateOverEpilogue(), llvm::LoopVectorizePass::processLoop(), processLoopInVPlanNativePath(), and llvm::LoopVectorizationCostModel::runtimeChecksRequired().
|
inline |
Definition at line 446 of file LoopVectorizationLegality.h.
Referenced by llvm::AArch64TTIImpl::preferPredicateOverEpilogue(), and llvm::ARMTTIImpl::preferPredicateOverEpilogue().
|
inline |
Definition at line 448 of file LoopVectorizationLegality.h.
Referenced by llvm::ARMTTIImpl::preferPredicateOverEpilogue().
|
inline |
Definition at line 388 of file LoopVectorizationLegality.h.
References llvm::LoopAccessInfo::getDepChecker(), and llvm::MemoryDepChecker::getMaxSafeVectorWidthInBits().
|
inline |
Definition at line 426 of file LoopVectorizationLegality.h.
References llvm::LoopAccessInfo::getNumLoads().
Referenced by llvm::LoopVectorizationCostModel::selectInterleaveCount().
|
inline |
Definition at line 425 of file LoopVectorizationLegality.h.
References llvm::LoopAccessInfo::getNumStores().
Referenced by llvm::LoopVectorizationCostModel::selectInterleaveCount().
const InductionDescriptor * llvm::LoopVectorizationLegality::getPointerInductionDescriptor | ( | PHINode * | Phi | ) | const |
Returns a pointer to the induction descriptor, if Phi
is pointer induction.
Definition at line 1348 of file LoopVectorizationLegality.cpp.
References llvm::MapVector< KeyT, ValueT, MapType, VectorType >::find(), getInductionVars(), llvm::InductionDescriptor::IK_PtrInduction, and isInductionPhi().
|
inline |
Definition at line 442 of file LoopVectorizationLegality.h.
Referenced by llvm::AArch64TTIImpl::preferPredicateOverEpilogue().
|
inline |
Returns the primary induction variable.
Definition at line 299 of file LoopVectorizationLegality.h.
Referenced by llvm::InnerLoopVectorizer::createInductionResumeVPValue(), and llvm::LoopVectorizationCostModel::isOptimizableIVTruncate().
|
inline |
Returns the reduction variables found in the loop.
Definition at line 302 of file LoopVectorizationLegality.h.
Referenced by canFoldTailByMasking(), canVectorizeFPMath(), llvm::LoopVectorizationCostModel::canVectorizeReductions(), llvm::LoopVectorizationCostModel::collectElementTypesForWidening(), llvm::LoopVectorizationCostModel::collectInLoopReductions(), llvm::LoopVectorizationCostModel::collectValuesToIgnore(), llvm::LoopVectorizationCostModel::getInstructionCost(), llvm::LoopVectorizationCostModel::getReductionPatternCost(), llvm::LoopVectorizationCostModel::getSmallestAndWidestTypes(), isInvariantAddressOfReduction(), isInvariantStoreOfReduction(), llvm::AArch64TTIImpl::preferPredicateOverEpilogue(), llvm::LoopVectorizationCostModel::selectInterleaveCount(), and llvm::VPRecipeBuilder::tryToCreateWidenRecipe().
|
inline |
Returns the information that we collected about runtime memory check.
Definition at line 378 of file LoopVectorizationLegality.h.
References llvm::LoopAccessInfo::getRuntimePointerChecking().
Referenced by llvm::LoopVectorizationCostModel::computeMaxVF(), llvm::LoopVectorizationCostModel::runtimeChecksRequired(), and llvm::LoopVectorizationCostModel::selectInterleaveCount().
|
inline |
Definition at line 452 of file LoopVectorizationLegality.h.
References llvm::PredicatedScalarEvolution::getSE().
Referenced by llvm::ARMTTIImpl::preferPredicateOverEpilogue().
|
inline |
Returns the destination of an uncountable early exiting block.
Definition at line 409 of file LoopVectorizationLegality.h.
References assert(), getUncountableExitBlocks(), and llvm::size().
|
inline |
Returns the uncountable early exiting block.
Definition at line 397 of file LoopVectorizationLegality.h.
References assert(), getUncountableExitingBlocks(), and llvm::size().
Referenced by blockNeedsPredication(), and llvm::VPRecipeBuilder::createEdgeMask().
|
inline |
Returns all the exit blocks from uncountable exiting blocks.
Definition at line 468 of file LoopVectorizationLegality.h.
Referenced by getUncountableEarlyExitBlock().
|
inline |
Returns all the exiting blocks with an uncountable exit.
Definition at line 463 of file LoopVectorizationLegality.h.
Referenced by getUncountableEarlyExitingBlock().
|
inline |
Returns the widest induction type.
Definition at line 311 of file LoopVectorizationLegality.h.
Referenced by llvm::LoopVectorizationPlanner::computeBestVF(), llvm::LoopVectorizationPlanner::emitInvalidCostRemarks(), llvm::LoopVectorizationPlanner::executePlan(), and llvm::LoopVectorizationPlanner::selectEpilogueVectorizationFactor().
|
inline |
Returns a list of all known histogram operations in the loop.
Definition at line 440 of file LoopVectorizationLegality.h.
Referenced by llvm::LoopVectorizePass::processLoop().
|
inline |
Returns true if the loop has an uncountable early exit, i.e.
an uncountable exit that isn't the latch block.
Definition at line 394 of file LoopVectorizationLegality.h.
Referenced by blockNeedsPredication(), llvm::LoopVectorizationPlanner::executePlan(), llvm::InnerLoopVectorizer::fixVectorizedLoop(), llvm::LoopVectorizePass::processLoop(), llvm::LoopVectorizationCostModel::requiresScalarEpilogue(), and llvm::LoopVectorizationCostModel::selectInterleaveCount().
|
inline |
Returns true if there is at least one function call in the loop which has a vectorized variant available.
Definition at line 423 of file LoopVectorizationLegality.h.
Returns True if V is a cast that is part of an induction def-use chain, and had been proven to be redundant under a runtime guard (in other words, the cast has the same SCEV expression as the induction phi).
Definition at line 1357 of file LoopVectorizationLegality.cpp.
References llvm::SmallPtrSetImpl< PtrType >::count().
Referenced by isInductionVariable().
Check if this pointer is consecutive when vectorizing.
This happens when the last index of the GEP is the induction variable, or that the pointer itself is an induction variable. This check allows us to vectorize A[idx] into a wide load/store. Returns: 0 - Stride is unknown or non-consecutive. 1 - Address is consecutive. -1 - Address is consecutive, and decreasing. NOTE: This method must only be used before modifying the original scalar loop. Do not use after invoking 'createVectorizedLoopSkeleton' (PR34965).
Definition at line 454 of file LoopVectorizationLegality.cpp.
References llvm::LoopBase< BlockT, LoopT >::getHeader(), llvm::getPtrStride(), llvm::LoopAccessInfo::getSymbolicStrides(), llvm::IRPass, Ptr, and llvm::shouldOptimizeForSize().
Referenced by llvm::LoopVectorizationCostModel::isLegalMaskedLoad(), llvm::LoopVectorizationCostModel::isLegalMaskedStore(), llvm::LoopVectorizationCostModel::memoryInstructionCanBeWidened(), and llvm::LoopVectorizationCostModel::setCostBasedWideningDecision().
Returns True if Phi is a fixed-order recurrence in this loop.
Definition at line 1367 of file LoopVectorizationLegality.cpp.
References llvm::SmallPtrSetImpl< PtrType >::count().
Referenced by llvm::LoopVectorizationCostModel::getInstructionCost(), and llvm::VPRecipeBuilder::tryToCreateWidenRecipe().
Returns True if V is a Phi node of an induction variable in this loop.
Definition at line 1327 of file LoopVectorizationLegality.cpp.
References llvm::MapVector< KeyT, ValueT, MapType, VectorType >::count().
Referenced by getIntOrFpInductionDescriptor(), getPointerInductionDescriptor(), isInductionVariable(), and llvm::LoopVectorizationCostModel::isOptimizableIVTruncate().
Returns True if V can be considered as an induction variable in this loop.
V can be the induction phi, or some redundant cast in the def-use chain of the inducion phi.
Definition at line 1363 of file LoopVectorizationLegality.cpp.
References isCastedInductionVariable(), and isInductionPhi().
Returns true if V
is invariant across all loop iterations according to SCEV.
Definition at line 472 of file LoopVectorizationLegality.cpp.
References llvm::LoopAccessInfo::isInvariant().
Referenced by llvm::LoopVectorizationCostModel::getDivRemSpeculationCost(), llvm::LoopVectorizationCostModel::isPredicatedInst(), isUniform(), and llvm::LoopVectorizationCostModel::shouldConsiderInvariant().
Returns True if given address is invariant and is used to store recurrent expression.
Definition at line 1314 of file LoopVectorizationLegality.cpp.
References llvm::any_of(), llvm::StoreInst::getPointerOperand(), getReductionVars(), llvm::ScalarEvolution::getSCEV(), llvm::PredicatedScalarEvolution::getSE(), llvm::RecurrenceDescriptor::IntermediateStore, and Reduction.
Referenced by llvm::LoopVectorizationCostModel::collectValuesToIgnore().
Returns True if given store is a final invariant store of one of the reductions found in the loop.
Definition at line 1307 of file LoopVectorizationLegality.cpp.
References llvm::any_of(), getReductionVars(), llvm::RecurrenceDescriptor::IntermediateStore, and Reduction.
|
inline |
Returns true if vector representation of the instruction I
requires mask.
Definition at line 417 of file LoopVectorizationLegality.h.
References I.
Referenced by llvm::LoopVectorizationCostModel::getInstructionCost(), llvm::LoopVectorizationCostModel::interleavedAccessCanBeWidened(), llvm::LoopVectorizationCostModel::isPredicatedInst(), and llvm::LoopVectorizationCostModel::setVectorizedCallDecision().
Returns True if PN is a reduction variable in this loop.
Definition at line 343 of file LoopVectorizationLegality.h.
References llvm::MapVector< KeyT, ValueT, MapType, VectorType >::count().
Referenced by llvm::LoopVectorizationCostModel::collectElementTypesForWidening(), and llvm::VPRecipeBuilder::tryToCreateWidenRecipe().
|
inline |
Definition at line 384 of file LoopVectorizationLegality.h.
References llvm::LoopAccessInfo::getDepChecker(), and llvm::MemoryDepChecker::isSafeForAnyVectorWidth().
Referenced by llvm::LoopVectorizationCostModel::selectInterleaveCount().
bool llvm::LoopVectorizationLegality::isUniform | ( | Value * | V, |
ElementCount | VF | ||
) | const |
Returns true if value V is uniform across VF
lanes, when VF
is provided, and otherwise if V
is invariant across all loop iterations.
Definition at line 567 of file LoopVectorizationLegality.cpp.
References llvm::all_of(), llvm::details::FixedOrScalableQuantity< LeafTy, ValueTy >::getKnownMinValue(), llvm::ScalarEvolution::getSCEV(), llvm::PredicatedScalarEvolution::getSE(), I, isInvariant(), llvm::details::FixedOrScalableQuantity< LeafTy, ValueTy >::isScalable(), llvm::ElementCount::isScalar(), llvm::ScalarEvolution::isSCEVable(), and llvm::reverse().
Referenced by isUniformMemOp().
bool llvm::LoopVectorizationLegality::isUniformMemOp | ( | Instruction & | I, |
ElementCount | VF | ||
) | const |
A uniform memory op is a load or store which accesses the same memory location on all VF
lanes, if VF
is provided and otherwise if the memory location is invariant.
Definition at line 600 of file LoopVectorizationLegality.cpp.
References blockNeedsPredication(), llvm::getLoadStorePointerOperand(), I, isUniform(), and Ptr.
Referenced by llvm::LoopVectorizationCostModel::setCostBasedWideningDecision().
void llvm::LoopVectorizationLegality::prepareToFoldTailByMasking | ( | ) |
Mark all respective loads/stores for masking.
Must only be called when tail-folding is possible.
Definition at line 1896 of file LoopVectorizationLegality.cpp.
References assert(), and llvm::LoopBase< BlockT, LoopT >::blocks().
Referenced by llvm::LoopVectorizationPlanner::plan().