LLVM  7.0.0svn
Classes | Public Types | Public Member Functions | Public Attributes | List of all members
llvm::LoopVectorizationCostModel Class Reference

LoopVectorizationCostModel - estimates the expected speedups due to vectorization. More...

Collaboration diagram for llvm::LoopVectorizationCostModel:
Collaboration graph
[legend]

Classes

struct  RegisterUsage
 A struct that represents some properties of the register usage of a loop. More...
 

Public Types

enum  InstWidening {
  CM_Unknown, CM_Widen, CM_Widen_Reverse, CM_Interleave,
  CM_GatherScatter, CM_Scalarize
}
 Decision that was taken during cost calculation for memory instruction. More...
 

Public Member Functions

 LoopVectorizationCostModel (Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, LoopVectorizationLegality *Legal, const TargetTransformInfo &TTI, const TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC, OptimizationRemarkEmitter *ORE, const Function *F, const LoopVectorizeHints *Hints)
 
Optional< unsignedcomputeMaxVF (bool OptForSize)
 
VectorizationFactor selectVectorizationFactor (unsigned MaxVF)
 
void selectUserVectorizationFactor (unsigned UserVF)
 Setup cost-based decisions for user vectorization factor. More...
 
std::pair< unsigned, unsignedgetSmallestAndWidestTypes ()
 
unsigned selectInterleaveCount (bool OptForSize, unsigned VF, unsigned LoopCost)
 
void setCostBasedWideningDecision (unsigned VF)
 Memory access instruction may be vectorized in more than one way. More...
 
SmallVector< RegisterUsage, 8 > calculateRegisterUsage (ArrayRef< unsigned > VFs)
 
void collectValuesToIgnore ()
 Collect values we want to ignore in the cost model. More...
 
const MapVector< Instruction *, uint64_t > & getMinimalBitwidths () const
 
bool isProfitableToScalarize (Instruction *I, unsigned VF) const
 
bool isUniformAfterVectorization (Instruction *I, unsigned VF) const
 Returns true if I is known to be uniform after vectorization. More...
 
bool isScalarAfterVectorization (Instruction *I, unsigned VF) const
 Returns true if I is known to be scalar after vectorization. More...
 
bool canTruncateToMinimalBitwidth (Instruction *I, unsigned VF) const
 
void setWideningDecision (Instruction *I, unsigned VF, InstWidening W, unsigned Cost)
 Save vectorization decision W and Cost taken by the cost model for instruction I and vector width VF. More...
 
void setWideningDecision (const InterleaveGroup *Grp, unsigned VF, InstWidening W, unsigned Cost)
 Save vectorization decision W and Cost taken by the cost model for interleaving group Grp and vector width VF. More...
 
InstWidening getWideningDecision (Instruction *I, unsigned VF)
 Return the cost model decision for the given instruction I and vector width VF. More...
 
unsigned getWideningCost (Instruction *I, unsigned VF)
 Return the vectorization cost for the given instruction I and vector width VF. More...
 
bool isOptimizableIVTruncate (Instruction *I, unsigned VF)
 Return True if instruction I is an optimizable truncate whose operand is an induction variable. More...
 
void collectInstsToScalarize (unsigned VF)
 Collects the instructions to scalarize for each predicated instruction in the loop. More...
 
void collectUniformsAndScalars (unsigned VF)
 Collect Uniform and Scalar values for the given VF. More...
 

Public Attributes

LoopTheLoop
 The loop that we evaluate. More...
 
PredicatedScalarEvolutionPSE
 Predicated scalar evolution analysis. More...
 
LoopInfoLI
 Loop Info analysis. More...
 
LoopVectorizationLegalityLegal
 Vectorization legality. More...
 
const TargetTransformInfoTTI
 Vector target information. More...
 
const TargetLibraryInfoTLI
 Target Library Info. More...
 
DemandedBitsDB
 Demanded bits analysis. More...
 
AssumptionCacheAC
 Assumption cache. More...
 
OptimizationRemarkEmitterORE
 Interface to emit optimization remarks. More...
 
const FunctionTheFunction
 
const LoopVectorizeHints * Hints
 Loop Vectorize Hint. More...
 
SmallPtrSet< const Value *, 16 > ValuesToIgnore
 Values to ignore in the cost model. More...
 
SmallPtrSet< const Value *, 16 > VecValuesToIgnore
 Values to ignore in the cost model when VF > 1. More...
 

Detailed Description

LoopVectorizationCostModel - estimates the expected speedups due to vectorization.

In many cases vectorization is not profitable. This can happen because of a number of reasons. In this class we mainly attempt to predict the expected speedup/slowdowns due to the supported instruction set. We use the TargetTransformInfo to query the different backends for the cost of different operations.

Definition at line 1854 of file LoopVectorize.cpp.

Member Enumeration Documentation

◆ InstWidening

Decision that was taken during cost calculation for memory instruction.

Enumerator
CM_Unknown 
CM_Widen 
CM_Widen_Reverse 
CM_Interleave 
CM_GatherScatter 
CM_Scalarize 

Definition at line 1966 of file LoopVectorize.cpp.

Constructor & Destructor Documentation

◆ LoopVectorizationCostModel()

llvm::LoopVectorizationCostModel::LoopVectorizationCostModel ( Loop L,
PredicatedScalarEvolution PSE,
LoopInfo LI,
LoopVectorizationLegality Legal,
const TargetTransformInfo TTI,
const TargetLibraryInfo TLI,
DemandedBits DB,
AssumptionCache AC,
OptimizationRemarkEmitter ORE,
const Function F,
const LoopVectorizeHints *  Hints 
)
inline

Definition at line 1856 of file LoopVectorize.cpp.

Member Function Documentation

◆ calculateRegisterUsage()

SmallVector< LoopVectorizationCostModel::RegisterUsage, 8 > LoopVectorizationCostModel::calculateRegisterUsage ( ArrayRef< unsigned VFs)

◆ canTruncateToMinimalBitwidth()

bool llvm::LoopVectorizationCostModel::canTruncateToMinimalBitwidth ( Instruction I,
unsigned  VF 
) const
inline
Returns
True if instruction I can be truncated to a smaller bitwidth for vectorization factor VF.

Definition at line 1960 of file LoopVectorize.cpp.

◆ collectInstsToScalarize()

void LoopVectorizationCostModel::collectInstsToScalarize ( unsigned  VF)

◆ collectUniformsAndScalars()

void llvm::LoopVectorizationCostModel::collectUniformsAndScalars ( unsigned  VF)
inline

Collect Uniform and Scalar values for the given VF.

The sets depend on CM decision for Load/Store instructions that may be vectorized as interleave, gather-scatter or scalarized.

Definition at line 2054 of file LoopVectorize.cpp.

References createMissedAnalysis(), and I.

◆ collectValuesToIgnore()

void LoopVectorizationCostModel::collectValuesToIgnore ( )

◆ computeMaxVF()

Optional< unsigned > LoopVectorizationCostModel::computeMaxVF ( bool  OptForSize)

◆ getMinimalBitwidths()

const MapVector<Instruction *, uint64_t>& llvm::LoopVectorizationCostModel::getMinimalBitwidths ( ) const
inline
Returns
The smallest bitwidth each instruction can be represented with. The vector equivalents of these instructions should be truncated to this type.

Definition at line 1926 of file LoopVectorize.cpp.

Referenced by llvm::InnerLoopVectorizer::truncateToMinimalBitwidths().

◆ getSmallestAndWidestTypes()

std::pair< unsigned, unsigned > LoopVectorizationCostModel::getSmallestAndWidestTypes ( )

◆ getWideningCost()

unsigned llvm::LoopVectorizationCostModel::getWideningCost ( Instruction I,
unsigned  VF 
)
inline

Return the vectorization cost for the given instruction I and vector width VF.

Definition at line 2014 of file LoopVectorize.cpp.

References assert().

◆ getWideningDecision()

InstWidening llvm::LoopVectorizationCostModel::getWideningDecision ( Instruction I,
unsigned  VF 
)
inline

Return the cost model decision for the given instruction I and vector width VF.

Return CM_Unknown if this instruction did not pass through the cost modeling.

Definition at line 2003 of file LoopVectorize.cpp.

References assert().

Referenced by llvm::InnerLoopVectorizer::vectorizeMemoryInstruction().

◆ isOptimizableIVTruncate()

bool llvm::LoopVectorizationCostModel::isOptimizableIVTruncate ( Instruction I,
unsigned  VF 
)
inline

Return True if instruction I is an optimizable truncate whose operand is an induction variable.

Such a truncate will be removed by adding a new induction variable with the destination type.

Definition at line 2024 of file LoopVectorize.cpp.

References llvm::dyn_cast(), llvm::LoopVectorizationLegality::getPrimaryInduction(), I, llvm::LoopVectorizationLegality::isInductionPhi(), llvm::TargetTransformInfo::isTruncateFree(), llvm::InnerLoopVectorizer::Legal, ToVectorTy(), and llvm::InnerLoopVectorizer::TTI.

◆ isProfitableToScalarize()

bool llvm::LoopVectorizationCostModel::isProfitableToScalarize ( Instruction I,
unsigned  VF 
) const
inline
Returns
True if it is more profitable to scalarize instruction I for vectorization factor VF.

Definition at line 1932 of file LoopVectorize.cpp.

References assert().

Referenced by llvm::InnerLoopVectorizer::shouldScalarizeInstruction().

◆ isScalarAfterVectorization()

bool llvm::LoopVectorizationCostModel::isScalarAfterVectorization ( Instruction I,
unsigned  VF 
) const
inline

Returns true if I is known to be scalar after vectorization.

Definition at line 1950 of file LoopVectorize.cpp.

References assert().

Referenced by llvm::InnerLoopVectorizer::shouldScalarizeInstruction().

◆ isUniformAfterVectorization()

bool llvm::LoopVectorizationCostModel::isUniformAfterVectorization ( Instruction I,
unsigned  VF 
) const
inline

◆ selectInterleaveCount()

unsigned LoopVectorizationCostModel::selectInterleaveCount ( bool  OptForSize,
unsigned  VF,
unsigned  LoopCost 
)
Returns
The desired interleave count. If interleave count has been specified by metadata it will be returned. Otherwise, the interleave count is computed and returned. VF and LoopCost are the selected vectorization factor and the cost of the selected VF.

Definition at line 6283 of file LoopVectorize.cpp.

References llvm::dbgs(), DEBUG, llvm::DenseMapBase< DerivedT, KeyT, ValueT, KeyInfoT, BucketT >::empty(), llvm::TargetTransformInfo::enableAggressiveInterleaving(), EnableIndVarRegisterHeur, EnableLoadStoreRuntimeInterleave, F(), ForceTargetMaxScalarInterleaveFactor, ForceTargetMaxVectorInterleaveFactor, ForceTargetNumScalarRegs, ForceTargetNumVectorRegs, llvm::TargetTransformInfo::getMaxInterleaveFactor(), llvm::LoopVectorizationLegality::getMaxSafeDepDistBytes(), llvm::TargetTransformInfo::getNumberOfRegisters(), llvm::LoopVectorizationLegality::getNumLoads(), llvm::LoopVectorizationLegality::getNumStores(), llvm::LoopVectorizationLegality::getReductionVars(), llvm::LoopVectorizationLegality::getRuntimePointerChecking(), llvm::PredicatedScalarEvolution::getSE(), llvm::ScalarEvolution::getSmallConstantTripCount(), llvm::InnerLoopVectorizer::Legal, llvm::LoopVectorizationCostModel::RegisterUsage::LoopInvariantRegs, llvm::max(), llvm::LoopVectorizationCostModel::RegisterUsage::MaxLocalUsers, MaxNestedScalarReductionIC, llvm::RuntimePointerChecking::Need, llvm::LoopVectorizationCostModel::RegisterUsage::NumInstructions, llvm::PowerOf2Floor(), llvm::InnerLoopVectorizer::PSE, SmallLoopCost, and llvm::InnerLoopVectorizer::TTI.

Referenced by llvm::LoopVectorizePass::processLoop().

◆ selectUserVectorizationFactor()

void llvm::LoopVectorizationCostModel::selectUserVectorizationFactor ( unsigned  UserVF)
inline

Setup cost-based decisions for user vectorization factor.

Definition at line 1877 of file LoopVectorize.cpp.

References llvm::InnerLoopVectorizer::VF.

◆ selectVectorizationFactor()

VectorizationFactor LoopVectorizationCostModel::selectVectorizationFactor ( unsigned  MaxVF)
Returns
The most profitable vectorization factor and the cost of that VF. This method checks every power of two up to MaxVF. If UserVF is not ZERO then this vectorization factor will be selected if vectorization is possible.

Definition at line 6184 of file LoopVectorize.cpp.

References C, llvm::InnerLoopVectorizer::Cost, llvm::dbgs(), and DEBUG.

◆ setCostBasedWideningDecision()

void LoopVectorizationCostModel::setCostBasedWideningDecision ( unsigned  VF)

Memory access instruction may be vectorized in more than one way.

Form of instruction after vectorization depends on cost. This function takes cost-based decisions for Load/Store instructions and collects them in a map. This decisions map is used for building the lists of loop-uniform and loop-scalar instructions. The calculated cost is saved with widening decision in order to avoid redundant calculations.

Definition at line 6975 of file LoopVectorize.cpp.

References llvm::MCID::Add, assert(), llvm::MCID::Call, llvm::InnerLoopVectorizer::Cost, llvm::dyn_cast(), llvm::SmallVectorBase::empty(), llvm::IntegerType::get(), llvm::VectorType::get(), llvm::TargetTransformInfo::getArithmeticInstrCost(), llvm::TargetTransformInfo::getCastInstrCost(), llvm::TargetTransformInfo::getCFInstrCost(), llvm::TargetTransformInfo::getCmpSelInstrCost(), llvm::SelectInst::getCondition(), llvm::Type::getContext(), llvm::Value::getContext(), llvm::Type::getInt1Ty(), llvm::LoopVectorizationLegality::getInterleavedAccessGroup(), getMemInstValueType(), llvm::PHINode::getNumIncomingValues(), llvm::Instruction::getOpcode(), llvm::User::getOperand(), llvm::Instruction::getParent(), getPointerOperand(), getReciprocalPredBlockProb(), llvm::TargetTransformInfo::getScalarizationOverhead(), getScalarizationOverhead(), llvm::PredicatedScalarEvolution::getSE(), llvm::TargetTransformInfo::getShuffleCost(), llvm::getSplatValue(), llvm::BranchInst::getSuccessor(), llvm::Value::getType(), llvm::ConstantInt::getValue(), getVectorCallCost(), getVectorIntrinsicCost(), llvm::getVectorIntrinsicIDForCall(), I, llvm::SmallPtrSetImpl< PtrType >::insert(), llvm::LoopVectorizationLegality::isAccessInterleaved(), llvm::BranchInst::isConditional(), llvm::LoopVectorizationLegality::isConsecutivePtr(), llvm::LoopVectorizationLegality::isFirstOrderRecurrence(), llvm::LoopVectorizationLegality::isLegalGatherOrScatter(), llvm::APInt::isPowerOf2(), llvm::LoopVectorizationLegality::isScalarWithPredication(), isStrideMul(), llvm::LoopVectorizationLegality::isUniform(), llvm::Type::isVectorTy(), largestIntegerVectorType(), llvm::InnerLoopVectorizer::Legal, LLVM_FALLTHROUGH, llvm::SPII::Load, llvm::max(), llvm::LoopVectorizationLegality::memoryInstructionCanBeWidened(), N, llvm::TargetTransformInfo::OK_AnyValue, llvm::TargetTransformInfo::OK_NonUniformConstantValue, llvm::TargetTransformInfo::OK_UniformConstantValue, llvm::TargetTransformInfo::OK_UniformValue, llvm::TargetTransformInfo::OP_None, llvm::TargetTransformInfo::OP_PowerOf2, llvm::User::operand_values(), llvm::User::operands(), llvm::SmallVectorImpl< T >::pop_back_val(), llvm::TargetTransformInfo::prefersVectorizedAddressing(), llvm::InnerLoopVectorizer::PSE, llvm::SmallVectorTemplateBase< T, isPodLike< T >::value >::push_back(), llvm::MCID::Select, SI, llvm::TargetTransformInfo::SK_ExtractSubvector, smallestIntegerVectorType(), llvm::SPII::Store, llvm::InnerLoopVectorizer::TLI, ToVectorTy(), llvm::InnerLoopVectorizer::TTI, and llvm::InnerLoopVectorizer::VF.

◆ setWideningDecision() [1/2]

void llvm::LoopVectorizationCostModel::setWideningDecision ( Instruction I,
unsigned  VF,
InstWidening  W,
unsigned  Cost 
)
inline

Save vectorization decision W and Cost taken by the cost model for instruction I and vector width VF.

Definition at line 1977 of file LoopVectorize.cpp.

References assert().

◆ setWideningDecision() [2/2]

void llvm::LoopVectorizationCostModel::setWideningDecision ( const InterleaveGroup Grp,
unsigned  VF,
InstWidening  W,
unsigned  Cost 
)
inline

Save vectorization decision W and Cost taken by the cost model for interleaving group Grp and vector width VF.

Broadcast this decicion to all instructions inside the group. But the cost will be assigned to one instruction only.

Definition at line 1985 of file LoopVectorize.cpp.

References assert(), llvm::InterleaveGroup::getFactor(), llvm::InterleaveGroup::getInsertPos(), llvm::InterleaveGroup::getMember(), and I.

Member Data Documentation

◆ AC

AssumptionCache* llvm::LoopVectorizationCostModel::AC

Assumption cache.

Definition at line 2212 of file LoopVectorize.cpp.

◆ DB

DemandedBits* llvm::LoopVectorizationCostModel::DB

Demanded bits analysis.

Definition at line 2209 of file LoopVectorize.cpp.

◆ Hints

const LoopVectorizeHints* llvm::LoopVectorizationCostModel::Hints

Loop Vectorize Hint.

Definition at line 2220 of file LoopVectorize.cpp.

◆ Legal

LoopVectorizationLegality* llvm::LoopVectorizationCostModel::Legal

Vectorization legality.

Definition at line 2200 of file LoopVectorize.cpp.

◆ LI

LoopInfo* llvm::LoopVectorizationCostModel::LI

Loop Info analysis.

Definition at line 2197 of file LoopVectorize.cpp.

◆ ORE

OptimizationRemarkEmitter* llvm::LoopVectorizationCostModel::ORE

Interface to emit optimization remarks.

Definition at line 2215 of file LoopVectorize.cpp.

◆ PSE

PredicatedScalarEvolution& llvm::LoopVectorizationCostModel::PSE

Predicated scalar evolution analysis.

Definition at line 2194 of file LoopVectorize.cpp.

◆ TheFunction

const Function* llvm::LoopVectorizationCostModel::TheFunction

Definition at line 2217 of file LoopVectorize.cpp.

◆ TheLoop

Loop* llvm::LoopVectorizationCostModel::TheLoop

The loop that we evaluate.

Definition at line 2191 of file LoopVectorize.cpp.

◆ TLI

const TargetLibraryInfo* llvm::LoopVectorizationCostModel::TLI

Target Library Info.

Definition at line 2206 of file LoopVectorize.cpp.

◆ TTI

const TargetTransformInfo& llvm::LoopVectorizationCostModel::TTI

Vector target information.

Definition at line 2203 of file LoopVectorize.cpp.

◆ ValuesToIgnore

SmallPtrSet<const Value *, 16> llvm::LoopVectorizationCostModel::ValuesToIgnore

Values to ignore in the cost model.

Definition at line 2223 of file LoopVectorize.cpp.

◆ VecValuesToIgnore

SmallPtrSet<const Value *, 16> llvm::LoopVectorizationCostModel::VecValuesToIgnore

Values to ignore in the cost model when VF > 1.

Definition at line 2226 of file LoopVectorize.cpp.


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