Bottom Up SLP Vectorizer. More...

Classes
struct	EdgeInfo
	This structure holds any data we need about the edges being traversed during buildTreeRec(). 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...
struct	StridedPtrInfo
	If we decide to generate strided load / store, this struct contains all the necessary info. 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 , CompressVectorize }
	Tracks the state we can represent the loads in the given sequence. More...
using	ValueList = SmallVector<Value *, 8>
using	InstrList = SmallVector<Instruction *, 16>
using	ValueSet = SmallPtrSet<Value *, 16>
using	StoreList = SmallVector<StoreInst *, 8>
using	ExtraValueToDebugLocsMap = SmallDenseSet<Value *, 4>
using	OrdersType = SmallVector<unsigned, 4>

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, Instruction *ReductionRoot=nullptr, ArrayRef< std::tuple< WeakTrackingVH, unsigned, bool, bool > > VectorValuesAndScales={})
	Vectorize the tree but with the list of externally used values `ExternallyUsedValues`.
InstructionCost	getSpillCost ()
InstructionCost	calculateTreeCostAndTrimNonProfitable (ArrayRef< Value * > VectorizedVals={}, Instruction *RdxRoot=nullptr)
	Calculates the cost of the subtrees, trims non-profitable ones and returns final cost.
InstructionCost	getTreeCost (InstructionCost TreeCost, ArrayRef< Value * > VectorizedVals={}, InstructionCost ReductionCost=TTI::TCC_Free, Instruction *RdxRoot=nullptr)
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	hasRuntimeCheckableBlockers () const
	Returns true if the last buildTree() observed a may-alias memory dependency between two distinct, range-checkable base objects, i.e.
void	setHasRuntimeCheckableBlockers (bool V)
	Records whether a may-alias dependency between distinct, range-checkable base objects has been observed, so the caller can decide to retry with runtime alias checks enabled.
bool	hasNonCheckableMemBlocker () const
	Returns true if the last buildTree() kept a may-alias memory dependency that is not runtime-checkable (call or a non-simple mem access).
void	setHasNonCheckableMemBlocker (bool V)
	Records that a non-runtime-checkable may-alias dependency was kept.
bool	isTryingRuntimeAliasChecks () const
	Returns true if the current vectorization attempt may drop runtime-checkable may-alias dependencies and guard the region with runtime alias checks.
void	setTryRuntimeAliasChecks (bool V)
	Enables or disables dropping runtime-checkable may-alias dependencies in favor of runtime alias checks for the current vectorization attempt.
void	resetRuntimeAliasCheckState ()
	Resets the runtime alias check data.
void	captureRuntimeCheckBodySnapshot ()
	Snapshots RTChecks.BB's body (non-PHI, non-terminator) into RTOrigBodyOrder in program order, for the scalar fallback.
bool	canVersionBlockForRuntimeChecks (BasicBlock *BB) const
	Returns true if `BB` satisfies the block-level preconditions for runtime alias check versioning (straight-line, outside any loop, duplicable, not a scalar fallback, function not optimized for size).
bool	canVersionForRuntimeChecks ()
	Returns true if the runtime alias checks can be safely emitted to guard the vectorized region.
bool	isScalarFallbackBlock (BasicBlock *BB) const
	Returns true if `BB` is a scalar fallback block created by runtime alias check versioning.
bool	runtimeChecksFailedForBlock (BasicBlock *BB) const
	Returns true if an optimistic runtime-checks versioning attempt already failed for `BB`, so further retries in the same block can be skipped.
void	markRuntimeChecksFailedForBlock (BasicBlock *BB)
	Records that an optimistic runtime-checks versioning attempt failed for `BB`.
InstructionCost	getRuntimeChecksCost () const
	Returns the modeled cost of the runtime alias checks collected during the last (optimistic) buildTree().
bool	hasRuntimeAliasChecks () const
	Returns true if the last (optimistic) buildTree() collected any runtime alias checks that must guard the vectorized region.
bool	isCFGChanged () const
	Returns true if vectorization changed the CFG (i.e.
ArrayRef< Value * >	getRootNodeScalars () const
	Returns the scalars of the root node.
unsigned	findRootLaneForValue (Value *V) const
	Returns the lane the given value is vectorized to in the root node.
std::optional< std::pair< Type *, bool > >	getRootNodeTypeWithNoCast () const
	Returns the type/is-signed info for the root node in the graph without casting.
bool	isSignedMinBitwidthRootNode () const
	Checks if the root graph node can be emitted with narrower bitwidth at codegen and returns it signedness, if so.
FixedVectorType *	getReductionType () const
	Returns reduction type after minbitdth analysis.
bool	isReducedBitcastRoot () const
	Returns true if the tree results in one of the reduced bitcasts variants.
bool	isReducedCmpBitcastRoot () const
	Returns true if the tree results in the reduced cmp bitcast root.
bool	isReductionTree () const
	Returns true if the tree is a reduction tree.
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
unsigned	getCanonicalGraphSize () const
	Returns the base graph size, before any transformations.
void	optimizeGatherSequence ()
	Perform LICM and CSE on the newly generated gather sequences.
std::optional< OrdersType >	findReusedOrderedScalars (const TreeEntry &TE, bool TopToBottom, bool IgnoreReorder)
	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, bool IgnoreReorder)
	Gets reordering data for the given tree entry.
bool	isProfitableToReorder () const
	Checks if it is profitable to reorder the current tree.
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	canVectorStoreInsertValue (const TreeEntry *E) const
FixedVectorType *	getInsertBuildVectorSrcTy (const TreeEntry *E) const
bool	isTreeTinyAndNotFullyVectorizable (bool ForReduction=false) const
bool	isTreeNotExtendable () const
	Checks if the graph and all its subgraphs cannot be better vectorized.
bool	isStridedLoad (ArrayRef< Value * > PointerOps, Type *ScalarTy, Align Alignment, const int64_t Diff, const size_t Sz) const
	Checks if strided loads can be generated out of `VL` loads with pointers `PointerOps:`
bool	analyzeConstantStrideCandidate (const ArrayRef< Value * > PointerOps, Type ElemTy, Align Alignment, const SmallVectorImpl< unsigned > &SortedIndices, const int64_t Diff, Value Ptr0, StridedPtrInfo &SPtrInfo) const
	Return true if an array of scalar loads can be replaced with a strided load (with constant stride).
bool	analyzeRtStrideCandidate (ArrayRef< Value * > PointerOps, Type *ScalarTy, Align CommonAlignment, SmallVectorImpl< unsigned > &SortedIndices, StridedPtrInfo &SPtrInfo, bool IsLoad) const
	Return true if an array of scalar loads can be replaced with a strided load (with run-time stride).
LoadsState	canVectorizeLoads (ArrayRef< Value * > VL, const Value VL0, SmallVectorImpl< unsigned > &Order, SmallVectorImpl< Value > &PointerOps, StridedPtrInfo &SPtrInfo, unsigned *BestVF=nullptr, bool TryRecursiveCheck=true) const
	Checks if the given array of loads can be represented as a vectorized, scatter or just simple gather.
bool	hasSameNode (const InstructionsState &S, ArrayRef< Value * > VL) const
	Checks whether some existing tree entry has scalars equal to `VL`.
template<typename T>
void	registerNonVectorizableLoads (ArrayRef< T * > VL)
	Registers non-vectorizable sequence of loads.
template<typename T>
bool	areKnownNonVectorizableLoads (ArrayRef< T * > VL) const
	Checks if the given loads sequence is known as not vectorizable.
OptimizationRemarkEmitter *	getORE ()
std::pair< std::optional< int >, 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, ArrayRef< std::tuple< WeakTrackingVH, unsigned, bool, bool > > VectorValuesAndScales)
	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 (const Value *V) const
	Check if the value is vectorized in the tree.
bool	canBuildSplitNode (ArrayRef< Value * > VL, const InstructionsState &LocalState, SmallVectorImpl< Value * > &Op1, SmallVectorImpl< Value * > &Op2, OrdersType &ReorderIndices) const
	Checks if it is legal and profitable to build SplitVectorize node for the given `VL`.
	~BoUpSLP ()
template<typename BVTy, typename ResTy, typename... Args>
ResTy	processBuildVector (const TreeEntry E, Type ScalarTy, Args &...Params)

Static Public Member Functions
static bool	isIdentityOrder (ArrayRef< unsigned > Order)
	Does this non-empty order represent an identity order?

Friends
struct	DenseMapInfo< EdgeInfo >
struct	GraphTraits< BoUpSLP * >
struct	DOTGraphTraits< BoUpSLP * >
raw_ostream &	operator<< (raw_ostream &OS, const BoUpSLP::ScheduleEntity &SE)
raw_ostream &	operator<< (raw_ostream &OS, const BoUpSLP::ScheduleData &SD)
raw_ostream &	operator<< (raw_ostream &OS, const BoUpSLP::ScheduleBundle &Bundle)
raw_ostream &	operator<< (raw_ostream &OS, const BoUpSLP::ScheduleCopyableData &SD)

Detailed Description

Bottom Up SLP Vectorizer.

Definition at line 1465 of file SLPVectorizer.cpp.

Member Typedef Documentation

◆ ExtraValueToDebugLocsMap

using slpvectorizer::BoUpSLP::ExtraValueToDebugLocsMap = SmallDenseSet<Value *, 4>

Definition at line 1500 of file SLPVectorizer.cpp.

◆ InstrList

using slpvectorizer::BoUpSLP::InstrList = SmallVector<Instruction *, 16>

Definition at line 1497 of file SLPVectorizer.cpp.

◆ OrdersType

using slpvectorizer::BoUpSLP::OrdersType = SmallVector<unsigned, 4>

Definition at line 1501 of file SLPVectorizer.cpp.

◆ StoreList

using slpvectorizer::BoUpSLP::StoreList = SmallVector<StoreInst *, 8>

Definition at line 1499 of file SLPVectorizer.cpp.

◆ ValueList

using slpvectorizer::BoUpSLP::ValueList = SmallVector<Value *, 8>

Definition at line 1496 of file SLPVectorizer.cpp.

◆ ValueSet

using slpvectorizer::BoUpSLP::ValueSet = SmallPtrSet<Value *, 16>

Definition at line 1498 of file SLPVectorizer.cpp.

Member Enumeration Documentation

◆ LoadsState

enum class slpvectorizer::BoUpSLP::LoadsState

strong

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

Enumerator
Gather
Vectorize
ScatterVectorize
StridedVectorize
CompressVectorize

Definition at line 1488 of file SLPVectorizer.cpp.

Constructor & Destructor Documentation

◆ BoUpSLP()

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 1503 of file SLPVectorizer.cpp.

References llvm::CodeMetrics::collectEphemeralValues(), getContext(), MaxVectorRegSizeOption, MinVectorRegSizeOption, and llvm::TargetTransformInfo::RGK_FixedWidthVector.

Referenced by slpvectorizer::BoUpSLP::LookAheadHeuristics::LookAheadHeuristics(), slpvectorizer::BoUpSLP::ShuffleCostEstimator::ShuffleCostEstimator(), slpvectorizer::BoUpSLP::ShuffleInstructionBuilder::ShuffleInstructionBuilder(), and slpvectorizer::BoUpSLP::VLOperands::VLOperands().

◆ ~BoUpSLP()

BoUpSLP::~BoUpSLP ( )

Definition at line 6245 of file SLPVectorizer.cpp.

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

Member Function Documentation

◆ analyzeConstantStrideCandidate()

bool BoUpSLP::analyzeConstantStrideCandidate	(	const ArrayRef< Value * >	PointerOps,
		Type *	ElemTy,
		Align	Alignment,
		const SmallVectorImpl< unsigned > &	SortedIndices,
		const int64_t	Diff,
		Value *	Ptr0,
		StridedPtrInfo &	SPtrInfo ) const

Return true if an array of scalar loads can be replaced with a strided load (with constant stride).

It is possible that the load gets "widened". Suppose that originally each load loads k bytes and PointerOps can be arranged as follows (s is constant): b + 0 * s + 0 b + 0 * s + 1 b + 0 * s + 2 ... b + 0 * s + (w - 1)

b + 1 * s + 0 b + 1 * s + 1 b + 1 * s + 2 ... b + 1 * s + (w - 1) ...

b + (n - 1) * s + 0 b + (n - 1) * s + 1 b + (n - 1) * s + 2 ... b + (n - 1) * s + (w - 1)

In this case we will generate a strided load of type <n x (k * w)>.

Parameters

PointerOps	list of pointer arguments of loads.
ElemTy	original scalar type of loads.
Alignment	alignment of the first load.
SortedIndices	is the order of PointerOps as returned by sortPtrAccesses
Diff	Pointer difference between the lowest and the highes pointer in PointerOps as returned by getPointersDiff.
Ptr0	first pointer in PointersOps.
PtrN	last pointer in PointersOps.
SPtrInfo	If the function return true, it also sets all the fields of SPtrInfo necessary to generate the strided load later.

Definition at line 7084 of file SLPVectorizer.cpp.

References assert(), llvm::cast(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::find_if(), llvm::Type::getIntNTy(), llvm::getPointersDiff(), llvm::ConstantInt::getSigned(), llvm::Value::getType(), getWidenedType(), I, isStridedLoad(), llvm::Type::isVectorTy(), llvm::Offset, llvm::seq(), llvm::ArrayRef< T >::size(), slpvectorizer::BoUpSLP::StridedPtrInfo::StrideVal, and slpvectorizer::BoUpSLP::StridedPtrInfo::Ty.

Referenced by canVectorizeLoads().

◆ analyzedReductionRoot()

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

inline

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

Definition at line 3349 of file SLPVectorizer.cpp.

References I.

◆ analyzedReductionVals()

void 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 3359 of file SLPVectorizer.cpp.

References llvm::hash_value().

◆ analyzeRtStrideCandidate()

bool BoUpSLP::analyzeRtStrideCandidate	(	ArrayRef< Value * >	PointerOps,
		Type *	ScalarTy,
		Align	CommonAlignment,
		SmallVectorImpl< unsigned > &	SortedIndices,
		StridedPtrInfo &	SPtrInfo,
		bool	IsLoad ) const

Return true if an array of scalar loads can be replaced with a strided load (with run-time stride).

Parameters

PointerOps	list of pointer arguments of loads.
ScalarTy	type of loads.
CommonAlignment	common alignement of loads as computed by computeCommonAlignment<LoadInst>.
SortedIndicies	is a list of indicies computed by this function such that the sequence PointerOps[SortedIndices[0]], / PointerOps[SortedIndicies[1]], ..., PointerOps[SortedIndices[n]] is ordered by the coefficient of the stride. For example, if PointerOps is base + stride, base, base + 2 * stride the SortedIndices will be [1, 0, 2]. We follow the convention that if SortedIndices has to be 0, 1, 2, 3, ... we return empty vector for SortedIndicies.
SPtrInfo	If the function return true, it also sets all the fields of SPtrInfo necessary to generate the strided load later.
IsLoad	Is this a strided load (true) or strided store (false)

Definition at line 7184 of file SLPVectorizer.cpp.

References llvm::Add, llvm::SmallVectorTemplateCommon< T, typename >::begin(), calculateRtStride(), llvm::cast(), llvm::SmallVectorImpl< T >::clear(), llvm::dyn_cast(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::end(), llvm::enumerate(), llvm::Type::getIntNTy(), llvm::slpvectorizer::getNumElements(), getWidenedType(), I, llvm::detail::DenseSetImpl< ValueT, MapTy >::insert(), llvm::Type::isVectorTy(), MinProfitableStridedLoads, MinProfitableStridedStores, llvm::Offset, llvm::SmallVectorImpl< T >::resize(), llvm::seq(), llvm::ArrayRef< T >::size(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::size(), llvm::detail::DenseSetImpl< ValueT, MapTy >::size(), llvm::SmallVectorTemplateCommon< T, typename >::size(), llvm::sort(), slpvectorizer::BoUpSLP::StridedPtrInfo::StrideSCEV, and slpvectorizer::BoUpSLP::StridedPtrInfo::Ty.

Referenced by canVectorizeLoads().

◆ areAnalyzedReductionVals()

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

inline

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

Definition at line 3354 of file SLPVectorizer.cpp.

References llvm::hash_value().

◆ areKnownNonVectorizableLoads()

template<typename T>

bool slpvectorizer::BoUpSLP::areKnownNonVectorizableLoads ( ArrayRef< T * > VL ) const

inline

Checks if the given loads sequence is known as not vectorizable.

Definition at line 2031 of file SLPVectorizer.cpp.

References llvm::hash_value().

Referenced by canVectorizeLoads().

◆ 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 9270 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::any_of(), assert(), llvm::dbgs(), llvm::slpvectorizer::doesInTreeUserNeedToExtract(), llvm::dyn_cast(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::detail::DenseSetImpl< ValueT, MapTy >::end(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::detail::DenseSetImpl< ValueT, MapTy >::find(), llvm::isa(), isDeleted(), LLVM_DEBUG, llvm::none_of(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace(), llvm::Value::users(), and llvm::slpvectorizer::UsesLimit.

◆ buildTree() [1/2]

void BoUpSLP::buildTree ( ArrayRef< Value * > Roots )

Construct a vectorizable tree that starts at Roots.

Definition at line 9538 of file SLPVectorizer.cpp.

References llvm::slpvectorizer::allSameType(), assert(), 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 9527 of file SLPVectorizer.cpp.

References llvm::slpvectorizer::allSameType(), assert(), and deleteTree().

◆ calculateTreeCostAndTrimNonProfitable()

InstructionCost BoUpSLP::calculateTreeCostAndTrimNonProfitable	(	ArrayRef< Value * >	VectorizedVals = {},
		Instruction *	RdxRoot = nullptr )

Calculates the cost of the subtrees, trims non-profitable ones and returns final cost.

Definition at line 18902 of file SLPVectorizer.cpp.

References llvm::any_of(), llvm::dbgs(), llvm::dyn_cast(), llvm::InstructionCost::getInvalid(), I, llvm::isa(), LLVM_DEBUG, and llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace().

◆ canBuildSplitNode()

bool BoUpSLP::canBuildSplitNode	(	ArrayRef< Value * >	VL,
		const InstructionsState &	LocalState,
		SmallVectorImpl< Value * > &	Op1,
		SmallVectorImpl< Value * > &	Op2,
		OrdersType &	ReorderIndices ) const

Checks if it is legal and profitable to build SplitVectorize node for the given VL.

Parameters

Op1	first homogeneous scalars.
Op2	second homogeneous scalars.
ReorderIndices	indices to reorder the scalars.

Returns: true if the node was successfully built.

Definition at line 11400 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::SmallVectorImpl< T >::assign(), llvm::cast(), llvm::SmallVectorImpl< T >::clear(), llvm::dbgs(), llvm::dyn_cast(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::enumerate(), llvm::from_range, llvm::ArrayRef< T >::front(), llvm::SmallVectorTemplateCommon< T, typename >::front(), llvm::slpvectorizer::getAltInstrMask(), llvm::slpvectorizer::getShuffleCost(), getValueType(), getWidenedType(), hasFullVectorsOrPowerOf2(), I, llvm::slpvectorizer::inversePermutation(), llvm::isa(), isAlternateInstruction(), isIdentityOrder(), isMainInstruction(), LLVM_DEBUG, llvm::PoisonMaskElem, llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::seq(), llvm::SmallBitVector::set(), llvm::ArrayRef< T >::size(), llvm::SmallPtrSetImplBase::size(), llvm::SmallVectorTemplateCommon< T, typename >::size(), llvm::TargetTransformInfo::SK_InsertSubvector, llvm::TargetTransformInfo::SK_PermuteTwoSrc, SplitAlternateInstructions, llvm::TargetTransformInfo::TCK_RecipThroughput, llvm::SmallBitVector::test(), and llvm::Values.

◆ 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 13334 of file SLPVectorizer.cpp.

References llvm::cast(), llvm::dyn_cast(), getWidenedType(), llvm::isa(), llvm::Type::isEmptyTy(), isValidElementType(), and N.

◆ canVectorizeLoads()

BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads	(	ArrayRef< Value * >	VL,
		const Value *	VL0,
		SmallVectorImpl< unsigned > &	Order,
		SmallVectorImpl< Value * > &	PointerOps,
		StridedPtrInfo &	SPtrInfo,
		unsigned *	BestVF = nullptr,
		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.
BestVF	return best vector factor, if recursive check found better vectorization sequences rather than masked gather.
TryRecursiveCheck	used to check if long masked gather can be represented as a serie of loads/insert subvector, if profitable.

Definition at line 7393 of file SLPVectorizer.cpp.

References llvm::accumulate(), llvm::all_of(), analyzeConstantStrideCandidate(), analyzeRtStrideCandidate(), llvm::any_of(), areKnownNonVectorizableLoads(), arePointersCompatible(), llvm::ArrayRef(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::back(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::CallingConv::C, canVectorizeLoads(), llvm::cast(), llvm::SmallVectorImpl< T >::clear(), llvm::APInt::clearAllBits(), CompressVectorize, computeCommonAlignment(), CostKind, llvm::count_if(), llvm::slpvectorizer::doesNotNeedToBeScheduled(), llvm::dyn_cast(), llvm::SmallVectorImpl< T >::emplace_back(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::ArrayRef< T >::front(), llvm::SmallVectorTemplateCommon< T, typename >::front(), Gather, GEP, llvm::APInt::getAllOnes(), getFloorFullVectorNumberOfElements(), llvm::slpvectorizer::getGEPCosts(), getMinVF(), llvm::APInt::getOneBitSet(), getParent(), llvm::getPointerOperand(), llvm::getPointersDiff(), getScalarizationOverhead(), llvm::slpvectorizer::getShuffleCost(), llvm::Value::getType(), llvm::getUnderlyingObject(), getWidenedType(), InstructionCost, llvm::isa(), llvm::APInt::isAllOnes(), llvm::IsaPred, isMaskedLoadCompress(), isReverseOrder(), llvm::APInt::isZero(), llvm_unreachable, P, llvm::SmallVectorImpl< T >::resize(), ScatterVectorize, llvm::seq(), llvm::APInt::setAllBits(), llvm::APInt::setBits(), llvm::ArrayRef< T >::size(), llvm::SmallVectorTemplateCommon< T, typename >::size(), llvm::TargetTransformInfo::SK_Broadcast, llvm::TargetTransformInfo::SK_InsertSubvector, llvm::TargetTransformInfo::SK_PermuteSingleSrc, llvm::ArrayRef< T >::slice(), SLPCostThreshold, llvm::sortPtrAccesses(), StridedVectorize, llvm::TargetTransformInfo::TCK_RecipThroughput, and Vectorize.

Referenced by canVectorizeLoads(), and getReorderingData().

◆ canVectorStoreInsertValue()

bool BoUpSLP::canVectorStoreInsertValue ( const TreeEntry * E ) const

Returns: true if the vectorized insertvalue result can be stored directly as a vector, i.e. every insertvalue with an external user is consumed by a single store only.

Definition at line 13367 of file SLPVectorizer.cpp.

References llvm::dyn_cast(), llvm::is_contained(), isDeleted(), and IV.

◆ canVersionBlockForRuntimeChecks()

bool BoUpSLP::canVersionBlockForRuntimeChecks ( BasicBlock * BB ) const

Returns true if BB satisfies the block-level preconditions for runtime alias check versioning (straight-line, outside any loop, duplicable, not a scalar fallback, function not optimized for size).

These checks do not depend on the collected checks, so they can gate the (expensive) optimistic retry before any tree is rebuilt.

Definition at line 24354 of file SLPVectorizer.cpp.

References llvm::any_of(), assert(), blockBodyHasVectorInstructions(), llvm::dyn_cast(), llvm::BasicBlock::getTerminatingMustTailCall(), llvm::BasicBlock::getTerminator(), and I.

Referenced by canVersionForRuntimeChecks().

◆ canVersionForRuntimeChecks()

bool BoUpSLP::canVersionForRuntimeChecks ( )

Returns true if the runtime alias checks can be safely emitted to guard the vectorized region.

Definition at line 24391 of file SLPVectorizer.cpp.

References llvm::any_of(), llvm::sampleprof::Base, canVersionBlockForRuntimeChecks(), llvm::SmallPtrSetImplBase::clear(), llvm::SmallPtrSetImpl< PtrType >::contains(), CostKind, llvm::dyn_cast(), llvm::BasicBlock::getFirstNonPHIIt(), llvm::details::FixedOrScalableQuantity< LeafTy, ValueTy >::getFixedValue(), llvm::getLoadStorePointerOperand(), llvm::getLoadStoreType(), getRuntimeChecksCost(), llvm::getUnderlyingObject(), I, llvm::SmallPtrSetImpl< PtrType >::insert(), llvm::isa(), llvm::details::FixedOrScalableQuantity< LeafTy, ValueTy >::isScalable(), llvm::InstructionCost::isValid(), isVectorized(), MaxMemDepDistance, P, SLPMaxRuntimeAliasChecks, SLPRuntimeAliasChecksMaxScalarCostPercent, llvm::TargetTransformInfo::TCK_RecipThroughput, and llvm::MapVector< KeyT, ValueT, MapType, VectorType, N >::try_emplace().

◆ captureRuntimeCheckBodySnapshot()

void BoUpSLP::captureRuntimeCheckBodySnapshot ( )

Snapshots RTChecks.BB's body (non-PHI, non-terminator) into RTOrigBodyOrder in program order, for the scalar fallback.

Definition at line 24314 of file SLPVectorizer.cpp.

References I, and llvm::isa().

◆ clearReductionData()

void slpvectorizer::BoUpSLP::clearReductionData ( )

inline

Clear the list of the analyzed reduction root instructions.

Definition at line 3363 of file SLPVectorizer.cpp.

◆ computeMinimumValueSizes()

void BoUpSLP::computeMinimumValueSizes ( )

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

Definition at line 27561 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::any_of(), assert(), llvm::bit_ceil(), llvm::SmallVectorImpl< T >::clear(), llvm::ComputeNumSignBits(), llvm::detail::DenseSetImpl< ValueT, MapTy >::contains(), llvm::dyn_cast(), llvm::IntegerType::get(), getNumberOfParts(), llvm::Type::getScalarType(), llvm::Value::getType(), getWidenedType(), I, llvm::isa(), llvm::isGather(), and llvm::SmallVectorTemplateBase< T, bool >::push_back().

◆ deleteTree()

void slpvectorizer::BoUpSLP::deleteTree ( )

inline

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

Definition at line 1743 of file SLPVectorizer.cpp.

Referenced by buildTree(), and buildTree().

◆ eraseInstruction()

void 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 3251 of file SLPVectorizer.cpp.

References I.

Referenced by optimizeGatherSequence(), removeInstructionsAndOperands(), and vectorizeTree().

◆ findBestRootPair()

std::pair< std::optional< int >, int > 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 3226 of file SLPVectorizer.cpp.

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

Referenced by transformNodes().

◆ findPartiallyOrderedLoads()

std::optional< BoUpSLP::OrdersType > BoUpSLP::findPartiallyOrderedLoads ( const TreeEntry & TE )

Sort loads into increasing pointers offsets to allow greater clustering.

Definition at line 7829 of file SLPVectorizer.cpp.

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

Referenced by getReorderingData().

◆ findReusedOrderedScalars()

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

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.

Parameters

TopToBottom	If true, used for the whole tree rotation, false - for sub-tree rotations.
IgnoreReorder	true, if the order of the root node might be ignored.

Definition at line 6345 of file SLPVectorizer.cpp.

References llvm::SmallBitVector::all(), llvm::all_of(), llvm::SmallBitVector::any(), llvm::any_of(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::count(), llvm::dyn_cast(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::end(), llvm::enumerate(), llvm::fill(), llvm::find(), llvm::SmallVectorTemplateCommon< T, typename >::front(), getNumberOfParts(), llvm::slpvectorizer::getNumElems(), llvm::slpvectorizer::getPartNumElems(), getWidenedType(), I, llvm::isa(), llvm::slpvectorizer::isConstant(), isValidElementType(), llvm::not_equal_to(), P, llvm::PoisonMaskElem, llvm::seq(), llvm::SmallBitVector::set(), llvm::SmallVectorTemplateCommon< T, typename >::size(), llvm::TargetTransformInfo::SK_PermuteSingleSrc, and llvm::SmallBitVector::test().

Referenced by getReorderingData().

◆ findRootLaneForValue()

unsigned slpvectorizer::BoUpSLP::findRootLaneForValue ( Value * V ) const

inline

Returns the lane the given value is vectorized to in the root node.

Definition at line 1661 of file SLPVectorizer.cpp.

References assert().

◆ getCanonicalGraphSize()

unsigned slpvectorizer::BoUpSLP::getCanonicalGraphSize ( ) const

inline

Returns the base graph size, before any transformations.

Definition at line 1786 of file SLPVectorizer.cpp.

Referenced by isTreeNotExtendable().

◆ getInsertBuildVectorSrcTy()

FixedVectorType * BoUpSLP::getInsertBuildVectorSrcTy ( const TreeEntry * E ) const

Returns: the source vector type for an InsertElement/InsertValue buildvector node E: the inserted vector type for insertelement, or a vector of the inserted scalar type wide enough to cover the highest inserted index for insertvalue.

Definition at line 13399 of file SLPVectorizer.cpp.

References assert(), llvm::cast(), llvm::dyn_cast(), llvm::Value::getType(), getValueType(), getWidenedType(), and I.

◆ getMaximumVF()

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

inline

Definition at line 1875 of file SLPVectorizer.cpp.

References MaxVFOption.

◆ getMaxVecRegSize()

unsigned slpvectorizer::BoUpSLP::getMaxVecRegSize ( ) const

inline

Definition at line 1862 of file SLPVectorizer.cpp.

◆ getMinVecRegSize()

unsigned slpvectorizer::BoUpSLP::getMinVecRegSize ( ) const

inline

Definition at line 1867 of file SLPVectorizer.cpp.

Referenced by getMinVF().

◆ getMinVF()

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

inline

Definition at line 1871 of file SLPVectorizer.cpp.

References getMinVecRegSize().

Referenced by canVectorizeLoads().

◆ getORE()

OptimizationRemarkEmitter * slpvectorizer::BoUpSLP::getORE ( )

inline

Definition at line 2035 of file SLPVectorizer.cpp.

◆ getReductionType()

FixedVectorType * slpvectorizer::BoUpSLP::getReductionType ( ) const

inline

Returns reduction type after minbitdth analysis.

Definition at line 1692 of file SLPVectorizer.cpp.

References llvm::cast(), llvm::IntegerType::get(), and getWidenedType().

◆ getReorderingData()

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

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.

IgnoreReorder true, if the root node order can be ignored.

Definition at line 7910 of file SLPVectorizer.cpp.

Referenced by reorderBottomToTop(), and reorderTopToBottom().

◆ getRootNodeScalars()

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

inline

Returns the scalars of the root node.

Definition at line 1655 of file SLPVectorizer.cpp.

References assert().

◆ getRootNodeTypeWithNoCast()

std::optional< std::pair< Type *, bool > > slpvectorizer::BoUpSLP::getRootNodeTypeWithNoCast ( ) const

inline

Returns the type/is-signed info for the root node in the graph without casting.

Definition at line 1668 of file SLPVectorizer.cpp.

References llvm::cast(), llvm::SmallVectorTemplateCommon< T, typename >::front(), and llvm::IntegerType::get().

◆ getRuntimeChecksCost()

InstructionCost BoUpSLP::getRuntimeChecksCost ( ) const

Returns the modeled cost of the runtime alias checks collected during the last (optimistic) buildTree().

Definition at line 24547 of file SLPVectorizer.cpp.

References llvm::Type::getInt1Ty(), llvm::CmpInst::ICMP_ULT, and llvm::TargetTransformInfo::TCK_RecipThroughput.

Referenced by canVersionForRuntimeChecks().

◆ getSpillCost()

InstructionCost BoUpSLP::getSpillCost ( )

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

Definition at line 18355 of file SLPVectorizer.cpp.

References llvm::slpvectorizer::allConstant(), llvm::SmallVectorImpl< T >::append(), assert(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::at(), BlockSize, llvm::cast(), Cleanup, llvm::Instruction::comesBefore(), llvm::SmallPtrSetImpl< PtrType >::contains(), llvm::dyn_cast(), llvm::ArrayRef< T >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::First, llvm::IntegerType::get(), llvm::IntrinsicCostAttributes::getArgTypes(), llvm::Type::getContext(), llvm::BasicBlock::getFirstNonPHIOrDbgOrAlloca(), llvm::BasicBlock::getParent(), llvm::ilist_detail::node_parent_access< NodeTy, ParentTy >::getParent(), llvm::BasicBlock::getTerminator(), getWidenedType(), I, II, llvm::InnerAnalysisManagerProxy< AnalysisManagerT, IRUnitT, ExtraArgTs... >::Key, llvm::SmallPtrSetImpl< PtrType >::insert(), llvm::isa(), isVectorized(), llvm::Type::isVectorTy(), llvm::Last, llvm::SmallVectorImpl< T >::pop_back_val(), llvm::pred_begin(), llvm::pred_end(), llvm::SmallVectorTemplateBase< T, bool >::push_back(), ScheduleRegionSizeBudget, llvm::BasicBlock::size(), llvm::TargetTransformInfo::TCK_RecipThroughput, and llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace().

Referenced by getTreeCost().

◆ getTreeCost()

InstructionCost BoUpSLP::getTreeCost	(	InstructionCost	TreeCost,
		ArrayRef< Value * >	VectorizedVals = {},
		InstructionCost	ReductionCost = TTI::TCC_Free,
		Instruction *	RdxRoot = nullptr )

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

Definition at line 19452 of file SLPVectorizer.cpp.

References llvm::SmallBitVector::all(), llvm::all_of(), llvm::any_of(), areTwoInsertFromSameBuildVector(), assert(), llvm::SmallVectorImpl< T >::assign(), llvm::sampleprof::Base, llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::CallingConv::C, llvm::cast(), llvm::cast_or_null(), llvm::detail::DenseSetImpl< ValueT, MapTy >::contains(), CostKind, llvm::count_if(), llvm::Data, llvm::dbgs(), llvm::dump(), llvm::dyn_cast(), llvm::dyn_cast_if_present(), llvm::dyn_cast_or_null(), llvm::SmallVectorImpl< T >::emplace_back(), llvm::ArrayRef< T >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::SmallVectorTemplateCommon< T, typename >::end(), llvm::enumerate(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::find_if(), llvm::find_if_not(), findInnermostNonInvariantLoop(), llvm::SmallVectorTemplateCommon< T, typename >::front(), llvm::IntegerType::get(), llvm::Type::getContext(), getElementIndex(), getExtractWithExtendCost(), llvm::ExtractValueInst::getIndexedType(), llvm::InstructionCost::getInvalid(), llvm::User::getOperand(), llvm::BasicBlock::getParent(), llvm::Type::getScalarType(), llvm::slpvectorizer::getShuffleCost(), getSpillCost(), llvm::BasicBlock::getTerminator(), llvm::InsertElementInst::getType(), llvm::getUnderlyingObject(), getValueType(), getVectorInstrCost(), getWidenedType(), llvm::APInt::getZero(), llvm::has_single_bit(), llvm::Value::hasOneUse(), I, II, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::insert(), llvm::detail::DenseSetImpl< ValueT, MapTy >::insert(), llvm::SmallPtrSetImpl< PtrType >::insert(), llvm::is_contained(), llvm::isa(), llvm::isa_and_nonnull(), llvm::isa_and_present(), llvm::BasicBlock::isEHPad(), isFirstInsertElement(), llvm::ShuffleVectorInst::isIdentityMask(), llvm::isKnownNonNegative(), isPoorThroughputOp(), llvm::InstructionCost::isValid(), isVectorized(), llvm::isVectorizedTy(), LLVM_DEBUG, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::lookup(), LoopAwareTripCount, llvm::MutableArrayRef(), llvm::TargetTransformInfo::None, llvm::none_of(), P, performExtractsShuffleAction(), PHI, llvm::PoisonMaskElem, llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::seq(), llvm::SmallBitVector::set(), llvm::SmallVectorTemplateCommon< T, typename >::size(), llvm::TargetTransformInfo::SK_PermuteSingleSrc, llvm::TargetTransformInfo::SK_PermuteTwoSrc, SLPCostThreshold, SLPInstCountCheck, SLPReVec, llvm::TargetTransformInfo::TCC_Basic, llvm::TargetTransformInfo::TCC_Free, llvm::TargetTransformInfo::TCK_RecipThroughput, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace(), llvm::Value::user_back(), llvm::slpvectorizer::UsesLimit, llvm::Vector, VectorizePoorThroughput, llvm::ViewGraph(), and ViewSLPTree.

◆ getTreeSize()

unsigned slpvectorizer::BoUpSLP::getTreeSize ( ) const

inline

Definition at line 1783 of file SLPVectorizer.cpp.

Referenced by isTreeNotExtendable().

◆ 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 27013 of file SLPVectorizer.cpp.

References llvm::dyn_cast(), llvm::SmallVectorImpl< T >::emplace_back(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::Type::getInt1Ty(), llvm::CallBase::getIntrinsicID(), llvm::Value::getType(), getVectorElementSize(), I, llvm::SmallPtrSetImpl< PtrType >::insert(), llvm::isa(), llvm::SmallVectorImpl< T >::pop_back_val(), RecursionMaxDepth, and llvm::Store.

Referenced by getVectorElementSize().

◆ hasNonCheckableMemBlocker()

bool slpvectorizer::BoUpSLP::hasNonCheckableMemBlocker ( ) const

inline

Returns true if the last buildTree() kept a may-alias memory dependency that is not runtime-checkable (call or a non-simple mem access).

Such a dependency cannot be dropped, so a runtime-checks retry cannot unblock the region and would be pure overhead.

Definition at line 1586 of file SLPVectorizer.cpp.

◆ hasRuntimeAliasChecks()

bool slpvectorizer::BoUpSLP::hasRuntimeAliasChecks ( ) const

inline

Returns true if the last (optimistic) buildTree() collected any runtime alias checks that must guard the vectorized region.

Definition at line 1648 of file SLPVectorizer.cpp.

◆ hasRuntimeCheckableBlockers()

bool slpvectorizer::BoUpSLP::hasRuntimeCheckableBlockers ( ) const

inline

Returns true if the last buildTree() observed a may-alias memory dependency between two distinct, range-checkable base objects, i.e.

a dependency that could be turned into a runtime alias check.

Definition at line 1571 of file SLPVectorizer.cpp.

◆ hasSameNode()

bool slpvectorizer::BoUpSLP::hasSameNode	(	const InstructionsState &	S,
		ArrayRef< Value * >	VL ) const

inline

Checks whether some existing tree entry has scalars equal to VL.

S is the common opcode of VL when one exists; an empty S means the values have no common opcode (mixed buildvector/gather candidates).

Definition at line 1995 of file SLPVectorizer.cpp.

References llvm::any_of(), llvm::SmallPtrSetImpl< PtrType >::insert(), and isConstant().

◆ isAnalyzedReductionRoot()

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

inline

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

Definition at line 3344 of file SLPVectorizer.cpp.

References I.

◆ isAnyGathered()

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

inline

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

Definition at line 3369 of file SLPVectorizer.cpp.

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

◆ isCFGChanged()

bool slpvectorizer::BoUpSLP::isCFGChanged ( ) const

inline

Returns true if vectorization changed the CFG (i.e.

a block was versioned with runtime alias checks). When true, CFG analyses must not be preserved.

Definition at line 1652 of file SLPVectorizer.cpp.

◆ isDeleted()

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

inline

Checks if the instruction is marked for deletion.

Definition at line 3246 of file SLPVectorizer.cpp.

References I.

Referenced by buildExternalUses(), canVectorStoreInsertValue(), and optimizeGatherSequence().

◆ isGathered()

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

inline

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

Definition at line 3373 of file SLPVectorizer.cpp.

◆ isIdentityOrder()

bool slpvectorizer::BoUpSLP::isIdentityOrder ( ArrayRef< unsigned > Order )

inlinestatic

Does this non-empty order represent an identity order?

Identity should be represented as an empty order, so this is used to decide if we can canonicalize a computed order. Undef elements (represented as size) are ignored.

Definition at line 1795 of file SLPVectorizer.cpp.

References llvm::all_of(), assert(), llvm::ArrayRef< T >::empty(), llvm::enumerate(), P, and llvm::ArrayRef< T >::size().

Referenced by canBuildSplitNode(), getReorderingData(), reorderBottomToTop(), and reorderTopToBottom().

◆ isNotScheduled()

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

inline

Checks if the specified value was not schedule.

Definition at line 3377 of file SLPVectorizer.cpp.

◆ isProfitableToReorder()

bool BoUpSLP::isProfitableToReorder ( ) const

Checks if it is profitable to reorder the current tree.

If the tree does not contain many profitable reordable nodes, better to skip it to save compile time.

Definition at line 8330 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::any_of(), llvm::ArrayRef(), llvm::count_if(), DisableTreeReorder, llvm::IsaPred, llvm::Instruction::isBinaryOp(), llvm::slpvectorizer::isCommutative(), and llvm::none_of().

◆ isReducedBitcastRoot()

bool slpvectorizer::BoUpSLP::isReducedBitcastRoot ( ) const

inline

Returns true if the tree results in one of the reduced bitcasts variants.

Definition at line 1709 of file SLPVectorizer.cpp.

◆ isReducedCmpBitcastRoot()

bool slpvectorizer::BoUpSLP::isReducedCmpBitcastRoot ( ) const

inline

Returns true if the tree results in the reduced cmp bitcast root.

Definition at line 1722 of file SLPVectorizer.cpp.

◆ isReductionTree()

bool slpvectorizer::BoUpSLP::isReductionTree ( ) const

inline

Returns true if the tree is a reduction tree.

Definition at line 1730 of file SLPVectorizer.cpp.

◆ isScalarFallbackBlock()

bool slpvectorizer::BoUpSLP::isScalarFallbackBlock ( BasicBlock * BB ) const

inline

Returns true if BB is a scalar fallback block created by runtime alias check versioning.

Definition at line 1626 of file SLPVectorizer.cpp.

◆ isSignedMinBitwidthRootNode()

bool 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 1687 of file SLPVectorizer.cpp.

◆ isStridedLoad()

bool BoUpSLP::isStridedLoad	(	ArrayRef< Value * >	PointerOps,
		Type *	ScalarTy,
		Align	Alignment,
		const int64_t	Diff,
		const size_t	Sz ) const

Checks if strided loads can be generated out of VL loads with pointers PointerOps:

Target with strided load support is detected.
The number of loads is greater than MinProfitableStridedLoads, or the potential stride <= MaxProfitableStride and the potential stride is power-of-2 (to avoid perf regressions for the very small number of loads) and max distance > number of loads, or potential stride is -1.
The loads are ordered, or number of unordered loads <= MaxProfitableUnorderedLoads, or loads are in reversed order. (this check is to avoid extra costs for very expensive shuffles).
Any pointer operand is an instruction with the users outside of the current graph (for masked gathers extra extractelement instructions might be required).

Definition at line 7053 of file SLPVectorizer.cpp.

References llvm::any_of(), getWidenedType(), llvm::has_single_bit(), llvm::isa(), MaxProfitableStride, and MinProfitableStridedLoads.

Referenced by analyzeConstantStrideCandidate().

◆ isTreeNotExtendable()

bool BoUpSLP::isTreeNotExtendable ( ) const

Checks if the graph and all its subgraphs cannot be better vectorized.

It may happen, if all gather nodes are loads and they cannot be "clusterized". In this case even subgraphs cannot be vectorized more effectively than the base graph.

Definition at line 18321 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::slpvectorizer::allConstant(), llvm::slpvectorizer::allSameBlock(), llvm::ArrayRef(), llvm::count_if(), getCanonicalGraphSize(), getSameOpcode(), getTreeSize(), llvm::isa(), llvm::IsaPred, isSplat(), and llvm::seq().

◆ 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 17965 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::slpvectorizer::allConstant(), llvm::slpvectorizer::allSameBlock(), llvm::any_of(), llvm::ArrayRef(), assert(), llvm::cast(), llvm::count_if(), llvm::dbgs(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::APInt::getAllOnes(), getWidenedType(), llvm::isa(), llvm::IsaPred, llvm::isGather(), isSplat(), LLVM_DEBUG, MinTreeSize, llvm::none_of(), llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::DebugCounter::shouldExecute(), llvm::SmallVectorTemplateCommon< T, typename >::size(), SLPCostThreshold, SLPReVec, and llvm::TargetTransformInfo::TCK_RecipThroughput.

◆ isTryingRuntimeAliasChecks()

bool slpvectorizer::BoUpSLP::isTryingRuntimeAliasChecks ( ) const

inline

Returns true if the current vectorization attempt may drop runtime-checkable may-alias dependencies and guard the region with runtime alias checks.

Definition at line 1594 of file SLPVectorizer.cpp.

◆ isVectorized()

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

inline

Check if the value is vectorized in the tree.

Definition at line 3382 of file SLPVectorizer.cpp.

References llvm::any_of(), and assert().

Referenced by canVersionForRuntimeChecks(), getSpillCost(), getTreeCost(), processBuildVector(), and vectorizeTree().

◆ markRuntimeChecksFailedForBlock()

void slpvectorizer::BoUpSLP::markRuntimeChecksFailedForBlock ( BasicBlock * BB )

inline

Records that an optimistic runtime-checks versioning attempt failed for BB.

Definition at line 1638 of file SLPVectorizer.cpp.

◆ optimizeGatherSequence()

void BoUpSLP::optimizeGatherSequence ( )

Perform LICM and CSE on the newly generated gather sequences.

Definition at line 25590 of file SLPVectorizer.cpp.

References A(), llvm::any_of(), assert(), llvm::SmallVectorImpl< T >::assign(), B(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::SmallPtrSetImplBase::clear(), llvm::dbgs(), llvm::dyn_cast(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::end(), eraseInstruction(), llvm::ilist_node_impl< OptionsT >::getIterator(), getNumberOfParts(), llvm::BasicBlock::getTerminator(), getWidenedType(), I, llvm::is_contained(), llvm::isa(), isDeleted(), 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::SmallVectorTemplateCommon< T, typename >::size(), and llvm::sort().

◆ processBuildVector()

template<typename BVTy, typename ResTy, typename... Args>

ResTy slpvectorizer::BoUpSLP::processBuildVector	(	const TreeEntry *	E,
		Type *	ScalarTy,
		Args &...	Params )

Definition at line 22281 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::any_of(), llvm::SmallVectorImpl< T >::append(), llvm::ArrayRef(), assert(), llvm::SmallVectorImpl< T >::assign(), llvm::SmallVectorTemplateCommon< T, typename >::back(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::cast(), llvm::SmallVectorImpl< T >::clear(), llvm::copy(), llvm::count_if(), llvm::dbgs(), llvm::dyn_cast(), llvm::ArrayRef< T >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::end(), llvm::enumerate(), llvm::equal(), llvm::find_if(), llvm::find_if_not(), llvm::for_each(), llvm::SmallVectorTemplateCommon< T, typename >::front(), llvm::PoisonValue::get(), llvm::UndefValue::get(), getNumberOfParts(), llvm::slpvectorizer::getNumElems(), llvm::slpvectorizer::getPartNumElems(), llvm::Value::getType(), getWidenedType(), I, if(), llvm::slpvectorizer::inversePermutation(), llvm::is_contained(), llvm::isa(), llvm::IsaPred, isConstant(), llvm::ShuffleVectorInst::isExtractSubvectorMask(), llvm::isGuaranteedNotToBePoison(), llvm::ShuffleVectorInst::isIdentityMask(), isSplat(), isVectorized(), LLVM_DEBUG, llvm::MutableArrayRef(), llvm::none_of(), P, llvm::PoisonMaskElem, llvm::SmallVectorTemplateBase< T, bool >::push_back(), llvm::slpvectorizer::reorderScalars(), llvm::seq(), llvm::slpvectorizer::shortBundleName(), llvm::SmallVectorTemplateCommon< T, typename >::size(), llvm::TargetTransformInfo::SK_PermuteSingleSrc, llvm::TargetTransformInfo::SK_PermuteTwoSrc, llvm::ArrayRef< T >::slice(), std::swap(), llvm::transform(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace(), llvm::V1, llvm::Values, and llvm::zip().

◆ registerNonVectorizableLoads()

template<typename T>

void slpvectorizer::BoUpSLP::registerNonVectorizableLoads ( ArrayRef< T * > VL )

inline

Registers non-vectorizable sequence of loads.

Definition at line 2025 of file SLPVectorizer.cpp.

References llvm::hash_value().

◆ removeInstructionsAndOperands()

template<typename T>

void slpvectorizer::BoUpSLP::removeInstructionsAndOperands	(	ArrayRef< T * >	DeadVals,
		ArrayRef< std::tuple< WeakTrackingVH, unsigned, bool, bool > >	VectorValuesAndScales )

inline

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

Definition at line 3258 of file SLPVectorizer.cpp.

References llvm::all_of(), assert(), llvm::cast(), llvm::cast_or_null(), llvm::dyn_cast(), llvm::dyn_cast_if_present(), llvm::ArrayRef< T >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), eraseInstruction(), I, llvm::detail::DenseSetImpl< ValueT, MapTy >::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 8821 of file SLPVectorizer.cpp.

References AbstractManglingParser< Derived, Alloc >::NumOps, AbstractManglingParser< Derived, Alloc >::Ops, llvm::all_of(), llvm::any_of(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::SmallPtrSetImplBase::clear(), combineOrders(), llvm::detail::DenseSetImpl< ValueT, MapTy >::contains(), llvm::detail::DenseSetImpl< ValueT, MapTy >::count(), llvm::count_if(), llvm::Data, llvm::dyn_cast(), llvm::ArrayRef< T >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::SmallPtrSetImpl< PtrType >::erase(), llvm::find_if_not(), llvm::slpvectorizer::fixupOrderingIndices(), Gather, getReorderingData(), getValueType(), llvm::getVectorIntrinsicIDForCall(), I, llvm::detail::DenseSetImpl< ValueT, MapTy >::insert(), llvm::SetVector< T, Vector, Set, N >::insert(), llvm::SmallPtrSetImpl< PtrType >::insert(), llvm::SmallPtrSetImpl< PtrType >::insert_range(), llvm::slpvectorizer::inversePermutation(), llvm::isa(), isConstant(), isIdentityOrder(), llvm::isVectorIntrinsicWithScalarOpAtArg(), llvm::make_second_range(), llvm::Intrinsic::not_intrinsic, P, llvm::PoisonMaskElem, llvm::SmallVectorTemplateBase< T, bool >::push_back(), reorderOrder(), reorderReuses(), llvm::slpvectorizer::reorderScalars(), llvm::seq(), llvm::ArrayRef< T >::size(), llvm::SmallVectorTemplateCommon< T, typename >::size(), llvm::SetVector< T, Vector, Set, N >::takeVector(), llvm::transform(), and Users.

◆ 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 8458 of file SLPVectorizer.cpp.

References llvm::slpvectorizer::addMask(), llvm::any_of(), assert(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::cast(), Cleanup, combineOrders(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::count(), llvm::ArrayRef< T >::empty(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::empty(), llvm::SmallVectorTemplateCommon< T, typename >::empty(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::erase(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::slpvectorizer::fixupOrderingIndices(), llvm::for_each(), llvm::slpvectorizer::getAltInstrMask(), getReorderingData(), getValueType(), getWidenedType(), I, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::insert(), llvm::slpvectorizer::inversePermutation(), llvm::isa(), isIdentityOrder(), llvm::PoisonMaskElem, RecursionMaxDepth, reorderOrder(), llvm::slpvectorizer::reorderScalars(), llvm::ArrayRef< T >::size(), SLPReVec, llvm::transform(), and llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace().

◆ resetRuntimeAliasCheckState()

void slpvectorizer::BoUpSLP::resetRuntimeAliasCheckState ( )

inline

Resets the runtime alias check data.

Definition at line 1601 of file SLPVectorizer.cpp.

◆ runtimeChecksFailedForBlock()

bool slpvectorizer::BoUpSLP::runtimeChecksFailedForBlock ( BasicBlock * BB ) const

inline

Returns true if an optimistic runtime-checks versioning attempt already failed for BB, so further retries in the same block can be skipped.

Definition at line 1632 of file SLPVectorizer.cpp.

◆ setHasNonCheckableMemBlocker()

void slpvectorizer::BoUpSLP::setHasNonCheckableMemBlocker ( bool V )

inline

Records that a non-runtime-checkable may-alias dependency was kept.

Definition at line 1589 of file SLPVectorizer.cpp.

◆ setHasRuntimeCheckableBlockers()

void slpvectorizer::BoUpSLP::setHasRuntimeCheckableBlockers ( bool V )

inline

Records whether a may-alias dependency between distinct, range-checkable base objects has been observed, so the caller can decide to retry with runtime alias checks enabled.

Definition at line 1578 of file SLPVectorizer.cpp.

◆ setTryRuntimeAliasChecks()

void slpvectorizer::BoUpSLP::setTryRuntimeAliasChecks ( bool V )

inline

Enables or disables dropping runtime-checkable may-alias dependencies in favor of runtime alias checks for the current vectorization attempt.

Definition at line 1598 of file SLPVectorizer.cpp.

◆ transformNodes()

void BoUpSLP::transformNodes ( )

Transforms graph nodes to target specific representations, if profitable.

Definition at line 14713 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::any_of(), CostKind, llvm::count_if(), llvm::dyn_cast(), llvm::SmallVectorImpl< T >::emplace_back(), llvm::ArrayRef< T >::empty(), findBestRootPair(), I, llvm::is_contained(), llvm::isa(), P, slpvectorizer::BoUpSLP::LookAheadHeuristics::ScoreSplatLoads, llvm::seq(), 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 24715 of file SLPVectorizer.cpp.

References vectorizeTree().

Referenced by vectorizeTree(), and vectorizeTree().

◆ vectorizeTree() [2/2]

Value * BoUpSLP::vectorizeTree	(	const ExtraValueToDebugLocsMap &	ExternallyUsedValues,
		Instruction *	ReductionRoot = nullptr,
		ArrayRef< std::tuple< WeakTrackingVH, unsigned, bool, bool > >	VectorValuesAndScales = {} )

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

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

Definition at line 24720 of file SLPVectorizer.cpp.

References llvm::all_of(), llvm::any_of(), areTwoInsertFromSameBuildVector(), assert(), llvm::SmallVectorImpl< T >::assign(), llvm::SmallVectorTemplateCommon< T, typename >::begin(), llvm::cast(), llvm::detail::DenseSetImpl< ValueT, MapTy >::contains(), llvm::detail::DenseSetImpl< ValueT, MapTy >::count(), createExtractVector(), llvm::Data, llvm::dbgs(), llvm::dyn_cast(), llvm::SmallVectorImpl< T >::emplace_back(), llvm::ArrayRef< T >::empty(), llvm::BasicBlock::end(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::end(), llvm::SmallVectorTemplateCommon< T, typename >::end(), eraseInstruction(), llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::find(), llvm::find_if(), llvm::get(), getElementIndex(), llvm::ilist_node_impl< OptionsT >::getIterator(), llvm::slpvectorizer::getNumElements(), llvm::User::getOperand(), llvm::ilist_detail::node_parent_access< NodeTy, ParentTy >::getParent(), llvm::InsertElementInst::getType(), llvm::Value::getType(), getWidenedType(), I, II, llvm::InnerAnalysisManagerProxy< AnalysisManagerT, IRUnitT, ExtraArgTs... >::Key, llvm::detail::DenseSetImpl< ValueT, MapTy >::insert(), llvm::is_contained(), llvm::isa(), llvm::isa_and_nonnull(), llvm::Type::isIntOrIntVectorTy(), llvm::isKnownNonNegative(), isVectorized(), llvm::isVectorizedTy(), IV, LLVM_DEBUG, llvm::mayHaveNonDefUseDependency(), llvm::Instruction::moveAfter(), PHI, llvm::PoisonMaskElem, llvm::Value::replaceAllUsesWith(), llvm::User::replaceUsesOfWith(), llvm::seq(), llvm::SmallVectorTemplateCommon< T, typename >::size(), SLPReVec, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::try_emplace(), llvm::User::User(), llvm::Value::users(), and vectorizeTree().

◆ DenseMapInfo< EdgeInfo >

friend struct DenseMapInfo< EdgeInfo >

friend

Definition at line 2035 of file SLPVectorizer.cpp.

◆ DOTGraphTraits< BoUpSLP * >

friend struct DOTGraphTraits< BoUpSLP * >

friend

Definition at line 5146 of file SLPVectorizer.cpp.

◆ GraphTraits< BoUpSLP * >

friend struct GraphTraits< BoUpSLP * >

friend

Definition at line 5146 of file SLPVectorizer.cpp.

◆ operator<< [1/4]

raw_ostream & operator<<	(	raw_ostream &	OS,
		const BoUpSLP::ScheduleBundle &	Bundle )

friend

Definition at line 5023 of file SLPVectorizer.cpp.

◆ operator<< [2/4]

raw_ostream & operator<<	(	raw_ostream &	OS,
		const BoUpSLP::ScheduleCopyableData &	SD )

friend

Definition at line 5146 of file SLPVectorizer.cpp.

◆ operator<< [3/4]

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

friend

Definition at line 4916 of file SLPVectorizer.cpp.

◆ operator<< [4/4]

raw_ostream & operator<<	(	raw_ostream &	OS,
		const BoUpSLP::ScheduleEntity &	SE )

friend

Definition at line 4753 of file SLPVectorizer.cpp.

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

lib/Transforms/Vectorize/SLPVectorizer.cpp

TopToBottom	If true, include the order of vectorized stores and insertelement nodes, otherwise skip them.
IgnoreReorder	true, if the root node order can be ignored.

Classes

Public Types

Public Member Functions

Static Public Member Functions

Friends

Detailed Description

Member Typedef Documentation

◆ ExtraValueToDebugLocsMap

◆ InstrList

◆ OrdersType

◆ StoreList

◆ ValueList

◆ ValueSet

Member Enumeration Documentation

◆ LoadsState

Constructor & Destructor Documentation

◆ BoUpSLP()

◆ ~BoUpSLP()

Member Function Documentation

◆ analyzeConstantStrideCandidate()

◆ analyzedReductionRoot()

◆ analyzedReductionVals()

◆ analyzeRtStrideCandidate()

◆ areAnalyzedReductionVals()

◆ areKnownNonVectorizableLoads()

◆ buildExternalUses()

◆ buildTree() [1/2]

◆ buildTree() [2/2]

◆ calculateTreeCostAndTrimNonProfitable()

◆ canBuildSplitNode()

◆ canMapToVector()

◆ canVectorizeLoads()

◆ canVectorStoreInsertValue()

◆ canVersionBlockForRuntimeChecks()

◆ canVersionForRuntimeChecks()

◆ captureRuntimeCheckBodySnapshot()

◆ clearReductionData()

◆ computeMinimumValueSizes()

◆ deleteTree()

◆ eraseInstruction()

◆ findBestRootPair()

◆ findPartiallyOrderedLoads()

◆ findReusedOrderedScalars()

◆ findRootLaneForValue()

◆ getCanonicalGraphSize()

◆ getInsertBuildVectorSrcTy()

◆ getMaximumVF()

◆ getMaxVecRegSize()

◆ getMinVecRegSize()

◆ getMinVF()

◆ getORE()

◆ getReductionType()

◆ getReorderingData()

◆ getRootNodeScalars()

◆ getRootNodeTypeWithNoCast()

◆ getRuntimeChecksCost()

◆ getSpillCost()

◆ getTreeCost()

◆ getTreeSize()

◆ getVectorElementSize()

◆ hasNonCheckableMemBlocker()

◆ hasRuntimeAliasChecks()

◆ hasRuntimeCheckableBlockers()

◆ hasSameNode()

◆ isAnalyzedReductionRoot()

◆ isAnyGathered()

◆ isCFGChanged()

◆ isDeleted()

◆ isGathered()

◆ isIdentityOrder()

◆ isNotScheduled()

◆ isProfitableToReorder()

◆ isReducedBitcastRoot()

◆ isReducedCmpBitcastRoot()

◆ isReductionTree()

◆ isScalarFallbackBlock()

◆ isSignedMinBitwidthRootNode()

◆ isStridedLoad()

◆ isTreeNotExtendable()

◆ isTreeTinyAndNotFullyVectorizable()