llvm::slpvectorizer::BoUpSLP Class Reference

Bottom Up SLP Vectorizer. More...

Classes
struct	EdgeInfo
	This structure holds any data we need about the edges being traversed during buildTree_rec(). More...
class	LookAheadHeuristics
	A helper class used for scoring candidates for two consecutive lanes. More...
class	ShuffleCostEstimator
	Merges shuffle masks and emits final shuffle instruction, if required. More...
class	ShuffleInstructionBuilder
	Merges shuffle masks and emits final shuffle instruction, if required. More...
class	VLOperands
	A helper data structure to hold the operands of a vector of instructions. More...

Public Types
enum class	LoadsState { Gather , Vectorize , ScatterVectorize , StridedVectorize }
	Tracks the state we can represent the loads in the given sequence. More...
using	ValueList = SmallVector< Value *, 4 >
using	ValueSet = SmallPtrSet< Value *, 8 >
using	ExtraValueToDebugLocsMap = MapVector< Value *, SmallVector< Instruction *, 2 > >
using	OrdersType = SmallVector< unsigned, 0 >

Public Member Functions
	BoUpSLP (Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti, TargetLibraryInfo *TLi, AAResults *Aa, LoopInfo *Li, DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB, const DataLayout *DL, OptimizationRemarkEmitter *ORE)
Value *	vectorizeTree ()
	Vectorize the tree that starts with the elements in VL.
Value *	vectorizeTree (const ExtraValueToDebugLocsMap &ExternallyUsedValues, SmallVectorImpl< std::pair< Value *, Value * > > &ReplacedExternals, Instruction *ReductionRoot=nullptr)
	Vectorize the tree but with the list of externally used values ExternallyUsedValues.
InstructionCost	getSpillCost () const
InstructionCost	getTreeCost (ArrayRef< Value * > VectorizedVals=std::nullopt)
void	buildTree (ArrayRef< Value * > Roots, const SmallDenseSet< Value * > &UserIgnoreLst)
	Construct a vectorizable tree that starts at Roots, ignoring users for the purpose of scheduling and extraction in the UserIgnoreLst.
void	buildTree (ArrayRef< Value * > Roots)
	Construct a vectorizable tree that starts at Roots.
bool	doesRootHaveInTreeUses () const
	Returns whether the root node has in-tree uses.
ArrayRef< Value * >	getRootNodeScalars () const
	Return the scalars of the root node.
bool	isSignedMinBitwidthRootNode () const
	Checks if the root graph node can be emitted with narrower bitwidth at codegen and returns it signedness, if so.
void	buildExternalUses (const ExtraValueToDebugLocsMap &ExternallyUsedValues={})
	Builds external uses of the vectorized scalars, i.e.
void	transformNodes ()
	Transforms graph nodes to target specific representations, if profitable.
void	deleteTree ()
	Clear the internal data structures that are created by 'buildTree'.
unsigned	getTreeSize () const
void	optimizeGatherSequence ()
	Perform LICM and CSE on the newly generated gather sequences.
std::optional< OrdersType >	findReusedOrderedScalars (const TreeEntry &TE)
	Checks if the specified gather tree entry TE can be represented as a shuffled vector entry + (possibly) permutation with other gathers.
std::optional< OrdersType >	findPartiallyOrderedLoads (const TreeEntry &TE)
	Sort loads into increasing pointers offsets to allow greater clustering.
std::optional< OrdersType >	getReorderingData (const TreeEntry &TE, bool TopToBottom)
	Gets reordering data for the given tree entry.
void	reorderTopToBottom ()
	Reorders the current graph to the most profitable order starting from the root node to the leaf nodes.
void	reorderBottomToTop (bool IgnoreReorder=false)
	Reorders the current graph to the most profitable order starting from leaves to the root.
unsigned	getVectorElementSize (Value *V)
void	computeMinimumValueSizes ()
	Compute the minimum type sizes required to represent the entries in a vectorizable tree.
unsigned	getMaxVecRegSize () const
unsigned	getMinVecRegSize () const
unsigned	getMinVF (unsigned Sz) const
unsigned	getMaximumVF (unsigned ElemWidth, unsigned Opcode) const
unsigned	canMapToVector (Type *T) const
	Check if homogeneous aggregate is isomorphic to some VectorType.
bool	isTreeTinyAndNotFullyVectorizable (bool ForReduction=false) const
bool	isLoadCombineReductionCandidate (RecurKind RdxKind) const
	Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values can be load combined in the backend.
bool	isLoadCombineCandidate (ArrayRef< Value * > Stores) const
	Assume that a vector of stores of bitwise-or/shifted/zexted loaded values can be load combined in the backend.
LoadsState	canVectorizeLoads (ArrayRef< Value * > VL, const Value *VL0, OrdersType &Order, SmallVectorImpl< Value * > &PointerOps, bool TryRecursiveCheck=true) const
	Checks if the given array of loads can be represented as a vectorized, scatter or just simple gather.
OptimizationRemarkEmitter *	getORE ()
std::optional< int >	findBestRootPair (ArrayRef< std::pair< Value *, Value * > > Candidates, int Limit=LookAheadHeuristics::ScoreFail) const
	Evaluate each pair in Candidates and return index into Candidates for a pair which have highest score deemed to have best chance to form root of profitable tree to vectorize.
bool	isDeleted (Instruction *I) const
	Checks if the instruction is marked for deletion.
void	eraseInstruction (Instruction *I)
	Removes an instruction from its block and eventually deletes it.
template<typename T >
void	removeInstructionsAndOperands (ArrayRef< T * > DeadVals)
	Remove instructions from the parent function and clear the operands of DeadVals instructions, marking for deletion trivially dead operands.
bool	isAnalyzedReductionRoot (Instruction *I) const
	Checks if the instruction was already analyzed for being possible reduction root.
void	analyzedReductionRoot (Instruction *I)
	Register given instruction as already analyzed for being possible reduction root.
bool	areAnalyzedReductionVals (ArrayRef< Value * > VL) const
	Checks if the provided list of reduced values was checked already for vectorization.
void	analyzedReductionVals (ArrayRef< Value * > VL)
	Adds the list of reduced values to list of already checked values for the vectorization.
void	clearReductionData ()
	Clear the list of the analyzed reduction root instructions.
bool	isAnyGathered (const SmallDenseSet< Value * > &Vals) const
	Checks if the given value is gathered in one of the nodes.
bool	isGathered (const Value *V) const
	Checks if the given value is gathered in one of the nodes.
bool	isNotScheduled (const Value *V) const
	Checks if the specified value was not schedule.
bool	isVectorized (Value *V) const
	Check if the value is vectorized in the tree.
	~BoUpSLP ()

Friends
struct	GraphTraits< BoUpSLP * >
struct	DOTGraphTraits< BoUpSLP * >
raw_ostream &	operator<< (raw_ostream &os, const BoUpSLP::ScheduleData &SD)

Detailed Description

Bottom Up SLP Vectorizer.

Definition at line 1231 of file SLPVectorizer.cpp.

Member Typedef Documentation

ExtraValueToDebugLocsMap

using llvm::slpvectorizer::BoUpSLP::ExtraValueToDebugLocsMap = MapVector<Value *, SmallVector<Instruction *, 2> >

Definition at line 1248 of file SLPVectorizer.cpp.

OrdersType

using llvm::slpvectorizer::BoUpSLP::OrdersType = SmallVector<unsigned, 0>

Definition at line 1250 of file SLPVectorizer.cpp.

ValueList

using llvm::slpvectorizer::BoUpSLP::ValueList = SmallVector<Value *, 4>

Definition at line 1246 of file SLPVectorizer.cpp.

ValueSet

using llvm::slpvectorizer::BoUpSLP::ValueSet = SmallPtrSet<Value *, 8>

Definition at line 1247 of file SLPVectorizer.cpp.

Member Enumeration Documentation

LoadsState

enum class llvm::slpvectorizer::BoUpSLP::LoadsState

strong

Tracks the state we can represent the loads in the given sequence.

Enumerator
Gather
Vectorize
ScatterVectorize
StridedVectorize

Definition at line 1239 of file SLPVectorizer.cpp.

Constructor & Destructor Documentation

BoUpSLP()

llvm::slpvectorizer::BoUpSLP::BoUpSLP ( Function *  Func, ScalarEvolution *  Se, TargetTransformInfo *  Tti, TargetLibraryInfo *  TLi, AAResults *  Aa, LoopInfo *  Li, DominatorTree *  Dt, AssumptionCache *  AC, DemandedBits *  DB, const DataLayout *  DL, OptimizationRemarkEmitter *  ORE  )

inline

Definition at line 1252 of file SLPVectorizer.cpp.

References llvm::CodeMetrics::collectEphemeralValues(), F, llvm::details::FixedOrScalableQuantity< LeafTy, ValueTy >::getFixedValue(), llvm::TargetTransformInfo::getMinVectorRegisterBitWidth(), llvm::TargetTransformInfo::getRegisterBitWidth(), MaxVectorRegSizeOption, MinVectorRegSizeOption, and llvm::TargetTransformInfo::RGK_FixedWidthVector.

~BoUpSLP()

BoUpSLP::~BoUpSLP

(

)

Definition at line 4246 of file SLPVectorizer.cpp.

References assert(), llvm::dbgs(), llvm::SmallVectorImpl< T >::emplace_back(), F, I, llvm::RecursivelyDeleteTriviallyDeadInstructions(), llvm::verifyFunction(), and llvm::wouldInstructionBeTriviallyDead().

Member Function Documentation

analyzedReductionRoot()

void llvm::slpvectorizer::BoUpSLP::analyzedReductionRoot ( Instruction *  I )

inline

Register given instruction as already analyzed for being possible reduction root.

Definition at line 2706 of file SLPVectorizer.cpp.

References I.

analyzedReductionVals()

void llvm::slpvectorizer::BoUpSLP::analyzedReductionVals ( ArrayRef< Value * >  VL )

inline

Adds the list of reduced values to list of already checked values for the vectorization.

Definition at line 2716 of file SLPVectorizer.cpp.

References llvm::hash_value(), and llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::insert().

areAnalyzedReductionVals()

bool llvm::slpvectorizer::BoUpSLP::areAnalyzedReductionVals ( ArrayRef< Value * >  VL ) const

inline

Checks if the provided list of reduced values was checked already for vectorization.

Definition at line 2711 of file SLPVectorizer.cpp.

References llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::contains(), and llvm::hash_value().

buildExternalUses()

void BoUpSLP::buildExternalUses

(

const ExtraValueToDebugLocsMap &

ExternallyUsedValues = {}

)

Builds external uses of the vectorized scalars, i.e.

the list of vectorized scalars to be extracted, their lanes and their scalar users. ExternallyUsedValues contains additional list of external uses to handle vectorization of reductions.

Definition at line 6027 of file SLPVectorizer.cpp.

References assert(), llvm::dbgs(), doesInTreeUserNeedToExtract(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::MapVector< KeyT, ValueT, MapType, VectorType >::end(), llvm::MapVector< KeyT, ValueT, MapType, VectorType >::find(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), isDeleted(), LLVM_DEBUG, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace(), and UsesLimit.

buildTree() [1/2]

void BoUpSLP::buildTree

(

ArrayRef< Value * >

Roots

)

Construct a vectorizable tree that starts at Roots.

Definition at line 6262 of file SLPVectorizer.cpp.

References allSameType(), and deleteTree().

buildTree() [2/2]

void BoUpSLP::buildTree	(	ArrayRef< Value * >	Roots,
		const SmallDenseSet< Value * > &	UserIgnoreLst
	)

Construct a vectorizable tree that starts at Roots, ignoring users for the purpose of scheduling and extraction in the UserIgnoreLst.

Definition at line 6253 of file SLPVectorizer.cpp.

References allSameType(), and deleteTree().

canMapToVector()

unsigned BoUpSLP::canMapToVector

(

Type *

T

)

const

Check if homogeneous aggregate is isomorphic to some VectorType.

Accepts homogeneous multidimensional aggregate of scalars/vectors like {[4 x i16], [4 x i16]}, { <2 x float>, <2 x float> }, {{{i16, i16}, {i16, i16}}, {{i16, i16}, {i16, i16}}} and so on.

Returns

number of elements in vector if isomorphism exists, 0 otherwise.

Definition at line 7686 of file SLPVectorizer.cpp.

References llvm::DataLayout::getTypeStoreSizeInBits(), getWidenedType(), isValidElementType(), and N.

canVectorizeLoads()

BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads	(	ArrayRef< Value * >	VL,
		const Value *	VL0,
		OrdersType &	Order,
		SmallVectorImpl< Value * > &	PointerOps,
		bool	TryRecursiveCheck = true
	)		const

Checks if the given array of loads can be represented as a vectorized, scatter or just simple gather.

Parameters

VL	list of loads.
VL0	main load value.
Order	returned order of load instructions.
PointerOps	returned list of pointer operands.
TryRecursiveCheck	used to check if long masked gather can be represented as a serie of loads/insert subvector, if profitable.

Definition at line 4695 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::any_of(), arePointersCompatible(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::back(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::bit_floor(), calculateRtStride(), canVectorizeLoads(), llvm::SmallVectorImpl< T >::clear(), CostKind, llvm::count_if(), DL, llvm::doesNotNeedToBeScheduled(), llvm::SmallVectorBase< Size_T >::empty(), End, llvm::enumerate(), llvm::TargetTransformInfo::forceScalarizeMaskedGather(), llvm::SmallVectorTemplateCommon< T, typename >::front(), llvm::ArrayRef< T >::front(), Gather, GEP, llvm::TargetTransformInfo::getGatherScatterOpCost(), getGEPCosts(), llvm::LoopInfoBase< BlockT, LoopT >::getLoopFor(), getMaximumVF(), llvm::TargetTransformInfo::getMemoryOpCost(), getMinVF(), getParent(), llvm::getPointerOperand(), llvm::getPointersDiff(), getShuffleCost(), llvm::TargetTransformInfo::getStridedMemoryOpCost(), llvm::Value::getType(), getWidenedType(), I, Idx, llvm::SmallSet< T, N, C >::insert(), llvm::TargetTransformInfo::isLegalMaskedGather(), llvm::TargetTransformInfo::isLegalStridedLoadStore(), llvm::isPowerOf2_32(), isReverseOrder(), llvm::TargetTransformInfo::isTypeLegal(), llvm_unreachable, MaxProfitableLoadStride, MinProfitableStridedLoads, P, Ptr, llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::SmallVectorImpl< T >::resize(), ScatterVectorize, llvm::ArrayRef< T >::size(), llvm::SmallSet< T, N, C >::size(), llvm::TargetTransformInfo::SK_InsertSubvector, llvm::ArrayRef< T >::slice(), llvm::sortPtrAccesses(), StridedVectorize, llvm::TargetTransformInfo::TCK_RecipThroughput, Vectorize, and VectorizeNonPowerOf2.

Referenced by canVectorizeLoads().

clearReductionData()

void llvm::slpvectorizer::BoUpSLP::clearReductionData ( )

inline

Clear the list of the analyzed reduction root instructions.

Definition at line 2720 of file SLPVectorizer.cpp.

References llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::clear().

computeMinimumValueSizes()

void BoUpSLP::computeMinimumValueSizes

(

)

Compute the minimum type sizes required to represent the entries in a vectorizable tree.

Definition at line 15930 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::any_of(), assert(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::begin(), llvm::ArrayRef< T >::begin(), llvm::bit_ceil(), llvm::SmallVectorImpl< T >::clear(), llvm::ComputeNumSignBits(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::contains(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::contains(), DL, llvm::slpvectorizer::BoUpSLP::EdgeInfo::EdgeIdx, llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::empty(), llvm::SmallVectorBase< Size_T >::empty(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::ArrayRef< T >::end(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::erase(), F, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::ArrayRef< T >::front(), llvm::IntegerType::get(), llvm::APInt::getBitsSetFrom(), llvm::DemandedBits::getDemandedBits(), llvm::TargetTransformInfo::getNumberOfParts(), getNumElements(), getOpcode(), getRdxKind(), llvm::Type::getScalarType(), llvm::Value::getType(), getWidenedType(), Idx, if(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::insert(), llvm::isGather(), llvm::RecurrenceDescriptor::isIntMinMaxRecurrenceKind(), llvm::isKnownNonNegative(), llvm::MaskedValueIsZero(), llvm::none_of(), llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::size(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace(), llvm::slpvectorizer::BoUpSLP::EdgeInfo::UserTE, and UsesLimit.

deleteTree()

void llvm::slpvectorizer::BoUpSLP::deleteTree ( )

inline

Clear the internal data structures that are created by 'buildTree'.

Definition at line 1338 of file SLPVectorizer.cpp.

References llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::clear(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::clear(), llvm::SetVector< T, Vector, Set, N >::clear(), llvm::SmallPtrSetImplBase::clear(), and llvm::SmallVectorImpl< T >::clear().

Referenced by buildTree().

doesRootHaveInTreeUses()

bool llvm::slpvectorizer::BoUpSLP::doesRootHaveInTreeUses ( ) const

inline

Returns whether the root node has in-tree uses.

Definition at line 1310 of file SLPVectorizer.cpp.

References llvm::SmallVectorBase< Size_T >::empty(), and llvm::SmallVectorTemplateCommon< T, typename >::front().

eraseInstruction()

void llvm::slpvectorizer::BoUpSLP::eraseInstruction ( Instruction *  I )

inline

Removes an instruction from its block and eventually deletes it.

It's like Instruction::eraseFromParent() except that the actual deletion is delayed until BoUpSLP is destructed.

Definition at line 2610 of file SLPVectorizer.cpp.

References I.

Referenced by optimizeGatherSequence(), and vectorizeTree().

findBestRootPair()

std::optional< int > llvm::slpvectorizer::BoUpSLP::findBestRootPair ( ArrayRef< std::pair< Value *, Value * > >  Candidates, int  Limit = LookAheadHeuristics::ScoreFail  ) const

inline

Evaluate each pair in Candidates and return index into Candidates for a pair which have highest score deemed to have best chance to form root of profitable tree to vectorize.

Return std::nullopt if no candidate scored above the LookAheadHeuristics::ScoreFail.

Parameters

Limit

Lower limit of the cost, considered to be good enough score.

Definition at line 2585 of file SLPVectorizer.cpp.

References llvm::slpvectorizer::BoUpSLP::LookAheadHeuristics::getScoreAtLevelRec(), I, and RootLookAheadMaxDepth.

findPartiallyOrderedLoads()

std::optional< BoUpSLP::OrdersType > BoUpSLP::findPartiallyOrderedLoads

(

const TreeEntry &

TE

)

Sort loads into increasing pointers offsets to allow greater clustering.

Definition at line 5050 of file SLPVectorizer.cpp.

References assert(), clusterSortPtrAccesses(), DL, llvm::SmallVectorTemplateBase< T, bool >::push_back(), and llvm::SmallVectorImpl< T >::reserve().

Referenced by getReorderingData().

findReusedOrderedScalars()

std::optional< BoUpSLP::OrdersType > BoUpSLP::findReusedOrderedScalars

(

const TreeEntry &

TE

)

Checks if the specified gather tree entry TE can be represented as a shuffled vector entry + (possibly) permutation with other gathers.

It implements the checks only for possibly ordered scalars (Loads, ExtractElement, ExtractValue), which can be part of the graph.

Definition at line 4346 of file SLPVectorizer.cpp.

References llvm::SmallBitVector::all(), llvm::all_of(), llvm::SmallBitVector::any(), llvm::any_of(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::MutableArrayRef< T >::begin(), llvm::count_if(), llvm::SmallVectorBase< Size_T >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::end(), llvm::MutableArrayRef< T >::end(), llvm::find(), llvm::SmallVectorTemplateCommon< T, typename >::front(), llvm::TargetTransformInfo::getNumberOfParts(), getNumElems(), getPartNumElems(), getWidenedType(), I, Idx, isConstant(), isValidElementType(), llvm::PoisonMaskElem, llvm::SmallBitVector::set(), llvm::SmallVectorBase< Size_T >::size(), llvm::TargetTransformInfo::SK_PermuteSingleSrc, and llvm::SmallBitVector::test().

Referenced by getReorderingData().

getMaximumVF()

unsigned llvm::slpvectorizer::BoUpSLP::getMaximumVF ( unsigned  ElemWidth, unsigned  Opcode  ) const

inline

Definition at line 1427 of file SLPVectorizer.cpp.

References llvm::TargetTransformInfo::getMaximumVF(), and MaxVFOption.

Referenced by canVectorizeLoads().

getMaxVecRegSize()

unsigned llvm::slpvectorizer::BoUpSLP::getMaxVecRegSize ( ) const

inline

Definition at line 1414 of file SLPVectorizer.cpp.

getMinVecRegSize()

unsigned llvm::slpvectorizer::BoUpSLP::getMinVecRegSize ( ) const

inline

Definition at line 1419 of file SLPVectorizer.cpp.

Referenced by getMinVF().

getMinVF()

unsigned llvm::slpvectorizer::BoUpSLP::getMinVF ( unsigned  Sz ) const

inline

Definition at line 1423 of file SLPVectorizer.cpp.

References getMinVecRegSize().

Referenced by canVectorizeLoads().

getORE()

OptimizationRemarkEmitter * llvm::slpvectorizer::BoUpSLP::getORE ( )

inline

Definition at line 1476 of file SLPVectorizer.cpp.

getReorderingData()

std::optional< BoUpSLP::OrdersType > BoUpSLP::getReorderingData	(	const TreeEntry &	TE,
		bool	TopToBottom
	)

Gets reordering data for the given tree entry.

If the entry is vectorized

• just return ReorderIndices, otherwise check if the scalars can be reordered and return the most optimal order.

  Returns

  std::nullopt if ordering is not important, empty order, if identity order is important, or the actual order.

  Parameters

  TopToBottom

  If true, include the order of vectorized stores and insertelement nodes, otherwise skip them.

Definition at line 5122 of file SLPVectorizer.cpp.

References addMask(), llvm::all_of(), allConstant(), allSameType(), llvm::any_of(), areTwoInsertFromSameBuildVector(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::count_if(), llvm::Data, llvm::divideCeil(), E, llvm::SmallVectorBase< Size_T >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::end(), llvm::enumerate(), llvm::find(), llvm::find_if(), findPartiallyOrderedLoads(), findReusedOrderedScalars(), fixupOrderingIndices(), llvm::PoisonValue::get(), getElementIndex(), getExtractIndex(), llvm::TargetTransformInfo::getNumberOfParts(), llvm::Value::getNumUses(), getShuffleCost(), llvm::TargetTransformInfo::getVectorInstrCost(), getWidenedType(), I, Idx, II, llvm::inversePermutation(), llvm::ShuffleVectorInst::isOneUseSingleSourceMask(), isReverseOrder(), isSplat(), llvm::PoisonMaskElem, reorderOrder(), llvm::SmallBitVector::set(), llvm::SmallVectorBase< Size_T >::size(), llvm::TargetTransformInfo::SK_PermuteSingleSrc, llvm::ArrayRef< T >::slice(), llvm::stable_sort(), llvm::TargetTransformInfo::TCK_RecipThroughput, llvm::SmallBitVector::test(), llvm::transform(), and llvm::Value::user_begin().

Referenced by reorderBottomToTop(), and reorderTopToBottom().

getRootNodeScalars()

ArrayRef< Value * > llvm::slpvectorizer::BoUpSLP::getRootNodeScalars ( ) const

inline

Return the scalars of the root node.

Definition at line 1316 of file SLPVectorizer.cpp.

References assert(), llvm::SmallVectorBase< Size_T >::empty(), and llvm::SmallVectorTemplateCommon< T, typename >::front().

getSpillCost()

InstructionCost BoUpSLP::getSpillCost

(

)

const

Returns

the cost incurred by unwanted spills and fills, caused by holding live values over call sites.

Definition at line 10461 of file SLPVectorizer.cpp.

References A, assert(), B, llvm::dbgs(), llvm::SmallPtrSetImpl< PtrType >::erase(), llvm::TargetTransformInfo::getCallInstrCost(), llvm::TargetTransformInfo::getCostOfKeepingLiveOverCall(), llvm::TargetTransformInfo::getIntrinsicInstrCost(), llvm::ilist_node_impl< OptionsT >::getIterator(), llvm::DominatorTreeBase< NodeT, IsPostDom >::getNode(), llvm::ilist_detail::node_parent_access< NodeTy, ParentTy >::getParent(), getWidenedType(), I, II, llvm::SmallPtrSetImpl< PtrType >::insert(), LLVM_DEBUG, llvm::User::operands(), llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::SmallPtrSetImplBase::size(), llvm::sort(), llvm::TargetTransformInfo::TCK_RecipThroughput, and X.

getTreeCost()

InstructionCost BoUpSLP::getTreeCost

(

ArrayRef< Value * >

VectorizedVals = std::nullopt

)

Returns

the vectorization cost of the subtree that starts at VL. A negative number means that this is profitable.

Definition at line 10732 of file SLPVectorizer.cpp.

References llvm::all_of(), areTwoInsertFromSameBuildVector(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::back(), llvm::sampleprof::Base, llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::CallingConv::C, CostKind, llvm::SmallPtrSetImpl< PtrType >::count(), llvm::count_if(), llvm::dbgs(), DL, llvm::SmallVectorImpl< T >::emplace_back(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::SmallVectorTemplateCommon< T, typename >::end(), llvm::enumerate(), F, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::find_if(), llvm::for_each(), llvm::IntegerType::get(), llvm::TargetTransformInfo::getCastInstrCost(), getElementIndex(), llvm::TargetTransformInfo::getExtractWithExtendCost(), llvm::TargetTransformInfo::getInstructionCost(), llvm::User::getOperand(), llvm::InsertElementInst::getType(), llvm::TargetTransformInfo::getVectorInstrCost(), getWidenedType(), llvm::APInt::getZero(), I, Idx, II, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::insert(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::insert(), llvm::SmallPtrSetImpl< PtrType >::insert(), isFirstInsertElement(), llvm::isPowerOf2_32(), LLVM_DEBUG, llvm::TargetTransformInfo::None, P, llvm::PoisonMaskElem, llvm::SmallVectorTemplateBase< T, bool >::push_back(), shortBundleName(), llvm::SmallVectorBase< Size_T >::size(), llvm::TargetTransformInfo::TCC_Basic, llvm::TargetTransformInfo::TCC_Free, and llvm::TargetTransformInfo::TCK_RecipThroughput.

getTreeSize()

unsigned llvm::slpvectorizer::BoUpSLP::getTreeSize ( ) const

inline

Definition at line 1361 of file SLPVectorizer.cpp.

References llvm::SmallVectorBase< Size_T >::size().

getVectorElementSize()

unsigned BoUpSLP::getVectorElementSize

(

Value *

V

)

Returns

The vector element size in bits to use when vectorizing the expression tree ending at V. If V is a store, the size is the width of the stored value. Otherwise, the size is the width of the largest loaded value reaching V. This method is used by the vectorizer to calculate vectorization factors.

Definition at line 15490 of file SLPVectorizer.cpp.

References DL, llvm::SmallVectorImpl< T >::emplace_back(), llvm::SmallVectorBase< Size_T >::empty(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::IRBuilderBase::getInt1Ty(), getVectorElementSize(), I, llvm::SmallPtrSetImpl< PtrType >::insert(), llvm::isa(), llvm::SmallVectorImpl< T >::pop_back_val(), and RecursionMaxDepth.

Referenced by getVectorElementSize().

isAnalyzedReductionRoot()

bool llvm::slpvectorizer::BoUpSLP::isAnalyzedReductionRoot ( Instruction *  I ) const

inline

Checks if the instruction was already analyzed for being possible reduction root.

Definition at line 2701 of file SLPVectorizer.cpp.

References I.

isAnyGathered()

bool llvm::slpvectorizer::BoUpSLP::isAnyGathered ( const SmallDenseSet< Value * > &  Vals ) const

inline

Checks if the given value is gathered in one of the nodes.

Definition at line 2726 of file SLPVectorizer.cpp.

References llvm::any_of(), and llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::contains().

isDeleted()

bool llvm::slpvectorizer::BoUpSLP::isDeleted ( Instruction *  I ) const

inline

Checks if the instruction is marked for deletion.

Definition at line 2605 of file SLPVectorizer.cpp.

References I.

Referenced by buildExternalUses(), and optimizeGatherSequence().

isGathered()

bool llvm::slpvectorizer::BoUpSLP::isGathered ( const Value *  V ) const

inline

Checks if the given value is gathered in one of the nodes.

Definition at line 2730 of file SLPVectorizer.cpp.

References llvm::SmallPtrSetImpl< PtrType >::contains().

isLoadCombineCandidate()

bool BoUpSLP::isLoadCombineCandidate

(

ArrayRef< Value * >

Stores

)

const

Assume that a vector of stores of bitwise-or/shifted/zexted loaded values can be load combined in the backend.

Load combining may not be allowed in the IR optimizer, so we do not want to alter the pattern. For example, partially transforming a scalar bswap() pattern into vector code is effectively impossible for the backend to undo. TODO: If load combining is allowed in the IR optimizer, this analysis may not be necessary.

Definition at line 10385 of file SLPVectorizer.cpp.

References isLoadCombineCandidateImpl(), llvm::PatternMatch::m_Store(), llvm::PatternMatch::m_Value(), llvm::PatternMatch::match(), llvm::ArrayRef< T >::size(), and X.

isLoadCombineReductionCandidate()

bool BoUpSLP::isLoadCombineReductionCandidate

(

RecurKind

RdxKind

)

const

Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values can be load combined in the backend.

Load combining may not be allowed in the IR optimizer, so we do not want to alter the pattern. For example, partially transforming a scalar bswap() pattern into vector code is effectively impossible for the backend to undo. TODO: If load combining is allowed in the IR optimizer, this analysis may not be necessary.

Definition at line 10375 of file SLPVectorizer.cpp.

References isLoadCombineCandidateImpl(), and llvm::Or.

isNotScheduled()

bool llvm::slpvectorizer::BoUpSLP::isNotScheduled ( const Value *  V ) const

inline

Checks if the specified value was not schedule.

Definition at line 2734 of file SLPVectorizer.cpp.

References llvm::SmallPtrSetImpl< PtrType >::contains().

isSignedMinBitwidthRootNode()

bool llvm::slpvectorizer::BoUpSLP::isSignedMinBitwidthRootNode ( ) const

inline

Checks if the root graph node can be emitted with narrower bitwidth at codegen and returns it signedness, if so.

Definition at line 1323 of file SLPVectorizer.cpp.

References llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::at(), and llvm::SmallVectorTemplateCommon< T, typename >::front().

isTreeTinyAndNotFullyVectorizable()

bool BoUpSLP::isTreeTinyAndNotFullyVectorizable

(

bool

ForReduction = false

)

const

Returns

True if the VectorizableTree is both tiny and not fully vectorizable. We do not vectorize such trees.

Definition at line 10398 of file SLPVectorizer.cpp.

References llvm::all_of(), allConstant(), allSameBlock(), llvm::any_of(), assert(), llvm::count_if(), isSplat(), MinTreeSize, and SLPCostThreshold.

isVectorized()

bool llvm::slpvectorizer::BoUpSLP::isVectorized ( Value *  V ) const

inline

Check if the value is vectorized in the tree.

Definition at line 2739 of file SLPVectorizer.cpp.

optimizeGatherSequence()

void BoUpSLP::optimizeGatherSequence

(

)

Perform LICM and CSE on the newly generated gather sequences.

Definition at line 14752 of file SLPVectorizer.cpp.

References A, llvm::any_of(), assert(), llvm::SmallVectorImpl< T >::assign(), B, llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::dbgs(), llvm::DominatorTree::dominates(), llvm::SmallVectorBase< Size_T >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::end(), eraseInstruction(), llvm::DominatorTreeBase< NodeT, IsPostDom >::getNode(), llvm::TargetTransformInfo::getNumberOfParts(), llvm::BasicBlock::getTerminator(), getWidenedType(), I, llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::insert(), llvm::is_contained(), isDeleted(), llvm::DominatorTree::isReachableFromEntry(), LLVM_DEBUG, llvm::make_early_inc_range(), N, llvm::PoisonMaskElem, llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::SmallVectorImpl< T >::reserve(), llvm::ArrayRef< T >::size(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::size(), llvm::SetVector< T, Vector, Set, N >::size(), llvm::SmallVectorBase< Size_T >::size(), and llvm::sort().

removeInstructionsAndOperands()

template<typename T >

void llvm::slpvectorizer::BoUpSLP::removeInstructionsAndOperands ( ArrayRef< T * >  DeadVals )

inline

Remove instructions from the parent function and clear the operands of DeadVals instructions, marking for deletion trivially dead operands.

Definition at line 2617 of file SLPVectorizer.cpp.

References llvm::all_of(), assert(), llvm::SmallVectorBase< Size_T >::empty(), llvm::ScalarEvolution::forgetValue(), I, llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::insert(), llvm::isInstructionTriviallyDead(), llvm::none_of(), llvm::SmallVectorImpl< T >::pop_back_val(), llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::salvageDebugInfo(), llvm::Value::use_empty(), and llvm::wouldInstructionBeTriviallyDead().

reorderBottomToTop()

void BoUpSLP::reorderBottomToTop

(

bool

IgnoreReorder = false

)

Reorders the current graph to the most profitable order starting from leaves to the root.

It allows to rotate small subgraphs and reduce the number of reshuffles if the leaf nodes use the same order. In this case we can merge the orders and just shuffle user node instead of shuffling its operands. Plus, even the leaf nodes have different orders, it allows to sink reordering in the graph closer to the root node and merge it later during analysis.

Definition at line 5747 of file SLPVectorizer.cpp.

References llvm::all_of(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::SmallPtrSetImplBase::clear(), combineOrders(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::contains(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::count(), llvm::count_if(), llvm::Data, llvm::drop_begin(), E, llvm::slpvectorizer::BoUpSLP::EdgeInfo::EdgeIdx, llvm::SetVector< T, Vector, Set, N >::empty(), fixupOrderingIndices(), Gather, getReorderingData(), I, Idx, llvm::SetVector< T, Vector, Set, N >::insert(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::insert(), llvm::SmallPtrSetImpl< PtrType >::insert(), llvm::inversePermutation(), P, llvm::PoisonMaskElem, llvm::SmallVectorTemplateBase< T, bool >::push_back(), reorderOrder(), reorderReuses(), llvm::reorderScalars(), llvm::ArrayRef< T >::size(), llvm::sort(), llvm::transform(), Users, and llvm::slpvectorizer::BoUpSLP::EdgeInfo::UserTE.

reorderTopToBottom()

void BoUpSLP::reorderTopToBottom

(

)

Reorders the current graph to the most profitable order starting from the root node to the leaf nodes.

The best order is chosen only from the nodes of the same size (vectorization factor). Smaller nodes are considered parts of subgraph with smaller VF and they are reordered independently. We can make it because we still need to extend smaller nodes to the wider VF and we can merge reordering shuffles with the widening shuffles.

Definition at line 5430 of file SLPVectorizer.cpp.

References addMask(), llvm::all_of(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), combineOrders(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::count(), E, llvm::ArrayRef< T >::empty(), llvm::SmallVectorBase< Size_T >::empty(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), fixupOrderingIndices(), llvm::for_each(), getAltInstrMask(), getReorderingData(), getWidenedType(), I, Idx, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::insert(), llvm::inversePermutation(), llvm::isa(), llvm::PoisonMaskElem, RecursionMaxDepth, reorderOrder(), llvm::reorderScalars(), llvm::ArrayRef< T >::size(), llvm::transform(), and llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace().

transformNodes()

void BoUpSLP::transformNodes

(

)

Transforms graph nodes to target specific representations, if profitable.

Definition at line 8360 of file SLPVectorizer.cpp.

References llvm::canConvertToMinOrMaxIntrinsic(), CostKind, llvm::TargetTransformInfo::getMemoryOpCost(), llvm::Type::getPointerAddressSpace(), getShuffleCost(), llvm::TargetTransformInfo::getStridedMemoryOpCost(), getWidenedType(), llvm::inversePermutation(), llvm::TargetTransformInfo::isLegalStridedLoadStore(), isReverseOrder(), llvm::Intrinsic::not_intrinsic, llvm::TargetTransformInfo::SK_Reverse, and llvm::TargetTransformInfo::TCK_RecipThroughput.

vectorizeTree() [1/2]

Value * BoUpSLP::vectorizeTree

(

)

Vectorize the tree that starts with the elements in VL.

Returns the vectorized root.

Definition at line 14097 of file SLPVectorizer.cpp.

References vectorizeTree().

Referenced by vectorizeTree().

vectorizeTree() [2/2]

Value * BoUpSLP::vectorizeTree	(	const ExtraValueToDebugLocsMap &	ExternallyUsedValues,
		SmallVectorImpl< std::pair< Value *, Value * > > &	ReplacedExternals,
		Instruction *	ReductionRoot = nullptr
	)

Vectorize the tree but with the list of externally used values ExternallyUsedValues.

Values in this MapVector can be replaced but the generated extractvalue instructions.

Parameters

ReplacedExternals

containd list of replaced external values {scalar, replace} after emitting extractelement for external uses.

Definition at line 14114 of file SLPVectorizer.cpp.

References llvm::slpvectorizer::BoUpSLP::ShuffleInstructionBuilder::add(), llvm::any_of(), areTwoInsertFromSameBuildVector(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::back(), llvm::ArrayRef< T >::back(), llvm::sampleprof::Base, llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::clear(), llvm::SmallPtrSetImpl< PtrType >::contains(), llvm::MapVector< KeyT, ValueT, MapType, VectorType >::count(), llvm::IRBuilderBase::CreateExtractElement(), llvm::IRBuilderBase::CreateExtractVector(), llvm::IRBuilderBase::CreateIntCast(), llvm::Data, llvm::dbgs(), DL, llvm::SmallVectorImpl< T >::emplace_back(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::SmallVectorTemplateCommon< T, typename >::end(), llvm::BasicBlock::end(), eraseInstruction(), F, llvm::slpvectorizer::BoUpSLP::ShuffleInstructionBuilder::finalize(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::find_if(), llvm::SmallVectorTemplateCommon< T, typename >::front(), llvm::ArrayRef< T >::front(), llvm::get(), llvm::FixedVectorType::get(), llvm::PoisonValue::get(), llvm::SetVector< T, Vector, Set, N >::getArrayRef(), getElementIndex(), llvm::IRBuilderBase::GetInsertBlock(), llvm::IRBuilderBase::GetInsertPoint(), llvm::IRBuilderBase::getInt32(), llvm::IRBuilderBase::getInt64(), llvm::ilist_node_impl< OptionsT >::getIterator(), llvm::FixedVectorType::getNumElements(), llvm::User::getOperand(), llvm::ilist_detail::node_parent_access< NodeTy, ParentTy >::getParent(), llvm::Type::getScalarType(), llvm::InsertElementInst::getType(), llvm::Value::getType(), getWidenedType(), I, Idx, II, llvm::SetVector< T, Vector, Set, N >::insert(), llvm::detail::DenseSetImpl< ValueT, MapTy, ValueInfoT >::insert(), llvm::is_contained(), isFirstInsertElement(), llvm::Type::isIntOrIntVectorTy(), llvm::isKnownNonNegative(), IV, LLVM_DEBUG, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::lookup(), llvm::mayHaveNonDefUseDependency(), llvm::Instruction::moveAfter(), PHI, llvm::PoisonMaskElem, llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::SmallVectorTemplateCommon< T, typename >::rbegin(), llvm::SmallVectorTemplateCommon< T, typename >::rend(), llvm::Value::replaceAllUsesWith(), llvm::User::replaceUsesOfWith(), llvm::reverse(), llvm::IRBuilderBase::SetCurrentDebugLocation(), llvm::IRBuilderBase::SetInsertPoint(), llvm::SmallVectorBase< Size_T >::size(), SLPReVec, llvm::sort(), llvm::Value::takeName(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace(), llvm::User::User(), UsesLimit, llvm::Vector, and vectorizeTree().

Friends And Related Function Documentation

DOTGraphTraits< BoUpSLP * >

friend struct DOTGraphTraits< BoUpSLP * >

friend

Definition at line 3750 of file SLPVectorizer.cpp.

GraphTraits< BoUpSLP * >

friend struct GraphTraits< BoUpSLP * >

friend

Definition at line 3750 of file SLPVectorizer.cpp.

operator<<

raw_ostream & operator<< ( raw_ostream &  os, const BoUpSLP::ScheduleData &  SD  )

friend

Definition at line 3750 of file SLPVectorizer.cpp.

---

The documentation for this class was generated from the following file:

• lib/Transforms/Vectorize/SLPVectorizer.cpp