//===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===// // // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. // See https://llvm.org/LICENSE.txt for license information. // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception // //===----------------------------------------------------------------------===// // // This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops // and generates target-independent LLVM-IR. // The vectorizer uses the TargetTransformInfo analysis to estimate the costs // of instructions in order to estimate the profitability of vectorization. // // The loop vectorizer combines consecutive loop iterations into a single // 'wide' iteration. After this transformation the index is incremented // by the SIMD vector width, and not by one. // // This pass has three parts: // 1. The main loop pass that drives the different parts. // 2. LoopVectorizationLegality - A unit that checks for the legality // of the vectorization. // 3. InnerLoopVectorizer - A unit that performs the actual // widening of instructions. // 4. LoopVectorizationCostModel - A unit that checks for the profitability // of vectorization. It decides on the optimal vector width, which // can be one, if vectorization is not profitable. // // There is a development effort going on to migrate loop vectorizer to the // VPlan infrastructure and to introduce outer loop vectorization support (see // docs/VectorizationPlan.rst and // http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this // purpose, we temporarily introduced the VPlan-native vectorization path: an // alternative vectorization path that is natively implemented on top of the // VPlan infrastructure. See EnableVPlanNativePath for enabling. // //===----------------------------------------------------------------------===// // // The reduction-variable vectorization is based on the paper: // D. Nuzman and R. Henderson. Multi-platform Auto-vectorization. // // Variable uniformity checks are inspired by: // Karrenberg, R. and Hack, S. Whole Function Vectorization. // // The interleaved access vectorization is based on the paper: // Dorit Nuzman, Ira Rosen and Ayal Zaks. Auto-Vectorization of Interleaved // Data for SIMD // // Other ideas/concepts are from: // A. Zaks and D. Nuzman. Autovectorization in GCC-two years later. // // S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua. An Evaluation of // Vectorizing Compilers. // //===----------------------------------------------------------------------===// #include "llvm/Transforms/Vectorize/LoopVectorize.h" #include "LoopVectorizationPlanner.h" #include "VPRecipeBuilder.h" #include "VPlan.h" #include "VPlanAnalysis.h" #include "VPlanCFG.h" #include "VPlanHelpers.h" #include "VPlanPatternMatch.h" #include "VPlanTransforms.h" #include "VPlanUtils.h" #include "VPlanVerifier.h" #include "llvm/ADT/APInt.h" #include "llvm/ADT/ArrayRef.h" #include "llvm/ADT/DenseMap.h" #include "llvm/ADT/DenseMapInfo.h" #include "llvm/ADT/Hashing.h" #include "llvm/ADT/MapVector.h" #include "llvm/ADT/STLExtras.h" #include "llvm/ADT/SmallPtrSet.h" #include "llvm/ADT/SmallVector.h" #include "llvm/ADT/Statistic.h" #include "llvm/ADT/StringRef.h" #include "llvm/ADT/Twine.h" #include "llvm/ADT/TypeSwitch.h" #include "llvm/ADT/iterator_range.h" #include "llvm/Analysis/AssumptionCache.h" #include "llvm/Analysis/BasicAliasAnalysis.h" #include "llvm/Analysis/BlockFrequencyInfo.h" #include "llvm/Analysis/CFG.h" #include "llvm/Analysis/CodeMetrics.h" #include "llvm/Analysis/DemandedBits.h" #include "llvm/Analysis/GlobalsModRef.h" #include "llvm/Analysis/LoopAccessAnalysis.h" #include "llvm/Analysis/LoopAnalysisManager.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/LoopIterator.h" #include "llvm/Analysis/OptimizationRemarkEmitter.h" #include "llvm/Analysis/ProfileSummaryInfo.h" #include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Analysis/ScalarEvolutionExpressions.h" #include "llvm/Analysis/ScalarEvolutionPatternMatch.h" #include "llvm/Analysis/TargetLibraryInfo.h" #include "llvm/Analysis/TargetTransformInfo.h" #include "llvm/Analysis/ValueTracking.h" #include "llvm/Analysis/VectorUtils.h" #include "llvm/IR/Attributes.h" #include "llvm/IR/BasicBlock.h" #include "llvm/IR/CFG.h" #include "llvm/IR/Constant.h" #include "llvm/IR/Constants.h" #include "llvm/IR/DataLayout.h" #include "llvm/IR/DebugInfo.h" #include "llvm/IR/DebugLoc.h" #include "llvm/IR/DerivedTypes.h" #include "llvm/IR/DiagnosticInfo.h" #include "llvm/IR/Dominators.h" #include "llvm/IR/Function.h" #include "llvm/IR/IRBuilder.h" #include "llvm/IR/InstrTypes.h" #include "llvm/IR/Instruction.h" #include "llvm/IR/Instructions.h" #include "llvm/IR/IntrinsicInst.h" #include "llvm/IR/Intrinsics.h" #include "llvm/IR/MDBuilder.h" #include "llvm/IR/Metadata.h" #include "llvm/IR/Module.h" #include "llvm/IR/Operator.h" #include "llvm/IR/PatternMatch.h" #include "llvm/IR/ProfDataUtils.h" #include "llvm/IR/Type.h" #include "llvm/IR/Use.h" #include "llvm/IR/User.h" #include "llvm/IR/Value.h" #include "llvm/IR/Verifier.h" #include "llvm/Support/Casting.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/InstructionCost.h" #include "llvm/Support/MathExtras.h" #include "llvm/Support/NativeFormatting.h" #include "llvm/Support/raw_ostream.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/Transforms/Utils/InjectTLIMappings.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/LoopSimplify.h" #include "llvm/Transforms/Utils/LoopUtils.h" #include "llvm/Transforms/Utils/LoopVersioning.h" #include "llvm/Transforms/Utils/ScalarEvolutionExpander.h" #include "llvm/Transforms/Utils/SizeOpts.h" #include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h" #include #include #include #include #include #include #include #include #include #include using namespace llvm; using namespace SCEVPatternMatch; #define LV_NAME "loop-vectorize" #define DEBUG_TYPE LV_NAME #ifndef NDEBUG const char VerboseDebug[] = DEBUG_TYPE "-verbose"; #endif /// @{ /// Metadata attribute names const char LLVMLoopVectorizeFollowupAll[] = "llvm.loop.vectorize.followup_all"; const char LLVMLoopVectorizeFollowupVectorized[] = "llvm.loop.vectorize.followup_vectorized"; const char LLVMLoopVectorizeFollowupEpilogue[] = "llvm.loop.vectorize.followup_epilogue"; /// @} STATISTIC(LoopsVectorized, "Number of loops vectorized"); STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization"); STATISTIC(LoopsEpilogueVectorized, "Number of epilogues vectorized"); STATISTIC(LoopsEarlyExitVectorized, "Number of early exit loops vectorized"); static cl::opt EnableEpilogueVectorization( "enable-epilogue-vectorization", cl::init(true), cl::Hidden, cl::desc("Enable vectorization of epilogue loops.")); static cl::opt EpilogueVectorizationForceVF( "epilogue-vectorization-force-VF", cl::init(1), cl::Hidden, cl::desc("When epilogue vectorization is enabled, and a value greater than " "1 is specified, forces the given VF for all applicable epilogue " "loops.")); static cl::opt EpilogueVectorizationMinVF( "epilogue-vectorization-minimum-VF", cl::Hidden, cl::desc("Only loops with vectorization factor equal to or larger than " "the specified value are considered for epilogue vectorization.")); /// Loops with a known constant trip count below this number are vectorized only /// if no scalar iteration overheads are incurred. static cl::opt TinyTripCountVectorThreshold( "vectorizer-min-trip-count", cl::init(16), cl::Hidden, cl::desc("Loops with a constant trip count that is smaller than this " "value are vectorized only if no scalar iteration overheads " "are incurred.")); static cl::opt VectorizeMemoryCheckThreshold( "vectorize-memory-check-threshold", cl::init(128), cl::Hidden, cl::desc("The maximum allowed number of runtime memory checks")); // Option prefer-predicate-over-epilogue indicates that an epilogue is undesired, // that predication is preferred, and this lists all options. I.e., the // vectorizer will try to fold the tail-loop (epilogue) into the vector body // and predicate the instructions accordingly. If tail-folding fails, there are // different fallback strategies depending on these values: namespace PreferPredicateTy { enum Option { ScalarEpilogue = 0, PredicateElseScalarEpilogue, PredicateOrDontVectorize }; } // namespace PreferPredicateTy static cl::opt PreferPredicateOverEpilogue( "prefer-predicate-over-epilogue", cl::init(PreferPredicateTy::ScalarEpilogue), cl::Hidden, cl::desc("Tail-folding and predication preferences over creating a scalar " "epilogue loop."), cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue, "scalar-epilogue", "Don't tail-predicate loops, create scalar epilogue"), clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue, "predicate-else-scalar-epilogue", "prefer tail-folding, create scalar epilogue if tail " "folding fails."), clEnumValN(PreferPredicateTy::PredicateOrDontVectorize, "predicate-dont-vectorize", "prefers tail-folding, don't attempt vectorization if " "tail-folding fails."))); static cl::opt ForceTailFoldingStyle( "force-tail-folding-style", cl::desc("Force the tail folding style"), cl::init(TailFoldingStyle::None), cl::values( clEnumValN(TailFoldingStyle::None, "none", "Disable tail folding"), clEnumValN( TailFoldingStyle::Data, "data", "Create lane mask for data only, using active.lane.mask intrinsic"), clEnumValN(TailFoldingStyle::DataWithoutLaneMask, "data-without-lane-mask", "Create lane mask with compare/stepvector"), clEnumValN(TailFoldingStyle::DataAndControlFlow, "data-and-control", "Create lane mask using active.lane.mask intrinsic, and use " "it for both data and control flow"), clEnumValN(TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck, "data-and-control-without-rt-check", "Similar to data-and-control, but remove the runtime check"), clEnumValN(TailFoldingStyle::DataWithEVL, "data-with-evl", "Use predicated EVL instructions for tail folding. If EVL " "is unsupported, fallback to data-without-lane-mask."))); static cl::opt MaximizeBandwidth( "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden, cl::desc("Maximize bandwidth when selecting vectorization factor which " "will be determined by the smallest type in loop.")); static cl::opt EnableInterleavedMemAccesses( "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden, cl::desc("Enable vectorization on interleaved memory accesses in a loop")); /// An interleave-group may need masking if it resides in a block that needs /// predication, or in order to mask away gaps. static cl::opt EnableMaskedInterleavedMemAccesses( "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden, cl::desc("Enable vectorization on masked interleaved memory accesses in a loop")); static cl::opt ForceTargetNumScalarRegs( "force-target-num-scalar-regs", cl::init(0), cl::Hidden, cl::desc("A flag that overrides the target's number of scalar registers.")); static cl::opt ForceTargetNumVectorRegs( "force-target-num-vector-regs", cl::init(0), cl::Hidden, cl::desc("A flag that overrides the target's number of vector registers.")); static cl::opt ForceTargetMaxScalarInterleaveFactor( "force-target-max-scalar-interleave", cl::init(0), cl::Hidden, cl::desc("A flag that overrides the target's max interleave factor for " "scalar loops.")); static cl::opt ForceTargetMaxVectorInterleaveFactor( "force-target-max-vector-interleave", cl::init(0), cl::Hidden, cl::desc("A flag that overrides the target's max interleave factor for " "vectorized loops.")); cl::opt llvm::ForceTargetInstructionCost( "force-target-instruction-cost", cl::init(0), cl::Hidden, cl::desc("A flag that overrides the target's expected cost for " "an instruction to a single constant value. Mostly " "useful for getting consistent testing.")); static cl::opt ForceTargetSupportsScalableVectors( "force-target-supports-scalable-vectors", cl::init(false), cl::Hidden, cl::desc( "Pretend that scalable vectors are supported, even if the target does " "not support them. This flag should only be used for testing.")); static cl::opt SmallLoopCost( "small-loop-cost", cl::init(20), cl::Hidden, cl::desc( "The cost of a loop that is considered 'small' by the interleaver.")); static cl::opt LoopVectorizeWithBlockFrequency( "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden, cl::desc("Enable the use of the block frequency analysis to access PGO " "heuristics minimizing code growth in cold regions and being more " "aggressive in hot regions.")); // Runtime interleave loops for load/store throughput. static cl::opt EnableLoadStoreRuntimeInterleave( "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden, cl::desc( "Enable runtime interleaving until load/store ports are saturated")); /// The number of stores in a loop that are allowed to need predication. static cl::opt NumberOfStoresToPredicate( "vectorize-num-stores-pred", cl::init(1), cl::Hidden, cl::desc("Max number of stores to be predicated behind an if.")); static cl::opt EnableIndVarRegisterHeur( "enable-ind-var-reg-heur", cl::init(true), cl::Hidden, cl::desc("Count the induction variable only once when interleaving")); static cl::opt EnableCondStoresVectorization( "enable-cond-stores-vec", cl::init(true), cl::Hidden, cl::desc("Enable if predication of stores during vectorization.")); static cl::opt MaxNestedScalarReductionIC( "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden, cl::desc("The maximum interleave count to use when interleaving a scalar " "reduction in a nested loop.")); static cl::opt PreferInLoopReductions("prefer-inloop-reductions", cl::init(false), cl::Hidden, cl::desc("Prefer in-loop vector reductions, " "overriding the targets preference.")); static cl::opt ForceOrderedReductions( "force-ordered-reductions", cl::init(false), cl::Hidden, cl::desc("Enable the vectorisation of loops with in-order (strict) " "FP reductions")); static cl::opt PreferPredicatedReductionSelect( "prefer-predicated-reduction-select", cl::init(false), cl::Hidden, cl::desc( "Prefer predicating a reduction operation over an after loop select.")); cl::opt llvm::EnableVPlanNativePath( "enable-vplan-native-path", cl::Hidden, cl::desc("Enable VPlan-native vectorization path with " "support for outer loop vectorization.")); cl::opt llvm::VerifyEachVPlan("vplan-verify-each", #ifdef EXPENSIVE_CHECKS cl::init(true), #else cl::init(false), #endif cl::Hidden, cl::desc("Verfiy VPlans after VPlan transforms.")); // This flag enables the stress testing of the VPlan H-CFG construction in the // VPlan-native vectorization path. It must be used in conjuction with // -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the // verification of the H-CFGs built. static cl::opt VPlanBuildStressTest( "vplan-build-stress-test", cl::init(false), cl::Hidden, cl::desc( "Build VPlan for every supported loop nest in the function and bail " "out right after the build (stress test the VPlan H-CFG construction " "in the VPlan-native vectorization path).")); cl::opt llvm::EnableLoopInterleaving( "interleave-loops", cl::init(true), cl::Hidden, cl::desc("Enable loop interleaving in Loop vectorization passes")); cl::opt llvm::EnableLoopVectorization( "vectorize-loops", cl::init(true), cl::Hidden, cl::desc("Run the Loop vectorization passes")); static cl::opt ForceSafeDivisor( "force-widen-divrem-via-safe-divisor", cl::Hidden, cl::desc( "Override cost based safe divisor widening for div/rem instructions")); static cl::opt UseWiderVFIfCallVariantsPresent( "vectorizer-maximize-bandwidth-for-vector-calls", cl::init(true), cl::Hidden, cl::desc("Try wider VFs if they enable the use of vector variants")); static cl::opt EnableEarlyExitVectorization( "enable-early-exit-vectorization", cl::init(true), cl::Hidden, cl::desc( "Enable vectorization of early exit loops with uncountable exits.")); // Likelyhood of bypassing the vectorized loop because there are zero trips left // after prolog. See `emitIterationCountCheck`. static constexpr uint32_t MinItersBypassWeights[] = {1, 127}; /// A helper function that returns true if the given type is irregular. The /// type is irregular if its allocated size doesn't equal the store size of an /// element of the corresponding vector type. static bool hasIrregularType(Type *Ty, const DataLayout &DL) { // Determine if an array of N elements of type Ty is "bitcast compatible" // with a vector. // This is only true if there is no padding between the array elements. return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty); } /// A version of ScalarEvolution::getSmallConstantTripCount that returns an /// ElementCount to include loops whose trip count is a function of vscale. static ElementCount getSmallConstantTripCount(ScalarEvolution *SE, const Loop *L) { if (unsigned ExpectedTC = SE->getSmallConstantTripCount(L)) return ElementCount::getFixed(ExpectedTC); const SCEV *BTC = SE->getBackedgeTakenCount(L); if (isa(BTC)) return ElementCount::getFixed(0); const SCEV *ExitCount = SE->getTripCountFromExitCount(BTC, BTC->getType(), L); if (isa(ExitCount)) return ElementCount::getScalable(1); const APInt *Scale; if (match(ExitCount, m_scev_Mul(m_scev_APInt(Scale), m_SCEVVScale()))) if (cast(ExitCount)->hasNoUnsignedWrap()) if (Scale->getActiveBits() <= 32) return ElementCount::getScalable(Scale->getZExtValue()); return ElementCount::getFixed(0); } /// Returns "best known" trip count, which is either a valid positive trip count /// or std::nullopt when an estimate cannot be made (including when the trip /// count would overflow), for the specified loop \p L as defined by the /// following procedure: /// 1) Returns exact trip count if it is known. /// 2) Returns expected trip count according to profile data if any. /// 3) Returns upper bound estimate if known, and if \p CanUseConstantMax. /// 4) Returns std::nullopt if all of the above failed. static std::optional getSmallBestKnownTC(PredicatedScalarEvolution &PSE, Loop *L, bool CanUseConstantMax = true) { // Check if exact trip count is known. if (auto ExpectedTC = getSmallConstantTripCount(PSE.getSE(), L)) return ExpectedTC; // Check if there is an expected trip count available from profile data. if (LoopVectorizeWithBlockFrequency) if (auto EstimatedTC = getLoopEstimatedTripCount(L)) return ElementCount::getFixed(*EstimatedTC); if (!CanUseConstantMax) return std::nullopt; // Check if upper bound estimate is known. if (unsigned ExpectedTC = PSE.getSmallConstantMaxTripCount()) return ElementCount::getFixed(ExpectedTC); return std::nullopt; } namespace { // Forward declare GeneratedRTChecks. class GeneratedRTChecks; using SCEV2ValueTy = DenseMap; } // namespace namespace llvm { AnalysisKey ShouldRunExtraVectorPasses::Key; /// InnerLoopVectorizer vectorizes loops which contain only one basic /// block to a specified vectorization factor (VF). /// This class performs the widening of scalars into vectors, or multiple /// scalars. This class also implements the following features: /// * It inserts an epilogue loop for handling loops that don't have iteration /// counts that are known to be a multiple of the vectorization factor. /// * It handles the code generation for reduction variables. /// * Scalarization (implementation using scalars) of un-vectorizable /// instructions. /// InnerLoopVectorizer does not perform any vectorization-legality /// checks, and relies on the caller to check for the different legality /// aspects. The InnerLoopVectorizer relies on the /// LoopVectorizationLegality class to provide information about the induction /// and reduction variables that were found to a given vectorization factor. class InnerLoopVectorizer { public: InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, const TargetTransformInfo *TTI, AssumptionCache *AC, ElementCount VecWidth, ElementCount MinProfitableTripCount, unsigned UnrollFactor, LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, GeneratedRTChecks &RTChecks, VPlan &Plan) : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TTI(TTI), AC(AC), VF(VecWidth), MinProfitableTripCount(MinProfitableTripCount), UF(UnrollFactor), Builder(PSE.getSE()->getContext()), Cost(CM), BFI(BFI), PSI(PSI), RTChecks(RTChecks), Plan(Plan), VectorPHVPBB(cast( Plan.getVectorLoopRegion()->getSinglePredecessor())) {} virtual ~InnerLoopVectorizer() = default; /// Create a new empty loop that will contain vectorized instructions later /// on, while the old loop will be used as the scalar remainder. Control flow /// is generated around the vectorized (and scalar epilogue) loops consisting /// of various checks and bypasses. Return the pre-header block of the new /// loop. In the case of epilogue vectorization, this function is overriden to /// handle the more complex control flow around the loops. virtual BasicBlock *createVectorizedLoopSkeleton(); /// Fix the vectorized code, taking care of header phi's, and more. void fixVectorizedLoop(VPTransformState &State); /// Fix the non-induction PHIs in \p Plan. void fixNonInductionPHIs(VPTransformState &State); /// Returns the original loop trip count. Value *getTripCount() const { return TripCount; } /// Used to set the trip count after ILV's construction and after the /// preheader block has been executed. Note that this always holds the trip /// count of the original loop for both main loop and epilogue vectorization. void setTripCount(Value *TC) { TripCount = TC; } /// Return the additional bypass block which targets the scalar loop by /// skipping the epilogue loop after completing the main loop. BasicBlock *getAdditionalBypassBlock() const { assert(AdditionalBypassBlock && "Trying to access AdditionalBypassBlock but it has not been set"); return AdditionalBypassBlock; } protected: friend class LoopVectorizationPlanner; // Create a check to see if the vector loop should be executed Value *createIterationCountCheck(ElementCount VF, unsigned UF) const; /// Emit a bypass check to see if the vector trip count is zero, including if /// it overflows. void emitIterationCountCheck(BasicBlock *Bypass); /// Emit basic blocks (prefixed with \p Prefix) for the iteration check, /// vector loop preheader, middle block and scalar preheader. void createVectorLoopSkeleton(StringRef Prefix); /// Allow subclasses to override and print debug traces before/after vplan /// execution, when trace information is requested. virtual void printDebugTracesAtStart() {} virtual void printDebugTracesAtEnd() {} /// Introduces a new VPIRBasicBlock for \p CheckIRBB to Plan between the /// vector preheader and its predecessor, also connecting the new block to the /// scalar preheader. void introduceCheckBlockInVPlan(BasicBlock *CheckIRBB); /// The original loop. Loop *OrigLoop; /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies /// dynamic knowledge to simplify SCEV expressions and converts them to a /// more usable form. PredicatedScalarEvolution &PSE; /// Loop Info. LoopInfo *LI; /// Dominator Tree. DominatorTree *DT; /// Target Transform Info. const TargetTransformInfo *TTI; /// Assumption Cache. AssumptionCache *AC; /// The vectorization SIMD factor to use. Each vector will have this many /// vector elements. ElementCount VF; ElementCount MinProfitableTripCount; /// The vectorization unroll factor to use. Each scalar is vectorized to this /// many different vector instructions. unsigned UF; /// The builder that we use IRBuilder<> Builder; // --- Vectorization state --- /// The vector-loop preheader. BasicBlock *LoopVectorPreHeader = nullptr; /// The scalar-loop preheader. BasicBlock *LoopScalarPreHeader = nullptr; /// Trip count of the original loop. Value *TripCount = nullptr; /// Trip count of the widened loop (TripCount - TripCount % (VF*UF)) Value *VectorTripCount = nullptr; /// The profitablity analysis. LoopVectorizationCostModel *Cost; /// BFI and PSI are used to check for profile guided size optimizations. BlockFrequencyInfo *BFI; ProfileSummaryInfo *PSI; /// Structure to hold information about generated runtime checks, responsible /// for cleaning the checks, if vectorization turns out unprofitable. GeneratedRTChecks &RTChecks; /// The additional bypass block which conditionally skips over the epilogue /// loop after executing the main loop. Needed to resume inductions and /// reductions during epilogue vectorization. BasicBlock *AdditionalBypassBlock = nullptr; VPlan &Plan; /// The vector preheader block of \p Plan, used as target for check blocks /// introduced during skeleton creation. VPBasicBlock *VectorPHVPBB; }; /// Encapsulate information regarding vectorization of a loop and its epilogue. /// This information is meant to be updated and used across two stages of /// epilogue vectorization. struct EpilogueLoopVectorizationInfo { ElementCount MainLoopVF = ElementCount::getFixed(0); unsigned MainLoopUF = 0; ElementCount EpilogueVF = ElementCount::getFixed(0); unsigned EpilogueUF = 0; BasicBlock *MainLoopIterationCountCheck = nullptr; BasicBlock *EpilogueIterationCountCheck = nullptr; Value *TripCount = nullptr; Value *VectorTripCount = nullptr; VPlan &EpiloguePlan; EpilogueLoopVectorizationInfo(ElementCount MVF, unsigned MUF, ElementCount EVF, unsigned EUF, VPlan &EpiloguePlan) : MainLoopVF(MVF), MainLoopUF(MUF), EpilogueVF(EVF), EpilogueUF(EUF), EpiloguePlan(EpiloguePlan) { assert(EUF == 1 && "A high UF for the epilogue loop is likely not beneficial."); } }; /// An extension of the inner loop vectorizer that creates a skeleton for a /// vectorized loop that has its epilogue (residual) also vectorized. /// The idea is to run the vplan on a given loop twice, firstly to setup the /// skeleton and vectorize the main loop, and secondly to complete the skeleton /// from the first step and vectorize the epilogue. This is achieved by /// deriving two concrete strategy classes from this base class and invoking /// them in succession from the loop vectorizer planner. class InnerLoopAndEpilogueVectorizer : public InnerLoopVectorizer { public: InnerLoopAndEpilogueVectorizer( Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, const TargetTransformInfo *TTI, AssumptionCache *AC, EpilogueLoopVectorizationInfo &EPI, LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, GeneratedRTChecks &Checks, VPlan &Plan, ElementCount VecWidth, ElementCount MinProfitableTripCount, unsigned UnrollFactor) : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TTI, AC, VecWidth, MinProfitableTripCount, UnrollFactor, CM, BFI, PSI, Checks, Plan), EPI(EPI) {} // Override this function to handle the more complex control flow around the // three loops. BasicBlock *createVectorizedLoopSkeleton() final { return createEpilogueVectorizedLoopSkeleton(); } /// The interface for creating a vectorized skeleton using one of two /// different strategies, each corresponding to one execution of the vplan /// as described above. virtual BasicBlock *createEpilogueVectorizedLoopSkeleton() = 0; /// Holds and updates state information required to vectorize the main loop /// and its epilogue in two separate passes. This setup helps us avoid /// regenerating and recomputing runtime safety checks. It also helps us to /// shorten the iteration-count-check path length for the cases where the /// iteration count of the loop is so small that the main vector loop is /// completely skipped. EpilogueLoopVectorizationInfo &EPI; }; /// A specialized derived class of inner loop vectorizer that performs /// vectorization of *main* loops in the process of vectorizing loops and their /// epilogues. class EpilogueVectorizerMainLoop : public InnerLoopAndEpilogueVectorizer { public: EpilogueVectorizerMainLoop(Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, const TargetTransformInfo *TTI, AssumptionCache *AC, EpilogueLoopVectorizationInfo &EPI, LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, GeneratedRTChecks &Check, VPlan &Plan) : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TTI, AC, EPI, CM, BFI, PSI, Check, Plan, EPI.MainLoopVF, EPI.MainLoopVF, EPI.MainLoopUF) {} /// Implements the interface for creating a vectorized skeleton using the /// *main loop* strategy (ie the first pass of vplan execution). BasicBlock *createEpilogueVectorizedLoopSkeleton() final; protected: /// Emits an iteration count bypass check once for the main loop (when \p /// ForEpilogue is false) and once for the epilogue loop (when \p /// ForEpilogue is true). BasicBlock *emitIterationCountCheck(BasicBlock *Bypass, bool ForEpilogue); void printDebugTracesAtStart() override; void printDebugTracesAtEnd() override; }; // A specialized derived class of inner loop vectorizer that performs // vectorization of *epilogue* loops in the process of vectorizing loops and // their epilogues. class EpilogueVectorizerEpilogueLoop : public InnerLoopAndEpilogueVectorizer { public: EpilogueVectorizerEpilogueLoop( Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, const TargetTransformInfo *TTI, AssumptionCache *AC, EpilogueLoopVectorizationInfo &EPI, LoopVectorizationCostModel *CM, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, GeneratedRTChecks &Checks, VPlan &Plan) : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TTI, AC, EPI, CM, BFI, PSI, Checks, Plan, EPI.EpilogueVF, EPI.EpilogueVF, EPI.EpilogueUF) { TripCount = EPI.TripCount; } /// Implements the interface for creating a vectorized skeleton using the /// *epilogue loop* strategy (ie the second pass of vplan execution). BasicBlock *createEpilogueVectorizedLoopSkeleton() final; protected: /// Emits an iteration count bypass check after the main vector loop has /// finished to see if there are any iterations left to execute by either /// the vector epilogue or the scalar epilogue. BasicBlock *emitMinimumVectorEpilogueIterCountCheck( BasicBlock *Bypass, BasicBlock *Insert); void printDebugTracesAtStart() override; void printDebugTracesAtEnd() override; }; } // end namespace llvm /// Look for a meaningful debug location on the instruction or its operands. static DebugLoc getDebugLocFromInstOrOperands(Instruction *I) { if (!I) return DebugLoc::getUnknown(); DebugLoc Empty; if (I->getDebugLoc() != Empty) return I->getDebugLoc(); for (Use &Op : I->operands()) { if (Instruction *OpInst = dyn_cast(Op)) if (OpInst->getDebugLoc() != Empty) return OpInst->getDebugLoc(); } return I->getDebugLoc(); } /// Write a \p DebugMsg about vectorization to the debug output stream. If \p I /// is passed, the message relates to that particular instruction. #ifndef NDEBUG static void debugVectorizationMessage(const StringRef Prefix, const StringRef DebugMsg, Instruction *I) { dbgs() << "LV: " << Prefix << DebugMsg; if (I != nullptr) dbgs() << " " << *I; else dbgs() << '.'; dbgs() << '\n'; } #endif /// Create an analysis remark that explains why vectorization failed /// /// \p PassName is the name of the pass (e.g. can be AlwaysPrint). \p /// RemarkName is the identifier for the remark. If \p I is passed it is an /// instruction that prevents vectorization. Otherwise \p TheLoop is used for /// the location of the remark. If \p DL is passed, use it as debug location for /// the remark. \return the remark object that can be streamed to. static OptimizationRemarkAnalysis createLVAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop, Instruction *I, DebugLoc DL = {}) { BasicBlock *CodeRegion = I ? I->getParent() : TheLoop->getHeader(); // If debug location is attached to the instruction, use it. Otherwise if DL // was not provided, use the loop's. if (I && I->getDebugLoc()) DL = I->getDebugLoc(); else if (!DL) DL = TheLoop->getStartLoc(); return OptimizationRemarkAnalysis(PassName, RemarkName, DL, CodeRegion); } namespace llvm { /// Return a value for Step multiplied by VF. Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF, int64_t Step) { assert(Ty->isIntegerTy() && "Expected an integer step"); return B.CreateElementCount(Ty, VF.multiplyCoefficientBy(Step)); } /// Return the runtime value for VF. Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF) { return B.CreateElementCount(Ty, VF); } void reportVectorizationFailure(const StringRef DebugMsg, const StringRef OREMsg, const StringRef ORETag, OptimizationRemarkEmitter *ORE, Loop *TheLoop, Instruction *I) { LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: ", DebugMsg, I)); LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); ORE->emit( createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I) << "loop not vectorized: " << OREMsg); } /// Reports an informative message: print \p Msg for debugging purposes as well /// as an optimization remark. Uses either \p I as location of the remark, or /// otherwise \p TheLoop. If \p DL is passed, use it as debug location for the /// remark. If \p DL is passed, use it as debug location for the remark. static void reportVectorizationInfo(const StringRef Msg, const StringRef ORETag, OptimizationRemarkEmitter *ORE, Loop *TheLoop, Instruction *I = nullptr, DebugLoc DL = {}) { LLVM_DEBUG(debugVectorizationMessage("", Msg, I)); LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE); ORE->emit(createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I, DL) << Msg); } /// Report successful vectorization of the loop. In case an outer loop is /// vectorized, prepend "outer" to the vectorization remark. static void reportVectorization(OptimizationRemarkEmitter *ORE, Loop *TheLoop, VectorizationFactor VF, unsigned IC) { LLVM_DEBUG(debugVectorizationMessage( "Vectorizing: ", TheLoop->isInnermost() ? "innermost loop" : "outer loop", nullptr)); StringRef LoopType = TheLoop->isInnermost() ? "" : "outer "; ORE->emit([&]() { return OptimizationRemark(LV_NAME, "Vectorized", TheLoop->getStartLoc(), TheLoop->getHeader()) << "vectorized " << LoopType << "loop (vectorization width: " << ore::NV("VectorizationFactor", VF.Width) << ", interleaved count: " << ore::NV("InterleaveCount", IC) << ")"; }); } } // end namespace llvm namespace llvm { // Loop vectorization cost-model hints how the scalar epilogue loop should be // lowered. enum ScalarEpilogueLowering { // The default: allowing scalar epilogues. CM_ScalarEpilogueAllowed, // Vectorization with OptForSize: don't allow epilogues. CM_ScalarEpilogueNotAllowedOptSize, // A special case of vectorisation with OptForSize: loops with a very small // trip count are considered for vectorization under OptForSize, thereby // making sure the cost of their loop body is dominant, free of runtime // guards and scalar iteration overheads. CM_ScalarEpilogueNotAllowedLowTripLoop, // Loop hint predicate indicating an epilogue is undesired. CM_ScalarEpilogueNotNeededUsePredicate, // Directive indicating we must either tail fold or not vectorize CM_ScalarEpilogueNotAllowedUsePredicate }; /// 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. class LoopVectorizationCostModel { friend class LoopVectorizationPlanner; public: LoopVectorizationCostModel(ScalarEpilogueLowering SEL, 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, InterleavedAccessInfo &IAI, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI) : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal), TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F), Hints(Hints), InterleaveInfo(IAI) { if (TTI.supportsScalableVectors() || ForceTargetSupportsScalableVectors) initializeVScaleForTuning(); CostKind = F->hasMinSize() ? TTI::TCK_CodeSize : TTI::TCK_RecipThroughput; // Query this against the original loop and save it here because the profile // of the original loop header may change as the transformation happens. OptForSize = llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI, PGSOQueryType::IRPass); } /// \return An upper bound for the vectorization factors (both fixed and /// scalable). If the factors are 0, vectorization and interleaving should be /// avoided up front. FixedScalableVFPair computeMaxVF(ElementCount UserVF, unsigned UserIC); /// \return True if runtime checks are required for vectorization, and false /// otherwise. bool runtimeChecksRequired(); /// Setup cost-based decisions for user vectorization factor. /// \return true if the UserVF is a feasible VF to be chosen. bool selectUserVectorizationFactor(ElementCount UserVF) { collectNonVectorizedAndSetWideningDecisions(UserVF); return expectedCost(UserVF).isValid(); } /// \return True if maximizing vector bandwidth is enabled by the target or /// user options, for the given register kind. bool useMaxBandwidth(TargetTransformInfo::RegisterKind RegKind); /// \return True if register pressure should be calculated for the given VF. bool shouldCalculateRegPressureForVF(ElementCount VF); /// \return The size (in bits) of the smallest and widest types in the code /// that needs to be vectorized. We ignore values that remain scalar such as /// 64 bit loop indices. std::pair getSmallestAndWidestTypes(); /// 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. void setCostBasedWideningDecision(ElementCount VF); /// A call may be vectorized in different ways depending on whether we have /// vectorized variants available and whether the target supports masking. /// This function analyzes all calls in the function at the supplied VF, /// makes a decision based on the costs of available options, and stores that /// decision in a map for use in planning and plan execution. void setVectorizedCallDecision(ElementCount VF); /// Collect values we want to ignore in the cost model. void collectValuesToIgnore(); /// Collect all element types in the loop for which widening is needed. void collectElementTypesForWidening(); /// Split reductions into those that happen in the loop, and those that happen /// outside. In loop reductions are collected into InLoopReductions. void collectInLoopReductions(); /// Returns true if we should use strict in-order reductions for the given /// RdxDesc. This is true if the -enable-strict-reductions flag is passed, /// the IsOrdered flag of RdxDesc is set and we do not allow reordering /// of FP operations. bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) const { return !Hints->allowReordering() && RdxDesc.isOrdered(); } /// \returns The smallest bitwidth each instruction can be represented with. /// The vector equivalents of these instructions should be truncated to this /// type. const MapVector &getMinimalBitwidths() const { return MinBWs; } /// \returns True if it is more profitable to scalarize instruction \p I for /// vectorization factor \p VF. bool isProfitableToScalarize(Instruction *I, ElementCount VF) const { assert(VF.isVector() && "Profitable to scalarize relevant only for VF > 1."); assert( TheLoop->isInnermost() && "cost-model should not be used for outer loops (in VPlan-native path)"); auto Scalars = InstsToScalarize.find(VF); assert(Scalars != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability"); return Scalars->second.contains(I); } /// Returns true if \p I is known to be uniform after vectorization. bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const { assert( TheLoop->isInnermost() && "cost-model should not be used for outer loops (in VPlan-native path)"); // Pseudo probe needs to be duplicated for each unrolled iteration and // vector lane so that profiled loop trip count can be accurately // accumulated instead of being under counted. if (isa(I)) return false; if (VF.isScalar()) return true; auto UniformsPerVF = Uniforms.find(VF); assert(UniformsPerVF != Uniforms.end() && "VF not yet analyzed for uniformity"); return UniformsPerVF->second.count(I); } /// Returns true if \p I is known to be scalar after vectorization. bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const { assert( TheLoop->isInnermost() && "cost-model should not be used for outer loops (in VPlan-native path)"); if (VF.isScalar()) return true; auto ScalarsPerVF = Scalars.find(VF); assert(ScalarsPerVF != Scalars.end() && "Scalar values are not calculated for VF"); return ScalarsPerVF->second.count(I); } /// \returns True if instruction \p I can be truncated to a smaller bitwidth /// for vectorization factor \p VF. bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const { return VF.isVector() && MinBWs.contains(I) && !isProfitableToScalarize(I, VF) && !isScalarAfterVectorization(I, VF); } /// Decision that was taken during cost calculation for memory instruction. enum InstWidening { CM_Unknown, CM_Widen, // For consecutive accesses with stride +1. CM_Widen_Reverse, // For consecutive accesses with stride -1. CM_Interleave, CM_GatherScatter, CM_Scalarize, CM_VectorCall, CM_IntrinsicCall }; /// Save vectorization decision \p W and \p Cost taken by the cost model for /// instruction \p I and vector width \p VF. void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W, InstructionCost Cost) { assert(VF.isVector() && "Expected VF >=2"); WideningDecisions[{I, VF}] = {W, Cost}; } /// Save vectorization decision \p W and \p Cost taken by the cost model for /// interleaving group \p Grp and vector width \p VF. void setWideningDecision(const InterleaveGroup *Grp, ElementCount VF, InstWidening W, InstructionCost Cost) { assert(VF.isVector() && "Expected VF >=2"); /// Broadcast this decicion to all instructions inside the group. /// When interleaving, the cost will only be assigned one instruction, the /// insert position. For other cases, add the appropriate fraction of the /// total cost to each instruction. This ensures accurate costs are used, /// even if the insert position instruction is not used. InstructionCost InsertPosCost = Cost; InstructionCost OtherMemberCost = 0; if (W != CM_Interleave) OtherMemberCost = InsertPosCost = Cost / Grp->getNumMembers(); ; for (unsigned Idx = 0; Idx < Grp->getFactor(); ++Idx) { if (auto *I = Grp->getMember(Idx)) { if (Grp->getInsertPos() == I) WideningDecisions[{I, VF}] = {W, InsertPosCost}; else WideningDecisions[{I, VF}] = {W, OtherMemberCost}; } } } /// Return the cost model decision for the given instruction \p I and vector /// width \p VF. Return CM_Unknown if this instruction did not pass /// through the cost modeling. InstWidening getWideningDecision(Instruction *I, ElementCount VF) const { assert(VF.isVector() && "Expected VF to be a vector VF"); assert( TheLoop->isInnermost() && "cost-model should not be used for outer loops (in VPlan-native path)"); std::pair InstOnVF(I, VF); auto Itr = WideningDecisions.find(InstOnVF); if (Itr == WideningDecisions.end()) return CM_Unknown; return Itr->second.first; } /// Return the vectorization cost for the given instruction \p I and vector /// width \p VF. InstructionCost getWideningCost(Instruction *I, ElementCount VF) { assert(VF.isVector() && "Expected VF >=2"); std::pair InstOnVF(I, VF); assert(WideningDecisions.contains(InstOnVF) && "The cost is not calculated"); return WideningDecisions[InstOnVF].second; } struct CallWideningDecision { InstWidening Kind; Function *Variant; Intrinsic::ID IID; std::optional MaskPos; InstructionCost Cost; }; void setCallWideningDecision(CallInst *CI, ElementCount VF, InstWidening Kind, Function *Variant, Intrinsic::ID IID, std::optional MaskPos, InstructionCost Cost) { assert(!VF.isScalar() && "Expected vector VF"); CallWideningDecisions[{CI, VF}] = {Kind, Variant, IID, MaskPos, Cost}; } CallWideningDecision getCallWideningDecision(CallInst *CI, ElementCount VF) const { assert(!VF.isScalar() && "Expected vector VF"); return CallWideningDecisions.at({CI, VF}); } /// Return True if instruction \p 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. bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) { // If the instruction is not a truncate, return false. auto *Trunc = dyn_cast(I); if (!Trunc) return false; // Get the source and destination types of the truncate. Type *SrcTy = toVectorTy(Trunc->getSrcTy(), VF); Type *DestTy = toVectorTy(Trunc->getDestTy(), VF); // If the truncate is free for the given types, return false. Replacing a // free truncate with an induction variable would add an induction variable // update instruction to each iteration of the loop. We exclude from this // check the primary induction variable since it will need an update // instruction regardless. Value *Op = Trunc->getOperand(0); if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy)) return false; // If the truncated value is not an induction variable, return false. return Legal->isInductionPhi(Op); } /// Collects the instructions to scalarize for each predicated instruction in /// the loop. void collectInstsToScalarize(ElementCount VF); /// Collect values that will not be widened, including Uniforms, Scalars, and /// Instructions to Scalarize for the given \p VF. /// The sets depend on CM decision for Load/Store instructions /// that may be vectorized as interleave, gather-scatter or scalarized. /// Also make a decision on what to do about call instructions in the loop /// at that VF -- scalarize, call a known vector routine, or call a /// vector intrinsic. void collectNonVectorizedAndSetWideningDecisions(ElementCount VF) { // Do the analysis once. if (VF.isScalar() || Uniforms.contains(VF)) return; setCostBasedWideningDecision(VF); collectLoopUniforms(VF); setVectorizedCallDecision(VF); collectLoopScalars(VF); collectInstsToScalarize(VF); } /// Returns true if the target machine supports masked store operation /// for the given \p DataType and kind of access to \p Ptr. bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment, unsigned AddressSpace) const { return Legal->isConsecutivePtr(DataType, Ptr) && TTI.isLegalMaskedStore(DataType, Alignment, AddressSpace); } /// Returns true if the target machine supports masked load operation /// for the given \p DataType and kind of access to \p Ptr. bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment, unsigned AddressSpace) const { return Legal->isConsecutivePtr(DataType, Ptr) && TTI.isLegalMaskedLoad(DataType, Alignment, AddressSpace); } /// Returns true if the target machine can represent \p V as a masked gather /// or scatter operation. bool isLegalGatherOrScatter(Value *V, ElementCount VF) { bool LI = isa(V); bool SI = isa(V); if (!LI && !SI) return false; auto *Ty = getLoadStoreType(V); Align Align = getLoadStoreAlignment(V); if (VF.isVector()) Ty = VectorType::get(Ty, VF); return (LI && TTI.isLegalMaskedGather(Ty, Align)) || (SI && TTI.isLegalMaskedScatter(Ty, Align)); } /// Returns true if the target machine supports all of the reduction /// variables found for the given VF. bool canVectorizeReductions(ElementCount VF) const { return (all_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool { const RecurrenceDescriptor &RdxDesc = Reduction.second; return TTI.isLegalToVectorizeReduction(RdxDesc, VF); })); } /// Given costs for both strategies, return true if the scalar predication /// lowering should be used for div/rem. This incorporates an override /// option so it is not simply a cost comparison. bool isDivRemScalarWithPredication(InstructionCost ScalarCost, InstructionCost SafeDivisorCost) const { switch (ForceSafeDivisor) { case cl::BOU_UNSET: return ScalarCost < SafeDivisorCost; case cl::BOU_TRUE: return false; case cl::BOU_FALSE: return true; } llvm_unreachable("impossible case value"); } /// Returns true if \p I is an instruction which requires predication and /// for which our chosen predication strategy is scalarization (i.e. we /// don't have an alternate strategy such as masking available). /// \p VF is the vectorization factor that will be used to vectorize \p I. bool isScalarWithPredication(Instruction *I, ElementCount VF) const; /// Returns true if \p I is an instruction that needs to be predicated /// at runtime. The result is independent of the predication mechanism. /// Superset of instructions that return true for isScalarWithPredication. bool isPredicatedInst(Instruction *I) const; /// Return the costs for our two available strategies for lowering a /// div/rem operation which requires speculating at least one lane. /// First result is for scalarization (will be invalid for scalable /// vectors); second is for the safe-divisor strategy. std::pair getDivRemSpeculationCost(Instruction *I, ElementCount VF) const; /// Returns true if \p I is a memory instruction with consecutive memory /// access that can be widened. bool memoryInstructionCanBeWidened(Instruction *I, ElementCount VF); /// Returns true if \p I is a memory instruction in an interleaved-group /// of memory accesses that can be vectorized with wide vector loads/stores /// and shuffles. bool interleavedAccessCanBeWidened(Instruction *I, ElementCount VF) const; /// Check if \p Instr belongs to any interleaved access group. bool isAccessInterleaved(Instruction *Instr) const { return InterleaveInfo.isInterleaved(Instr); } /// Get the interleaved access group that \p Instr belongs to. const InterleaveGroup * getInterleavedAccessGroup(Instruction *Instr) const { return InterleaveInfo.getInterleaveGroup(Instr); } /// Returns true if we're required to use a scalar epilogue for at least /// the final iteration of the original loop. bool requiresScalarEpilogue(bool IsVectorizing) const { if (!isScalarEpilogueAllowed()) { LLVM_DEBUG(dbgs() << "LV: Loop does not require scalar epilogue\n"); return false; } // If we might exit from anywhere but the latch and early exit vectorization // is disabled, we must run the exiting iteration in scalar form. if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch() && !(EnableEarlyExitVectorization && Legal->hasUncountableEarlyExit())) { LLVM_DEBUG(dbgs() << "LV: Loop requires scalar epilogue: not exiting " "from latch block\n"); return true; } if (IsVectorizing && InterleaveInfo.requiresScalarEpilogue()) { LLVM_DEBUG(dbgs() << "LV: Loop requires scalar epilogue: " "interleaved group requires scalar epilogue\n"); return true; } LLVM_DEBUG(dbgs() << "LV: Loop does not require scalar epilogue\n"); return false; } /// Returns true if a scalar epilogue is not allowed due to optsize or a /// loop hint annotation. bool isScalarEpilogueAllowed() const { return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed; } /// Returns the TailFoldingStyle that is best for the current loop. TailFoldingStyle getTailFoldingStyle(bool IVUpdateMayOverflow = true) const { if (!ChosenTailFoldingStyle) return TailFoldingStyle::None; return IVUpdateMayOverflow ? ChosenTailFoldingStyle->first : ChosenTailFoldingStyle->second; } /// Selects and saves TailFoldingStyle for 2 options - if IV update may /// overflow or not. /// \param IsScalableVF true if scalable vector factors enabled. /// \param UserIC User specific interleave count. void setTailFoldingStyles(bool IsScalableVF, unsigned UserIC) { assert(!ChosenTailFoldingStyle && "Tail folding must not be selected yet."); if (!Legal->canFoldTailByMasking()) { ChosenTailFoldingStyle = {TailFoldingStyle::None, TailFoldingStyle::None}; return; } // Default to TTI preference, but allow command line override. ChosenTailFoldingStyle = { TTI.getPreferredTailFoldingStyle(/*IVUpdateMayOverflow=*/true), TTI.getPreferredTailFoldingStyle(/*IVUpdateMayOverflow=*/false)}; if (ForceTailFoldingStyle.getNumOccurrences()) ChosenTailFoldingStyle = {ForceTailFoldingStyle.getValue(), ForceTailFoldingStyle.getValue()}; if (ChosenTailFoldingStyle->first != TailFoldingStyle::DataWithEVL && ChosenTailFoldingStyle->second != TailFoldingStyle::DataWithEVL) return; // Override EVL styles if needed. // FIXME: Investigate opportunity for fixed vector factor. bool EVLIsLegal = UserIC <= 1 && IsScalableVF && TTI.hasActiveVectorLength() && !EnableVPlanNativePath; if (EVLIsLegal) return; // If for some reason EVL mode is unsupported, fallback to a scalar epilogue // if it's allowed, or DataWithoutLaneMask otherwise. if (ScalarEpilogueStatus == CM_ScalarEpilogueAllowed || ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) ChosenTailFoldingStyle = {TailFoldingStyle::None, TailFoldingStyle::None}; else ChosenTailFoldingStyle = {TailFoldingStyle::DataWithoutLaneMask, TailFoldingStyle::DataWithoutLaneMask}; LLVM_DEBUG( dbgs() << "LV: Preference for VP intrinsics indicated. Will " "not try to generate VP Intrinsics " << (UserIC > 1 ? "since interleave count specified is greater than 1.\n" : "due to non-interleaving reasons.\n")); } /// Returns true if all loop blocks should be masked to fold tail loop. bool foldTailByMasking() const { // TODO: check if it is possible to check for None style independent of // IVUpdateMayOverflow flag in getTailFoldingStyle. return getTailFoldingStyle() != TailFoldingStyle::None; } /// Return maximum safe number of elements to be processed per vector /// iteration, which do not prevent store-load forwarding and are safe with /// regard to the memory dependencies. Required for EVL-based VPlans to /// correctly calculate AVL (application vector length) as min(remaining AVL, /// MaxSafeElements). /// TODO: need to consider adjusting cost model to use this value as a /// vectorization factor for EVL-based vectorization. std::optional getMaxSafeElements() const { return MaxSafeElements; } /// Returns true if the instructions in this block requires predication /// for any reason, e.g. because tail folding now requires a predicate /// or because the block in the original loop was predicated. bool blockNeedsPredicationForAnyReason(BasicBlock *BB) const { return foldTailByMasking() || Legal->blockNeedsPredication(BB); } /// Returns true if VP intrinsics with explicit vector length support should /// be generated in the tail folded loop. bool foldTailWithEVL() const { return getTailFoldingStyle() == TailFoldingStyle::DataWithEVL; } /// Returns true if the Phi is part of an inloop reduction. bool isInLoopReduction(PHINode *Phi) const { return InLoopReductions.contains(Phi); } /// Returns true if the predicated reduction select should be used to set the /// incoming value for the reduction phi. bool usePredicatedReductionSelect() const { // Force to use predicated reduction select since the EVL of the // second-to-last iteration might not be VF*UF. if (foldTailWithEVL()) return true; return PreferPredicatedReductionSelect || TTI.preferPredicatedReductionSelect(); } /// Estimate cost of an intrinsic call instruction CI if it were vectorized /// with factor VF. Return the cost of the instruction, including /// scalarization overhead if it's needed. InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const; /// Estimate cost of a call instruction CI if it were vectorized with factor /// VF. Return the cost of the instruction, including scalarization overhead /// if it's needed. InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF) const; /// Invalidates decisions already taken by the cost model. void invalidateCostModelingDecisions() { WideningDecisions.clear(); CallWideningDecisions.clear(); Uniforms.clear(); Scalars.clear(); } /// Returns the expected execution cost. The unit of the cost does /// not matter because we use the 'cost' units to compare different /// vector widths. The cost that is returned is *not* normalized by /// the factor width. InstructionCost expectedCost(ElementCount VF); bool hasPredStores() const { return NumPredStores > 0; } /// Returns true if epilogue vectorization is considered profitable, and /// false otherwise. /// \p VF is the vectorization factor chosen for the original loop. /// \p Multiplier is an aditional scaling factor applied to VF before /// comparing to EpilogueVectorizationMinVF. bool isEpilogueVectorizationProfitable(const ElementCount VF, const unsigned IC) const; /// Returns the execution time cost of an instruction for a given vector /// width. Vector width of one means scalar. InstructionCost getInstructionCost(Instruction *I, ElementCount VF); /// Return the cost of instructions in an inloop reduction pattern, if I is /// part of that pattern. std::optional getReductionPatternCost(Instruction *I, ElementCount VF, Type *VectorTy) const; /// Returns true if \p Op should be considered invariant and if it is /// trivially hoistable. bool shouldConsiderInvariant(Value *Op); /// Return the value of vscale used for tuning the cost model. std::optional getVScaleForTuning() const { return VScaleForTuning; } private: unsigned NumPredStores = 0; /// Used to store the value of vscale used for tuning the cost model. It is /// initialized during object construction. std::optional VScaleForTuning; /// Initializes the value of vscale used for tuning the cost model. If /// vscale_range.min == vscale_range.max then return vscale_range.max, else /// return the value returned by the corresponding TTI method. void initializeVScaleForTuning() { const Function *Fn = TheLoop->getHeader()->getParent(); if (Fn->hasFnAttribute(Attribute::VScaleRange)) { auto Attr = Fn->getFnAttribute(Attribute::VScaleRange); auto Min = Attr.getVScaleRangeMin(); auto Max = Attr.getVScaleRangeMax(); if (Max && Min == Max) { VScaleForTuning = Max; return; } } VScaleForTuning = TTI.getVScaleForTuning(); } /// \return An upper bound for the vectorization factors for both /// fixed and scalable vectorization, where the minimum-known number of /// elements is a power-of-2 larger than zero. If scalable vectorization is /// disabled or unsupported, then the scalable part will be equal to /// ElementCount::getScalable(0). FixedScalableVFPair computeFeasibleMaxVF(unsigned MaxTripCount, ElementCount UserVF, bool FoldTailByMasking); /// If \p VF > MaxTripcount, clamps it to the next lower VF that is <= /// MaxTripCount. ElementCount clampVFByMaxTripCount(ElementCount VF, unsigned MaxTripCount, bool FoldTailByMasking) const; /// \return the maximized element count based on the targets vector /// registers and the loop trip-count, but limited to a maximum safe VF. /// This is a helper function of computeFeasibleMaxVF. ElementCount getMaximizedVFForTarget(unsigned MaxTripCount, unsigned SmallestType, unsigned WidestType, ElementCount MaxSafeVF, bool FoldTailByMasking); /// Checks if scalable vectorization is supported and enabled. Caches the /// result to avoid repeated debug dumps for repeated queries. bool isScalableVectorizationAllowed(); /// \return the maximum legal scalable VF, based on the safe max number /// of elements. ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements); /// Calculate vectorization cost of memory instruction \p I. InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF); /// The cost computation for scalarized memory instruction. InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF); /// The cost computation for interleaving group of memory instructions. InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF); /// The cost computation for Gather/Scatter instruction. InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF); /// The cost computation for widening instruction \p I with consecutive /// memory access. InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF); /// The cost calculation for Load/Store instruction \p I with uniform pointer - /// Load: scalar load + broadcast. /// Store: scalar store + (loop invariant value stored? 0 : extract of last /// element) InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF); /// Estimate the overhead of scalarizing an instruction. This is a /// convenience wrapper for the type-based getScalarizationOverhead API. InstructionCost getScalarizationOverhead(Instruction *I, ElementCount VF) const; /// Returns true if an artificially high cost for emulated masked memrefs /// should be used. bool useEmulatedMaskMemRefHack(Instruction *I, ElementCount VF); /// Map of scalar integer values to the smallest bitwidth they can be legally /// represented as. The vector equivalents of these values should be truncated /// to this type. MapVector MinBWs; /// A type representing the costs for instructions if they were to be /// scalarized rather than vectorized. The entries are Instruction-Cost /// pairs. using ScalarCostsTy = MapVector; /// A set containing all BasicBlocks that are known to present after /// vectorization as a predicated block. DenseMap> PredicatedBBsAfterVectorization; /// Records whether it is allowed to have the original scalar loop execute at /// least once. This may be needed as a fallback loop in case runtime /// aliasing/dependence checks fail, or to handle the tail/remainder /// iterations when the trip count is unknown or doesn't divide by the VF, /// or as a peel-loop to handle gaps in interleave-groups. /// Under optsize and when the trip count is very small we don't allow any /// iterations to execute in the scalar loop. ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; /// Control finally chosen tail folding style. The first element is used if /// the IV update may overflow, the second element - if it does not. std::optional> ChosenTailFoldingStyle; /// true if scalable vectorization is supported and enabled. std::optional IsScalableVectorizationAllowed; /// Maximum safe number of elements to be processed per vector iteration, /// which do not prevent store-load forwarding and are safe with regard to the /// memory dependencies. Required for EVL-based veectorization, where this /// value is used as the upper bound of the safe AVL. std::optional MaxSafeElements; /// A map holding scalar costs for different vectorization factors. The /// presence of a cost for an instruction in the mapping indicates that the /// instruction will be scalarized when vectorizing with the associated /// vectorization factor. The entries are VF-ScalarCostTy pairs. MapVector InstsToScalarize; /// Holds the instructions known to be uniform after vectorization. /// The data is collected per VF. DenseMap> Uniforms; /// Holds the instructions known to be scalar after vectorization. /// The data is collected per VF. DenseMap> Scalars; /// Holds the instructions (address computations) that are forced to be /// scalarized. DenseMap> ForcedScalars; /// PHINodes of the reductions that should be expanded in-loop. SmallPtrSet InLoopReductions; /// A Map of inloop reduction operations and their immediate chain operand. /// FIXME: This can be removed once reductions can be costed correctly in /// VPlan. This was added to allow quick lookup of the inloop operations. DenseMap InLoopReductionImmediateChains; /// Returns the expected difference in cost from scalarizing the expression /// feeding a predicated instruction \p PredInst. The instructions to /// scalarize and their scalar costs are collected in \p ScalarCosts. A /// non-negative return value implies the expression will be scalarized. /// Currently, only single-use chains are considered for scalarization. InstructionCost computePredInstDiscount(Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF); /// Collect the instructions that are uniform after vectorization. An /// instruction is uniform if we represent it with a single scalar value in /// the vectorized loop corresponding to each vector iteration. Examples of /// uniform instructions include pointer operands of consecutive or /// interleaved memory accesses. Note that although uniformity implies an /// instruction will be scalar, the reverse is not true. In general, a /// scalarized instruction will be represented by VF scalar values in the /// vectorized loop, each corresponding to an iteration of the original /// scalar loop. void collectLoopUniforms(ElementCount VF); /// Collect the instructions that are scalar after vectorization. An /// instruction is scalar if it is known to be uniform or will be scalarized /// during vectorization. collectLoopScalars should only add non-uniform nodes /// to the list if they are used by a load/store instruction that is marked as /// CM_Scalarize. Non-uniform scalarized instructions will be represented by /// VF values in the vectorized loop, each corresponding to an iteration of /// the original scalar loop. void collectLoopScalars(ElementCount VF); /// Keeps cost model vectorization decision and cost for instructions. /// Right now it is used for memory instructions only. using DecisionList = DenseMap, std::pair>; DecisionList WideningDecisions; using CallDecisionList = DenseMap, CallWideningDecision>; CallDecisionList CallWideningDecisions; /// Returns true if \p V is expected to be vectorized and it needs to be /// extracted. bool needsExtract(Value *V, ElementCount VF) const { Instruction *I = dyn_cast(V); if (VF.isScalar() || !I || !TheLoop->contains(I) || TheLoop->isLoopInvariant(I) || getWideningDecision(I, VF) == CM_Scalarize) return false; // Assume we can vectorize V (and hence we need extraction) if the // scalars are not computed yet. This can happen, because it is called // via getScalarizationOverhead from setCostBasedWideningDecision, before // the scalars are collected. That should be a safe assumption in most // cases, because we check if the operands have vectorizable types // beforehand in LoopVectorizationLegality. return !Scalars.contains(VF) || !isScalarAfterVectorization(I, VF); }; /// Returns a range containing only operands needing to be extracted. SmallVector filterExtractingOperands(Instruction::op_range Ops, ElementCount VF) const { return SmallVector(make_filter_range( Ops, [this, VF](Value *V) { return this->needsExtract(V, VF); })); } public: /// The loop that we evaluate. Loop *TheLoop; /// Predicated scalar evolution analysis. PredicatedScalarEvolution &PSE; /// Loop Info analysis. LoopInfo *LI; /// Vectorization legality. LoopVectorizationLegality *Legal; /// Vector target information. const TargetTransformInfo &TTI; /// Target Library Info. const TargetLibraryInfo *TLI; /// Demanded bits analysis. DemandedBits *DB; /// Assumption cache. AssumptionCache *AC; /// Interface to emit optimization remarks. OptimizationRemarkEmitter *ORE; const Function *TheFunction; /// Loop Vectorize Hint. const LoopVectorizeHints *Hints; /// The interleave access information contains groups of interleaved accesses /// with the same stride and close to each other. InterleavedAccessInfo &InterleaveInfo; /// Values to ignore in the cost model. SmallPtrSet ValuesToIgnore; /// Values to ignore in the cost model when VF > 1. SmallPtrSet VecValuesToIgnore; /// All element types found in the loop. SmallPtrSet ElementTypesInLoop; /// The kind of cost that we are calculating TTI::TargetCostKind CostKind; /// Whether this loop should be optimized for size based on function attribute /// or profile information. bool OptForSize; /// The highest VF possible for this loop, without using MaxBandwidth. FixedScalableVFPair MaxPermissibleVFWithoutMaxBW; }; } // end namespace llvm namespace { /// Helper struct to manage generating runtime checks for vectorization. /// /// The runtime checks are created up-front in temporary blocks to allow better /// estimating the cost and un-linked from the existing IR. After deciding to /// vectorize, the checks are moved back. If deciding not to vectorize, the /// temporary blocks are completely removed. class GeneratedRTChecks { /// Basic block which contains the generated SCEV checks, if any. BasicBlock *SCEVCheckBlock = nullptr; /// The value representing the result of the generated SCEV checks. If it is /// nullptr no SCEV checks have been generated. Value *SCEVCheckCond = nullptr; /// Basic block which contains the generated memory runtime checks, if any. BasicBlock *MemCheckBlock = nullptr; /// The value representing the result of the generated memory runtime checks. /// If it is nullptr no memory runtime checks have been generated. Value *MemRuntimeCheckCond = nullptr; DominatorTree *DT; LoopInfo *LI; TargetTransformInfo *TTI; SCEVExpander SCEVExp; SCEVExpander MemCheckExp; bool CostTooHigh = false; Loop *OuterLoop = nullptr; PredicatedScalarEvolution &PSE; /// The kind of cost that we are calculating TTI::TargetCostKind CostKind; public: GeneratedRTChecks(PredicatedScalarEvolution &PSE, DominatorTree *DT, LoopInfo *LI, TargetTransformInfo *TTI, const DataLayout &DL, TTI::TargetCostKind CostKind) : DT(DT), LI(LI), TTI(TTI), SCEVExp(*PSE.getSE(), DL, "scev.check"), MemCheckExp(*PSE.getSE(), DL, "scev.check"), PSE(PSE), CostKind(CostKind) {} /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can /// accurately estimate the cost of the runtime checks. The blocks are /// un-linked from the IR and are added back during vector code generation. If /// there is no vector code generation, the check blocks are removed /// completely. void create(Loop *L, const LoopAccessInfo &LAI, const SCEVPredicate &UnionPred, ElementCount VF, unsigned IC) { // Hard cutoff to limit compile-time increase in case a very large number of // runtime checks needs to be generated. // TODO: Skip cutoff if the loop is guaranteed to execute, e.g. due to // profile info. CostTooHigh = LAI.getNumRuntimePointerChecks() > VectorizeMemoryCheckThreshold; if (CostTooHigh) return; BasicBlock *LoopHeader = L->getHeader(); BasicBlock *Preheader = L->getLoopPreheader(); // Use SplitBlock to create blocks for SCEV & memory runtime checks to // ensure the blocks are properly added to LoopInfo & DominatorTree. Those // may be used by SCEVExpander. The blocks will be un-linked from their // predecessors and removed from LI & DT at the end of the function. if (!UnionPred.isAlwaysTrue()) { SCEVCheckBlock = SplitBlock(Preheader, Preheader->getTerminator(), DT, LI, nullptr, "vector.scevcheck"); SCEVCheckCond = SCEVExp.expandCodeForPredicate( &UnionPred, SCEVCheckBlock->getTerminator()); if (isa(SCEVCheckCond)) { // Clean up directly after expanding the predicate to a constant, to // avoid further expansions re-using anything left over from SCEVExp. SCEVExpanderCleaner SCEVCleaner(SCEVExp); SCEVCleaner.cleanup(); } } const auto &RtPtrChecking = *LAI.getRuntimePointerChecking(); if (RtPtrChecking.Need) { auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader; MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr, "vector.memcheck"); auto DiffChecks = RtPtrChecking.getDiffChecks(); if (DiffChecks) { Value *RuntimeVF = nullptr; MemRuntimeCheckCond = addDiffRuntimeChecks( MemCheckBlock->getTerminator(), *DiffChecks, MemCheckExp, [VF, &RuntimeVF](IRBuilderBase &B, unsigned Bits) { if (!RuntimeVF) RuntimeVF = getRuntimeVF(B, B.getIntNTy(Bits), VF); return RuntimeVF; }, IC); } else { MemRuntimeCheckCond = addRuntimeChecks( MemCheckBlock->getTerminator(), L, RtPtrChecking.getChecks(), MemCheckExp, VectorizerParams::HoistRuntimeChecks); } assert(MemRuntimeCheckCond && "no RT checks generated although RtPtrChecking " "claimed checks are required"); } if (!MemCheckBlock && !SCEVCheckBlock) return; // Unhook the temporary block with the checks, update various places // accordingly. if (SCEVCheckBlock) SCEVCheckBlock->replaceAllUsesWith(Preheader); if (MemCheckBlock) MemCheckBlock->replaceAllUsesWith(Preheader); if (SCEVCheckBlock) { SCEVCheckBlock->getTerminator()->moveBefore( Preheader->getTerminator()->getIterator()); auto *UI = new UnreachableInst(Preheader->getContext(), SCEVCheckBlock); UI->setDebugLoc(DebugLoc::getTemporary()); Preheader->getTerminator()->eraseFromParent(); } if (MemCheckBlock) { MemCheckBlock->getTerminator()->moveBefore( Preheader->getTerminator()->getIterator()); auto *UI = new UnreachableInst(Preheader->getContext(), MemCheckBlock); UI->setDebugLoc(DebugLoc::getTemporary()); Preheader->getTerminator()->eraseFromParent(); } DT->changeImmediateDominator(LoopHeader, Preheader); if (MemCheckBlock) { DT->eraseNode(MemCheckBlock); LI->removeBlock(MemCheckBlock); } if (SCEVCheckBlock) { DT->eraseNode(SCEVCheckBlock); LI->removeBlock(SCEVCheckBlock); } // Outer loop is used as part of the later cost calculations. OuterLoop = L->getParentLoop(); } InstructionCost getCost() { if (SCEVCheckBlock || MemCheckBlock) LLVM_DEBUG(dbgs() << "Calculating cost of runtime checks:\n"); if (CostTooHigh) { InstructionCost Cost; Cost.setInvalid(); LLVM_DEBUG(dbgs() << " number of checks exceeded threshold\n"); return Cost; } InstructionCost RTCheckCost = 0; if (SCEVCheckBlock) for (Instruction &I : *SCEVCheckBlock) { if (SCEVCheckBlock->getTerminator() == &I) continue; InstructionCost C = TTI->getInstructionCost(&I, CostKind); LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n"); RTCheckCost += C; } if (MemCheckBlock) { InstructionCost MemCheckCost = 0; for (Instruction &I : *MemCheckBlock) { if (MemCheckBlock->getTerminator() == &I) continue; InstructionCost C = TTI->getInstructionCost(&I, CostKind); LLVM_DEBUG(dbgs() << " " << C << " for " << I << "\n"); MemCheckCost += C; } // If the runtime memory checks are being created inside an outer loop // we should find out if these checks are outer loop invariant. If so, // the checks will likely be hoisted out and so the effective cost will // reduce according to the outer loop trip count. if (OuterLoop) { ScalarEvolution *SE = MemCheckExp.getSE(); // TODO: If profitable, we could refine this further by analysing every // individual memory check, since there could be a mixture of loop // variant and invariant checks that mean the final condition is // variant. const SCEV *Cond = SE->getSCEV(MemRuntimeCheckCond); if (SE->isLoopInvariant(Cond, OuterLoop)) { // It seems reasonable to assume that we can reduce the effective // cost of the checks even when we know nothing about the trip // count. Assume that the outer loop executes at least twice. unsigned BestTripCount = 2; // Get the best known TC estimate. if (auto EstimatedTC = getSmallBestKnownTC( PSE, OuterLoop, /* CanUseConstantMax = */ false)) if (EstimatedTC->isFixed()) BestTripCount = EstimatedTC->getFixedValue(); InstructionCost NewMemCheckCost = MemCheckCost / BestTripCount; // Let's ensure the cost is always at least 1. NewMemCheckCost = std::max(NewMemCheckCost.getValue(), (InstructionCost::CostType)1); if (BestTripCount > 1) LLVM_DEBUG(dbgs() << "We expect runtime memory checks to be hoisted " << "out of the outer loop. Cost reduced from " << MemCheckCost << " to " << NewMemCheckCost << '\n'); MemCheckCost = NewMemCheckCost; } } RTCheckCost += MemCheckCost; } if (SCEVCheckBlock || MemCheckBlock) LLVM_DEBUG(dbgs() << "Total cost of runtime checks: " << RTCheckCost << "\n"); return RTCheckCost; } /// Remove the created SCEV & memory runtime check blocks & instructions, if /// unused. ~GeneratedRTChecks() { SCEVExpanderCleaner SCEVCleaner(SCEVExp); SCEVExpanderCleaner MemCheckCleaner(MemCheckExp); bool SCEVChecksUsed = !SCEVCheckBlock || !pred_empty(SCEVCheckBlock); bool MemChecksUsed = !MemCheckBlock || !pred_empty(MemCheckBlock); if (SCEVChecksUsed) SCEVCleaner.markResultUsed(); if (MemChecksUsed) { MemCheckCleaner.markResultUsed(); } else { auto &SE = *MemCheckExp.getSE(); // Memory runtime check generation creates compares that use expanded // values. Remove them before running the SCEVExpanderCleaners. for (auto &I : make_early_inc_range(reverse(*MemCheckBlock))) { if (MemCheckExp.isInsertedInstruction(&I)) continue; SE.forgetValue(&I); I.eraseFromParent(); } } MemCheckCleaner.cleanup(); SCEVCleaner.cleanup(); if (!SCEVChecksUsed) SCEVCheckBlock->eraseFromParent(); if (!MemChecksUsed) MemCheckBlock->eraseFromParent(); } /// Retrieves the SCEVCheckCond and SCEVCheckBlock that were generated as IR /// outside VPlan. std::pair getSCEVChecks() { using namespace llvm::PatternMatch; if (!SCEVCheckCond || match(SCEVCheckCond, m_ZeroInt())) return {nullptr, nullptr}; return {SCEVCheckCond, SCEVCheckBlock}; } /// Retrieves the MemCheckCond and MemCheckBlock that were generated as IR /// outside VPlan. std::pair getMemRuntimeChecks() { using namespace llvm::PatternMatch; if (MemRuntimeCheckCond && match(MemRuntimeCheckCond, m_ZeroInt())) return {nullptr, nullptr}; return {MemRuntimeCheckCond, MemCheckBlock}; } /// Return true if any runtime checks have been added bool hasChecks() const { using namespace llvm::PatternMatch; return (SCEVCheckCond && !match(SCEVCheckCond, m_ZeroInt())) || MemRuntimeCheckCond; } }; } // namespace static bool useActiveLaneMask(TailFoldingStyle Style) { return Style == TailFoldingStyle::Data || Style == TailFoldingStyle::DataAndControlFlow || Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; } static bool useActiveLaneMaskForControlFlow(TailFoldingStyle Style) { return Style == TailFoldingStyle::DataAndControlFlow || Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; } // Return true if \p OuterLp is an outer loop annotated with hints for explicit // vectorization. The loop needs to be annotated with #pragma omp simd // simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the // vector length information is not provided, vectorization is not considered // explicit. Interleave hints are not allowed either. These limitations will be // relaxed in the future. // Please, note that we are currently forced to abuse the pragma 'clang // vectorize' semantics. This pragma provides *auto-vectorization hints* // (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd' // provides *explicit vectorization hints* (LV can bypass legal checks and // assume that vectorization is legal). However, both hints are implemented // using the same metadata (llvm.loop.vectorize, processed by // LoopVectorizeHints). This will be fixed in the future when the native IR // representation for pragma 'omp simd' is introduced. static bool isExplicitVecOuterLoop(Loop *OuterLp, OptimizationRemarkEmitter *ORE) { assert(!OuterLp->isInnermost() && "This is not an outer loop"); LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE); // Only outer loops with an explicit vectorization hint are supported. // Unannotated outer loops are ignored. if (Hints.getForce() == LoopVectorizeHints::FK_Undefined) return false; Function *Fn = OuterLp->getHeader()->getParent(); if (!Hints.allowVectorization(Fn, OuterLp, true /*VectorizeOnlyWhenForced*/)) { LLVM_DEBUG(dbgs() << "LV: Loop hints prevent outer loop vectorization.\n"); return false; } if (Hints.getInterleave() > 1) { // TODO: Interleave support is future work. LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Interleave is not supported for " "outer loops.\n"); Hints.emitRemarkWithHints(); return false; } return true; } static void collectSupportedLoops(Loop &L, LoopInfo *LI, OptimizationRemarkEmitter *ORE, SmallVectorImpl &V) { // Collect inner loops and outer loops without irreducible control flow. For // now, only collect outer loops that have explicit vectorization hints. If we // are stress testing the VPlan H-CFG construction, we collect the outermost // loop of every loop nest. if (L.isInnermost() || VPlanBuildStressTest || (EnableVPlanNativePath && isExplicitVecOuterLoop(&L, ORE))) { LoopBlocksRPO RPOT(&L); RPOT.perform(LI); if (!containsIrreducibleCFG(RPOT, *LI)) { V.push_back(&L); // TODO: Collect inner loops inside marked outer loops in case // vectorization fails for the outer loop. Do not invoke // 'containsIrreducibleCFG' again for inner loops when the outer loop is // already known to be reducible. We can use an inherited attribute for // that. return; } } for (Loop *InnerL : L) collectSupportedLoops(*InnerL, LI, ORE, V); } //===----------------------------------------------------------------------===// // Implementation of LoopVectorizationLegality, InnerLoopVectorizer and // LoopVectorizationCostModel and LoopVectorizationPlanner. //===----------------------------------------------------------------------===// /// Compute the transformed value of Index at offset StartValue using step /// StepValue. /// For integer induction, returns StartValue + Index * StepValue. /// For pointer induction, returns StartValue[Index * StepValue]. /// FIXME: The newly created binary instructions should contain nsw/nuw /// flags, which can be found from the original scalar operations. static Value * emitTransformedIndex(IRBuilderBase &B, Value *Index, Value *StartValue, Value *Step, InductionDescriptor::InductionKind InductionKind, const BinaryOperator *InductionBinOp) { using namespace llvm::PatternMatch; Type *StepTy = Step->getType(); Value *CastedIndex = StepTy->isIntegerTy() ? B.CreateSExtOrTrunc(Index, StepTy) : B.CreateCast(Instruction::SIToFP, Index, StepTy); if (CastedIndex != Index) { CastedIndex->setName(CastedIndex->getName() + ".cast"); Index = CastedIndex; } // Note: the IR at this point is broken. We cannot use SE to create any new // SCEV and then expand it, hoping that SCEV's simplification will give us // a more optimal code. Unfortunately, attempt of doing so on invalid IR may // lead to various SCEV crashes. So all we can do is to use builder and rely // on InstCombine for future simplifications. Here we handle some trivial // cases only. auto CreateAdd = [&B](Value *X, Value *Y) { assert(X->getType() == Y->getType() && "Types don't match!"); if (match(X, m_ZeroInt())) return Y; if (match(Y, m_ZeroInt())) return X; return B.CreateAdd(X, Y); }; // We allow X to be a vector type, in which case Y will potentially be // splatted into a vector with the same element count. auto CreateMul = [&B](Value *X, Value *Y) { assert(X->getType()->getScalarType() == Y->getType() && "Types don't match!"); if (match(X, m_One())) return Y; if (match(Y, m_One())) return X; VectorType *XVTy = dyn_cast(X->getType()); if (XVTy && !isa(Y->getType())) Y = B.CreateVectorSplat(XVTy->getElementCount(), Y); return B.CreateMul(X, Y); }; switch (InductionKind) { case InductionDescriptor::IK_IntInduction: { assert(!isa(Index->getType()) && "Vector indices not supported for integer inductions yet"); assert(Index->getType() == StartValue->getType() && "Index type does not match StartValue type"); if (isa(Step) && cast(Step)->isMinusOne()) return B.CreateSub(StartValue, Index); auto *Offset = CreateMul(Index, Step); return CreateAdd(StartValue, Offset); } case InductionDescriptor::IK_PtrInduction: return B.CreatePtrAdd(StartValue, CreateMul(Index, Step)); case InductionDescriptor::IK_FpInduction: { assert(!isa(Index->getType()) && "Vector indices not supported for FP inductions yet"); assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value"); assert(InductionBinOp && (InductionBinOp->getOpcode() == Instruction::FAdd || InductionBinOp->getOpcode() == Instruction::FSub) && "Original bin op should be defined for FP induction"); Value *MulExp = B.CreateFMul(Step, Index); return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp, "induction"); } case InductionDescriptor::IK_NoInduction: return nullptr; } llvm_unreachable("invalid enum"); } static std::optional getMaxVScale(const Function &F, const TargetTransformInfo &TTI) { if (std::optional MaxVScale = TTI.getMaxVScale()) return MaxVScale; if (F.hasFnAttribute(Attribute::VScaleRange)) return F.getFnAttribute(Attribute::VScaleRange).getVScaleRangeMax(); return std::nullopt; } /// For the given VF and UF and maximum trip count computed for the loop, return /// whether the induction variable might overflow in the vectorized loop. If not, /// then we know a runtime overflow check always evaluates to false and can be /// removed. static bool isIndvarOverflowCheckKnownFalse( const LoopVectorizationCostModel *Cost, ElementCount VF, std::optional UF = std::nullopt) { // Always be conservative if we don't know the exact unroll factor. unsigned MaxUF = UF ? *UF : Cost->TTI.getMaxInterleaveFactor(VF); IntegerType *IdxTy = Cost->Legal->getWidestInductionType(); APInt MaxUIntTripCount = IdxTy->getMask(); // We know the runtime overflow check is known false iff the (max) trip-count // is known and (max) trip-count + (VF * UF) does not overflow in the type of // the vector loop induction variable. if (unsigned TC = Cost->PSE.getSmallConstantMaxTripCount()) { uint64_t MaxVF = VF.getKnownMinValue(); if (VF.isScalable()) { std::optional MaxVScale = getMaxVScale(*Cost->TheFunction, Cost->TTI); if (!MaxVScale) return false; MaxVF *= *MaxVScale; } return (MaxUIntTripCount - TC).ugt(MaxVF * MaxUF); } return false; } // Return whether we allow using masked interleave-groups (for dealing with // strided loads/stores that reside in predicated blocks, or for dealing // with gaps). static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) { // If an override option has been passed in for interleaved accesses, use it. if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0) return EnableMaskedInterleavedMemAccesses; return TTI.enableMaskedInterleavedAccessVectorization(); } void InnerLoopVectorizer::introduceCheckBlockInVPlan(BasicBlock *CheckIRBB) { // Note: The block with the minimum trip-count check is already connected // during earlier VPlan construction. VPBlockBase *ScalarPH = Plan.getScalarPreheader(); VPBlockBase *PreVectorPH = VectorPHVPBB->getSinglePredecessor(); assert(PreVectorPH->getNumSuccessors() == 2 && "Expected 2 successors"); assert(PreVectorPH->getSuccessors()[0] == ScalarPH && "Unexpected successor"); VPIRBasicBlock *CheckVPIRBB = Plan.createVPIRBasicBlock(CheckIRBB); VPBlockUtils::insertOnEdge(PreVectorPH, VectorPHVPBB, CheckVPIRBB); PreVectorPH = CheckVPIRBB; VPBlockUtils::connectBlocks(PreVectorPH, ScalarPH); PreVectorPH->swapSuccessors(); // We just connected a new block to the scalar preheader. Update all // VPPhis by adding an incoming value for it, replicating the last value. unsigned NumPredecessors = ScalarPH->getNumPredecessors(); for (VPRecipeBase &R : cast(ScalarPH)->phis()) { assert(isa(&R) && "Phi expected to be VPPhi"); assert(cast(&R)->getNumIncoming() == NumPredecessors - 1 && "must have incoming values for all operands"); R.addOperand(R.getOperand(NumPredecessors - 2)); } } Value *InnerLoopVectorizer::createIterationCountCheck(ElementCount VF, unsigned UF) const { // Generate code to check if the loop's trip count is less than VF * UF, or // equal to it in case a scalar epilogue is required; this implies that the // vector trip count is zero. This check also covers the case where adding one // to the backedge-taken count overflowed leading to an incorrect trip count // of zero. In this case we will also jump to the scalar loop. auto P = Cost->requiresScalarEpilogue(VF.isVector()) ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; // Reuse existing vector loop preheader for TC checks. // Note that new preheader block is generated for vector loop. BasicBlock *const TCCheckBlock = LoopVectorPreHeader; IRBuilder<> Builder(TCCheckBlock->getTerminator()); // If tail is to be folded, vector loop takes care of all iterations. Value *Count = getTripCount(); Type *CountTy = Count->getType(); Value *CheckMinIters = Builder.getFalse(); auto CreateStep = [&]() -> Value * { // Create step with max(MinProTripCount, UF * VF). if (UF * VF.getKnownMinValue() >= MinProfitableTripCount.getKnownMinValue()) return createStepForVF(Builder, CountTy, VF, UF); Value *MinProfTC = createStepForVF(Builder, CountTy, MinProfitableTripCount, 1); if (!VF.isScalable()) return MinProfTC; return Builder.CreateBinaryIntrinsic( Intrinsic::umax, MinProfTC, createStepForVF(Builder, CountTy, VF, UF)); }; TailFoldingStyle Style = Cost->getTailFoldingStyle(); if (Style == TailFoldingStyle::None) { Value *Step = CreateStep(); ScalarEvolution &SE = *PSE.getSE(); // TODO: Emit unconditional branch to vector preheader instead of // conditional branch with known condition. const SCEV *TripCountSCEV = SE.applyLoopGuards(SE.getSCEV(Count), OrigLoop); // Check if the trip count is < the step. if (SE.isKnownPredicate(P, TripCountSCEV, SE.getSCEV(Step))) { // TODO: Ensure step is at most the trip count when determining max VF and // UF, w/o tail folding. CheckMinIters = Builder.getTrue(); } else if (!SE.isKnownPredicate(CmpInst::getInversePredicate(P), TripCountSCEV, SE.getSCEV(Step))) { // Generate the minimum iteration check only if we cannot prove the // check is known to be true, or known to be false. CheckMinIters = Builder.CreateICmp(P, Count, Step, "min.iters.check"); } // else step known to be < trip count, use CheckMinIters preset to false. } else if (VF.isScalable() && !TTI->isVScaleKnownToBeAPowerOfTwo() && !isIndvarOverflowCheckKnownFalse(Cost, VF, UF) && Style != TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck) { // vscale is not necessarily a power-of-2, which means we cannot guarantee // an overflow to zero when updating induction variables and so an // additional overflow check is required before entering the vector loop. // Get the maximum unsigned value for the type. Value *MaxUIntTripCount = ConstantInt::get(CountTy, cast(CountTy)->getMask()); Value *LHS = Builder.CreateSub(MaxUIntTripCount, Count); // Don't execute the vector loop if (UMax - n) < (VF * UF). CheckMinIters = Builder.CreateICmp(ICmpInst::ICMP_ULT, LHS, CreateStep()); } return CheckMinIters; } void InnerLoopVectorizer::emitIterationCountCheck(BasicBlock *Bypass) { BasicBlock *const TCCheckBlock = LoopVectorPreHeader; Value *CheckMinIters = createIterationCountCheck(VF, UF); // Create new preheader for vector loop. LoopVectorPreHeader = SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), static_cast(nullptr), LI, nullptr, "vector.ph"); BranchInst &BI = *BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters); if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator())) setBranchWeights(BI, MinItersBypassWeights, /*IsExpected=*/false); ReplaceInstWithInst(TCCheckBlock->getTerminator(), &BI); assert(cast(Plan.getEntry())->getIRBasicBlock() == TCCheckBlock && "Plan's entry must be TCCCheckBlock"); } /// Replace \p VPBB with a VPIRBasicBlock wrapping \p IRBB. All recipes from \p /// VPBB are moved to the end of the newly created VPIRBasicBlock. VPBB must /// have a single predecessor, which is rewired to the new VPIRBasicBlock. All /// successors of VPBB, if any, are rewired to the new VPIRBasicBlock. static VPIRBasicBlock *replaceVPBBWithIRVPBB(VPBasicBlock *VPBB, BasicBlock *IRBB) { VPIRBasicBlock *IRVPBB = VPBB->getPlan()->createVPIRBasicBlock(IRBB); auto IP = IRVPBB->begin(); for (auto &R : make_early_inc_range(VPBB->phis())) R.moveBefore(*IRVPBB, IP); for (auto &R : make_early_inc_range(make_range(VPBB->getFirstNonPhi(), VPBB->end()))) R.moveBefore(*IRVPBB, IRVPBB->end()); VPBlockUtils::reassociateBlocks(VPBB, IRVPBB); // VPBB is now dead and will be cleaned up when the plan gets destroyed. return IRVPBB; } void InnerLoopVectorizer::createVectorLoopSkeleton(StringRef Prefix) { LoopVectorPreHeader = OrigLoop->getLoopPreheader(); assert(LoopVectorPreHeader && "Invalid loop structure"); assert((OrigLoop->getUniqueLatchExitBlock() || Cost->requiresScalarEpilogue(VF.isVector())) && "loops not exiting via the latch without required epilogue?"); LoopScalarPreHeader = SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->getTerminator(), DT, LI, nullptr, Twine(Prefix) + "scalar.ph"); // NOTE: The Plan's scalar preheader VPBB isn't replaced with a VPIRBasicBlock // wrapping LoopScalarPreHeader here at the moment, because the Plan's scalar // preheader may be unreachable at this point. Instead it is replaced in // createVectorizedLoopSkeleton. } /// Return the expanded step for \p ID using \p ExpandedSCEVs to look up SCEV /// expansion results. static Value *getExpandedStep(const InductionDescriptor &ID, const SCEV2ValueTy &ExpandedSCEVs) { const SCEV *Step = ID.getStep(); if (auto *C = dyn_cast(Step)) return C->getValue(); if (auto *U = dyn_cast(Step)) return U->getValue(); Value *V = ExpandedSCEVs.lookup(Step); assert(V && "SCEV must be expanded at this point"); return V; } /// Knowing that loop \p L executes a single vector iteration, add instructions /// that will get simplified and thus should not have any cost to \p /// InstsToIgnore. static void addFullyUnrolledInstructionsToIgnore( Loop *L, const LoopVectorizationLegality::InductionList &IL, SmallPtrSetImpl &InstsToIgnore) { auto *Cmp = L->getLatchCmpInst(); if (Cmp) InstsToIgnore.insert(Cmp); for (const auto &KV : IL) { // Extract the key by hand so that it can be used in the lambda below. Note // that captured structured bindings are a C++20 extension. const PHINode *IV = KV.first; // Get next iteration value of the induction variable. Instruction *IVInst = cast(IV->getIncomingValueForBlock(L->getLoopLatch())); if (all_of(IVInst->users(), [&](const User *U) { return U == IV || U == Cmp; })) InstsToIgnore.insert(IVInst); } } BasicBlock *InnerLoopVectorizer::createVectorizedLoopSkeleton() { /* In this function we generate a new loop. The new loop will contain the vectorized instructions while the old loop will continue to run the scalar remainder. [ ] <-- old preheader - loop iteration number check and SCEVs in Plan's / | preheader are expanded here. Eventually all required SCEV / | expansion should happen here. / v | [ ] <-- vector loop bypass (may consist of multiple blocks). | / | | / v || [ ] <-- vector pre header. |/ | | v | [ ] \ | [ ]_| <-- vector loop (created during VPlan execution). | | | v \ -[ ] <--- middle-block (wrapped in VPIRBasicBlock with the branch to | | successors created during VPlan execution) \/ | /\ v | ->[ ] <--- new preheader (wrapped in VPIRBasicBlock). | | (opt) v <-- edge from middle to exit iff epilogue is not required. | [ ] \ | [ ]_| <-- old scalar loop to handle remainder (scalar epilogue, header | | wrapped in VPIRBasicBlock). \ | \ v >[ ] <-- exit block(s). (wrapped in VPIRBasicBlock) ... */ // Create an empty vector loop, and prepare basic blocks for the runtime // checks. createVectorLoopSkeleton(""); // Now, compare the new count to zero. If it is zero skip the vector loop and // jump to the scalar loop. This check also covers the case where the // backedge-taken count is uint##_max: adding one to it will overflow leading // to an incorrect trip count of zero. In this (rare) case we will also jump // to the scalar loop. emitIterationCountCheck(LoopScalarPreHeader); replaceVPBBWithIRVPBB(Plan.getScalarPreheader(), LoopScalarPreHeader); replaceVPBBWithIRVPBB(VectorPHVPBB, LoopVectorPreHeader); return LoopVectorPreHeader; } namespace { struct CSEDenseMapInfo { static bool canHandle(const Instruction *I) { return isa(I) || isa(I) || isa(I) || isa(I); } static inline Instruction *getEmptyKey() { return DenseMapInfo::getEmptyKey(); } static inline Instruction *getTombstoneKey() { return DenseMapInfo::getTombstoneKey(); } static unsigned getHashValue(const Instruction *I) { assert(canHandle(I) && "Unknown instruction!"); return hash_combine(I->getOpcode(), hash_combine_range(I->operand_values())); } static bool isEqual(const Instruction *LHS, const Instruction *RHS) { if (LHS == getEmptyKey() || RHS == getEmptyKey() || LHS == getTombstoneKey() || RHS == getTombstoneKey()) return LHS == RHS; return LHS->isIdenticalTo(RHS); } }; } // end anonymous namespace ///Perform cse of induction variable instructions. static void cse(BasicBlock *BB) { // Perform simple cse. SmallDenseMap CSEMap; for (Instruction &In : llvm::make_early_inc_range(*BB)) { if (!CSEDenseMapInfo::canHandle(&In)) continue; // Check if we can replace this instruction with any of the // visited instructions. if (Instruction *V = CSEMap.lookup(&In)) { In.replaceAllUsesWith(V); In.eraseFromParent(); continue; } CSEMap[&In] = &In; } } /// This function attempts to return a value that represents the ElementCount /// at runtime. For fixed-width VFs we know this precisely at compile /// time, but for scalable VFs we calculate it based on an estimate of the /// vscale value. static unsigned estimateElementCount(ElementCount VF, std::optional VScale) { unsigned EstimatedVF = VF.getKnownMinValue(); if (VF.isScalable()) if (VScale) EstimatedVF *= *VScale; assert(EstimatedVF >= 1 && "Estimated VF shouldn't be less than 1"); return EstimatedVF; } InstructionCost LoopVectorizationCostModel::getVectorCallCost(CallInst *CI, ElementCount VF) const { // We only need to calculate a cost if the VF is scalar; for actual vectors // we should already have a pre-calculated cost at each VF. if (!VF.isScalar()) return getCallWideningDecision(CI, VF).Cost; Type *RetTy = CI->getType(); if (RecurrenceDescriptor::isFMulAddIntrinsic(CI)) if (auto RedCost = getReductionPatternCost(CI, VF, RetTy)) return *RedCost; SmallVector Tys; for (auto &ArgOp : CI->args()) Tys.push_back(ArgOp->getType()); InstructionCost ScalarCallCost = TTI.getCallInstrCost(CI->getCalledFunction(), RetTy, Tys, CostKind); // If this is an intrinsic we may have a lower cost for it. if (getVectorIntrinsicIDForCall(CI, TLI)) { InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF); return std::min(ScalarCallCost, IntrinsicCost); } return ScalarCallCost; } static Type *maybeVectorizeType(Type *Ty, ElementCount VF) { if (VF.isScalar() || !canVectorizeTy(Ty)) return Ty; return toVectorizedTy(Ty, VF); } InstructionCost LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const { Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); assert(ID && "Expected intrinsic call!"); Type *RetTy = maybeVectorizeType(CI->getType(), VF); FastMathFlags FMF; if (auto *FPMO = dyn_cast(CI)) FMF = FPMO->getFastMathFlags(); SmallVector Arguments(CI->args()); FunctionType *FTy = CI->getCalledFunction()->getFunctionType(); SmallVector ParamTys; std::transform(FTy->param_begin(), FTy->param_end(), std::back_inserter(ParamTys), [&](Type *Ty) { return maybeVectorizeType(Ty, VF); }); IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF, dyn_cast(CI), InstructionCost::getInvalid(), TLI); return TTI.getIntrinsicInstrCost(CostAttrs, CostKind); } void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) { // Fix widened non-induction PHIs by setting up the PHI operands. fixNonInductionPHIs(State); // After vectorization, the exit blocks of the original loop will have // additional predecessors. Invalidate SCEVs for the exit phis in case SE // looked through single-entry phis. SmallVector ExitBlocks; OrigLoop->getExitBlocks(ExitBlocks); for (BasicBlock *Exit : ExitBlocks) for (PHINode &PN : Exit->phis()) PSE.getSE()->forgetLcssaPhiWithNewPredecessor(OrigLoop, &PN); // Forget the original basic block. PSE.getSE()->forgetLoop(OrigLoop); PSE.getSE()->forgetBlockAndLoopDispositions(); // Don't apply optimizations below when no (vector) loop remains, as they all // require one at the moment. VPBasicBlock *HeaderVPBB = vputils::getFirstLoopHeader(*State.Plan, State.VPDT); if (!HeaderVPBB) return; BasicBlock *HeaderBB = State.CFG.VPBB2IRBB[HeaderVPBB]; // Remove redundant induction instructions. cse(HeaderBB); // Set/update profile weights for the vector and remainder loops as original // loop iterations are now distributed among them. Note that original loop // becomes the scalar remainder loop after vectorization. // // For cases like foldTailByMasking() and requiresScalarEpiloque() we may // end up getting slightly roughened result but that should be OK since // profile is not inherently precise anyway. Note also possible bypass of // vector code caused by legality checks is ignored, assigning all the weight // to the vector loop, optimistically. // // For scalable vectorization we can't know at compile time how many // iterations of the loop are handled in one vector iteration, so instead // use the value of vscale used for tuning. Loop *VectorLoop = LI->getLoopFor(HeaderBB); unsigned EstimatedVFxUF = estimateElementCount(VF * UF, Cost->getVScaleForTuning()); setProfileInfoAfterUnrolling(OrigLoop, VectorLoop, OrigLoop, EstimatedVFxUF); } void InnerLoopVectorizer::fixNonInductionPHIs(VPTransformState &State) { auto Iter = vp_depth_first_shallow(Plan.getEntry()); for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly(Iter)) { for (VPRecipeBase &P : VPBB->phis()) { VPWidenPHIRecipe *VPPhi = dyn_cast(&P); if (!VPPhi) continue; PHINode *NewPhi = cast(State.get(VPPhi)); // Make sure the builder has a valid insert point. Builder.SetInsertPoint(NewPhi); for (unsigned Idx = 0; Idx < VPPhi->getNumIncoming(); ++Idx) { VPValue *Inc = VPPhi->getIncomingValue(Idx); const VPBasicBlock *VPBB = VPPhi->getIncomingBlock(Idx); NewPhi->addIncoming(State.get(Inc), State.CFG.VPBB2IRBB[VPBB]); } } } } void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) { // We should not collect Scalars more than once per VF. Right now, this // function is called from collectUniformsAndScalars(), which already does // this check. Collecting Scalars for VF=1 does not make any sense. assert(VF.isVector() && !Scalars.contains(VF) && "This function should not be visited twice for the same VF"); // This avoids any chances of creating a REPLICATE recipe during planning // since that would result in generation of scalarized code during execution, // which is not supported for scalable vectors. if (VF.isScalable()) { Scalars[VF].insert_range(Uniforms[VF]); return; } SmallSetVector Worklist; // These sets are used to seed the analysis with pointers used by memory // accesses that will remain scalar. SmallSetVector ScalarPtrs; SmallPtrSet PossibleNonScalarPtrs; auto *Latch = TheLoop->getLoopLatch(); // A helper that returns true if the use of Ptr by MemAccess will be scalar. // The pointer operands of loads and stores will be scalar as long as the // memory access is not a gather or scatter operation. The value operand of a // store will remain scalar if the store is scalarized. auto IsScalarUse = [&](Instruction *MemAccess, Value *Ptr) { InstWidening WideningDecision = getWideningDecision(MemAccess, VF); assert(WideningDecision != CM_Unknown && "Widening decision should be ready at this moment"); if (auto *Store = dyn_cast(MemAccess)) if (Ptr == Store->getValueOperand()) return WideningDecision == CM_Scalarize; assert(Ptr == getLoadStorePointerOperand(MemAccess) && "Ptr is neither a value or pointer operand"); return WideningDecision != CM_GatherScatter; }; // A helper that returns true if the given value is a getelementptr // instruction contained in the loop. auto IsLoopVaryingGEP = [&](Value *V) { return isa(V) && !TheLoop->isLoopInvariant(V); }; // A helper that evaluates a memory access's use of a pointer. If the use will // be a scalar use and the pointer is only used by memory accesses, we place // the pointer in ScalarPtrs. Otherwise, the pointer is placed in // PossibleNonScalarPtrs. auto EvaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) { // We only care about bitcast and getelementptr instructions contained in // the loop. if (!IsLoopVaryingGEP(Ptr)) return; // If the pointer has already been identified as scalar (e.g., if it was // also identified as uniform), there's nothing to do. auto *I = cast(Ptr); if (Worklist.count(I)) return; // If the use of the pointer will be a scalar use, and all users of the // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise, // place the pointer in PossibleNonScalarPtrs. if (IsScalarUse(MemAccess, Ptr) && all_of(I->users(), IsaPred)) ScalarPtrs.insert(I); else PossibleNonScalarPtrs.insert(I); }; // We seed the scalars analysis with three classes of instructions: (1) // instructions marked uniform-after-vectorization and (2) bitcast, // getelementptr and (pointer) phi instructions used by memory accesses // requiring a scalar use. // // (1) Add to the worklist all instructions that have been identified as // uniform-after-vectorization. Worklist.insert_range(Uniforms[VF]); // (2) Add to the worklist all bitcast and getelementptr instructions used by // memory accesses requiring a scalar use. The pointer operands of loads and // stores will be scalar unless the operation is a gather or scatter. // The value operand of a store will remain scalar if the store is scalarized. for (auto *BB : TheLoop->blocks()) for (auto &I : *BB) { if (auto *Load = dyn_cast(&I)) { EvaluatePtrUse(Load, Load->getPointerOperand()); } else if (auto *Store = dyn_cast(&I)) { EvaluatePtrUse(Store, Store->getPointerOperand()); EvaluatePtrUse(Store, Store->getValueOperand()); } } for (auto *I : ScalarPtrs) if (!PossibleNonScalarPtrs.count(I)) { LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *I << "\n"); Worklist.insert(I); } // Insert the forced scalars. // FIXME: Currently VPWidenPHIRecipe() often creates a dead vector // induction variable when the PHI user is scalarized. auto ForcedScalar = ForcedScalars.find(VF); if (ForcedScalar != ForcedScalars.end()) for (auto *I : ForcedScalar->second) { LLVM_DEBUG(dbgs() << "LV: Found (forced) scalar instruction: " << *I << "\n"); Worklist.insert(I); } // Expand the worklist by looking through any bitcasts and getelementptr // instructions we've already identified as scalar. This is similar to the // expansion step in collectLoopUniforms(); however, here we're only // expanding to include additional bitcasts and getelementptr instructions. unsigned Idx = 0; while (Idx != Worklist.size()) { Instruction *Dst = Worklist[Idx++]; if (!IsLoopVaryingGEP(Dst->getOperand(0))) continue; auto *Src = cast(Dst->getOperand(0)); if (llvm::all_of(Src->users(), [&](User *U) -> bool { auto *J = cast(U); return !TheLoop->contains(J) || Worklist.count(J) || ((isa(J) || isa(J)) && IsScalarUse(J, Src)); })) { Worklist.insert(Src); LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Src << "\n"); } } // An induction variable will remain scalar if all users of the induction // variable and induction variable update remain scalar. for (const auto &Induction : Legal->getInductionVars()) { auto *Ind = Induction.first; auto *IndUpdate = cast(Ind->getIncomingValueForBlock(Latch)); // If tail-folding is applied, the primary induction variable will be used // to feed a vector compare. if (Ind == Legal->getPrimaryInduction() && foldTailByMasking()) continue; // Returns true if \p Indvar is a pointer induction that is used directly by // load/store instruction \p I. auto IsDirectLoadStoreFromPtrIndvar = [&](Instruction *Indvar, Instruction *I) { return Induction.second.getKind() == InductionDescriptor::IK_PtrInduction && (isa(I) || isa(I)) && Indvar == getLoadStorePointerOperand(I) && IsScalarUse(I, Indvar); }; // Determine if all users of the induction variable are scalar after // vectorization. bool ScalarInd = all_of(Ind->users(), [&](User *U) -> bool { auto *I = cast(U); return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || IsDirectLoadStoreFromPtrIndvar(Ind, I); }); if (!ScalarInd) continue; // If the induction variable update is a fixed-order recurrence, neither the // induction variable or its update should be marked scalar after // vectorization. auto *IndUpdatePhi = dyn_cast(IndUpdate); if (IndUpdatePhi && Legal->isFixedOrderRecurrence(IndUpdatePhi)) continue; // Determine if all users of the induction variable update instruction are // scalar after vectorization. bool ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool { auto *I = cast(U); return I == Ind || !TheLoop->contains(I) || Worklist.count(I) || IsDirectLoadStoreFromPtrIndvar(IndUpdate, I); }); if (!ScalarIndUpdate) continue; // The induction variable and its update instruction will remain scalar. Worklist.insert(Ind); Worklist.insert(IndUpdate); LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n"); LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdate << "\n"); } Scalars[VF].insert_range(Worklist); } bool LoopVectorizationCostModel::isScalarWithPredication( Instruction *I, ElementCount VF) const { if (!isPredicatedInst(I)) return false; // Do we have a non-scalar lowering for this predicated // instruction? No - it is scalar with predication. switch(I->getOpcode()) { default: return true; case Instruction::Call: if (VF.isScalar()) return true; return getCallWideningDecision(cast(I), VF).Kind == CM_Scalarize; case Instruction::Load: case Instruction::Store: { auto *Ptr = getLoadStorePointerOperand(I); auto *Ty = getLoadStoreType(I); unsigned AS = getLoadStoreAddressSpace(I); Type *VTy = Ty; if (VF.isVector()) VTy = VectorType::get(Ty, VF); const Align Alignment = getLoadStoreAlignment(I); return isa(I) ? !(isLegalMaskedLoad(Ty, Ptr, Alignment, AS) || TTI.isLegalMaskedGather(VTy, Alignment)) : !(isLegalMaskedStore(Ty, Ptr, Alignment, AS) || TTI.isLegalMaskedScatter(VTy, Alignment)); } case Instruction::UDiv: case Instruction::SDiv: case Instruction::SRem: case Instruction::URem: { // We have the option to use the safe-divisor idiom to avoid predication. // The cost based decision here will always select safe-divisor for // scalable vectors as scalarization isn't legal. const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost); } } } // TODO: Fold into LoopVectorizationLegality::isMaskRequired. bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const { // TODO: We can use the loop-preheader as context point here and get // context sensitive reasoning for isSafeToSpeculativelyExecute. if (isSafeToSpeculativelyExecute(I) || (isa(I) && !Legal->isMaskRequired(I)) || isa(I)) return false; // If the instruction was executed conditionally in the original scalar loop, // predication is needed with a mask whose lanes are all possibly inactive. if (Legal->blockNeedsPredication(I->getParent())) return true; // If we're not folding the tail by masking, predication is unnecessary. if (!foldTailByMasking()) return false; // All that remain are instructions with side-effects originally executed in // the loop unconditionally, but now execute under a tail-fold mask (only) // having at least one active lane (the first). If the side-effects of the // instruction are invariant, executing it w/o (the tail-folding) mask is safe // - it will cause the same side-effects as when masked. switch(I->getOpcode()) { default: llvm_unreachable( "instruction should have been considered by earlier checks"); case Instruction::Call: // Side-effects of a Call are assumed to be non-invariant, needing a // (fold-tail) mask. assert(Legal->isMaskRequired(I) && "should have returned earlier for calls not needing a mask"); return true; case Instruction::Load: // If the address is loop invariant no predication is needed. return !Legal->isInvariant(getLoadStorePointerOperand(I)); case Instruction::Store: { // For stores, we need to prove both speculation safety (which follows from // the same argument as loads), but also must prove the value being stored // is correct. The easiest form of the later is to require that all values // stored are the same. return !(Legal->isInvariant(getLoadStorePointerOperand(I)) && TheLoop->isLoopInvariant(cast(I)->getValueOperand())); } case Instruction::UDiv: case Instruction::SDiv: case Instruction::SRem: case Instruction::URem: // If the divisor is loop-invariant no predication is needed. return !Legal->isInvariant(I->getOperand(1)); } } std::pair LoopVectorizationCostModel::getDivRemSpeculationCost(Instruction *I, ElementCount VF) const { assert(I->getOpcode() == Instruction::UDiv || I->getOpcode() == Instruction::SDiv || I->getOpcode() == Instruction::SRem || I->getOpcode() == Instruction::URem); assert(!isSafeToSpeculativelyExecute(I)); // Scalarization isn't legal for scalable vector types InstructionCost ScalarizationCost = InstructionCost::getInvalid(); if (!VF.isScalable()) { // Get the scalarization cost and scale this amount by the probability of // executing the predicated block. If the instruction is not predicated, // we fall through to the next case. ScalarizationCost = 0; // These instructions have a non-void type, so account for the phi nodes // that we will create. This cost is likely to be zero. The phi node // cost, if any, should be scaled by the block probability because it // models a copy at the end of each predicated block. ScalarizationCost += VF.getFixedValue() * TTI.getCFInstrCost(Instruction::PHI, CostKind); // The cost of the non-predicated instruction. ScalarizationCost += VF.getFixedValue() * TTI.getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind); // The cost of insertelement and extractelement instructions needed for // scalarization. ScalarizationCost += getScalarizationOverhead(I, VF); // Scale the cost by the probability of executing the predicated blocks. // This assumes the predicated block for each vector lane is equally // likely. ScalarizationCost = ScalarizationCost / getPredBlockCostDivisor(CostKind); } InstructionCost SafeDivisorCost = 0; auto *VecTy = toVectorTy(I->getType(), VF); // The cost of the select guard to ensure all lanes are well defined // after we speculate above any internal control flow. SafeDivisorCost += TTI.getCmpSelInstrCost(Instruction::Select, VecTy, toVectorTy(Type::getInt1Ty(I->getContext()), VF), CmpInst::BAD_ICMP_PREDICATE, CostKind); // Certain instructions can be cheaper to vectorize if they have a constant // second vector operand. One example of this are shifts on x86. Value *Op2 = I->getOperand(1); auto Op2Info = TTI.getOperandInfo(Op2); if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && Legal->isInvariant(Op2)) Op2Info.Kind = TargetTransformInfo::OK_UniformValue; SmallVector Operands(I->operand_values()); SafeDivisorCost += TTI.getArithmeticInstrCost( I->getOpcode(), VecTy, CostKind, {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, Op2Info, Operands, I); return {ScalarizationCost, SafeDivisorCost}; } bool LoopVectorizationCostModel::interleavedAccessCanBeWidened( Instruction *I, ElementCount VF) const { assert(isAccessInterleaved(I) && "Expecting interleaved access."); assert(getWideningDecision(I, VF) == CM_Unknown && "Decision should not be set yet."); auto *Group = getInterleavedAccessGroup(I); assert(Group && "Must have a group."); unsigned InterleaveFactor = Group->getFactor(); // If the instruction's allocated size doesn't equal its type size, it // requires padding and will be scalarized. auto &DL = I->getDataLayout(); auto *ScalarTy = getLoadStoreType(I); if (hasIrregularType(ScalarTy, DL)) return false; // For scalable vectors, the interleave factors must be <= 8 since we require // the (de)interleaveN intrinsics instead of shufflevectors. if (VF.isScalable() && InterleaveFactor > 8) return false; // If the group involves a non-integral pointer, we may not be able to // losslessly cast all values to a common type. bool ScalarNI = DL.isNonIntegralPointerType(ScalarTy); for (unsigned Idx = 0; Idx < InterleaveFactor; Idx++) { Instruction *Member = Group->getMember(Idx); if (!Member) continue; auto *MemberTy = getLoadStoreType(Member); bool MemberNI = DL.isNonIntegralPointerType(MemberTy); // Don't coerce non-integral pointers to integers or vice versa. if (MemberNI != ScalarNI) // TODO: Consider adding special nullptr value case here return false; if (MemberNI && ScalarNI && ScalarTy->getPointerAddressSpace() != MemberTy->getPointerAddressSpace()) return false; } // Check if masking is required. // A Group may need masking for one of two reasons: it resides in a block that // needs predication, or it was decided to use masking to deal with gaps // (either a gap at the end of a load-access that may result in a speculative // load, or any gaps in a store-access). bool PredicatedAccessRequiresMasking = blockNeedsPredicationForAnyReason(I->getParent()) && Legal->isMaskRequired(I); bool LoadAccessWithGapsRequiresEpilogMasking = isa(I) && Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed(); bool StoreAccessWithGapsRequiresMasking = isa(I) && !Group->isFull(); if (!PredicatedAccessRequiresMasking && !LoadAccessWithGapsRequiresEpilogMasking && !StoreAccessWithGapsRequiresMasking) return true; // If masked interleaving is required, we expect that the user/target had // enabled it, because otherwise it either wouldn't have been created or // it should have been invalidated by the CostModel. assert(useMaskedInterleavedAccesses(TTI) && "Masked interleave-groups for predicated accesses are not enabled."); if (Group->isReverse()) return false; auto *Ty = getLoadStoreType(I); const Align Alignment = getLoadStoreAlignment(I); unsigned AS = getLoadStoreAddressSpace(I); return isa(I) ? TTI.isLegalMaskedLoad(Ty, Alignment, AS) : TTI.isLegalMaskedStore(Ty, Alignment, AS); } bool LoopVectorizationCostModel::memoryInstructionCanBeWidened( Instruction *I, ElementCount VF) { // Get and ensure we have a valid memory instruction. assert((isa(I)) && "Invalid memory instruction"); auto *Ptr = getLoadStorePointerOperand(I); auto *ScalarTy = getLoadStoreType(I); // In order to be widened, the pointer should be consecutive, first of all. if (!Legal->isConsecutivePtr(ScalarTy, Ptr)) return false; // If the instruction is a store located in a predicated block, it will be // scalarized. if (isScalarWithPredication(I, VF)) return false; // If the instruction's allocated size doesn't equal it's type size, it // requires padding and will be scalarized. auto &DL = I->getDataLayout(); if (hasIrregularType(ScalarTy, DL)) return false; return true; } void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) { // We should not collect Uniforms more than once per VF. Right now, // this function is called from collectUniformsAndScalars(), which // already does this check. Collecting Uniforms for VF=1 does not make any // sense. assert(VF.isVector() && !Uniforms.contains(VF) && "This function should not be visited twice for the same VF"); // Visit the list of Uniforms. If we find no uniform value, we won't // analyze again. Uniforms.count(VF) will return 1. Uniforms[VF].clear(); // Now we know that the loop is vectorizable! // Collect instructions inside the loop that will remain uniform after // vectorization. // Global values, params and instructions outside of current loop are out of // scope. auto IsOutOfScope = [&](Value *V) -> bool { Instruction *I = dyn_cast(V); return (!I || !TheLoop->contains(I)); }; // Worklist containing uniform instructions demanding lane 0. SetVector Worklist; // Add uniform instructions demanding lane 0 to the worklist. Instructions // that require predication must not be considered uniform after // vectorization, because that would create an erroneous replicating region // where only a single instance out of VF should be formed. auto AddToWorklistIfAllowed = [&](Instruction *I) -> void { if (IsOutOfScope(I)) { LLVM_DEBUG(dbgs() << "LV: Found not uniform due to scope: " << *I << "\n"); return; } if (isPredicatedInst(I)) { LLVM_DEBUG( dbgs() << "LV: Found not uniform due to requiring predication: " << *I << "\n"); return; } LLVM_DEBUG(dbgs() << "LV: Found uniform instruction: " << *I << "\n"); Worklist.insert(I); }; // Start with the conditional branches exiting the loop. If the branch // condition is an instruction contained in the loop that is only used by the // branch, it is uniform. Note conditions from uncountable early exits are not // uniform. SmallVector Exiting; TheLoop->getExitingBlocks(Exiting); for (BasicBlock *E : Exiting) { if (Legal->hasUncountableEarlyExit() && TheLoop->getLoopLatch() != E) continue; auto *Cmp = dyn_cast(E->getTerminator()->getOperand(0)); if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse()) AddToWorklistIfAllowed(Cmp); } auto PrevVF = VF.divideCoefficientBy(2); // Return true if all lanes perform the same memory operation, and we can // thus choose to execute only one. auto IsUniformMemOpUse = [&](Instruction *I) { // If the value was already known to not be uniform for the previous // (smaller VF), it cannot be uniform for the larger VF. if (PrevVF.isVector()) { auto Iter = Uniforms.find(PrevVF); if (Iter != Uniforms.end() && !Iter->second.contains(I)) return false; } if (!Legal->isUniformMemOp(*I, VF)) return false; if (isa(I)) // Loading the same address always produces the same result - at least // assuming aliasing and ordering which have already been checked. return true; // Storing the same value on every iteration. return TheLoop->isLoopInvariant(cast(I)->getValueOperand()); }; auto IsUniformDecision = [&](Instruction *I, ElementCount VF) { InstWidening WideningDecision = getWideningDecision(I, VF); assert(WideningDecision != CM_Unknown && "Widening decision should be ready at this moment"); if (IsUniformMemOpUse(I)) return true; return (WideningDecision == CM_Widen || WideningDecision == CM_Widen_Reverse || WideningDecision == CM_Interleave); }; // Returns true if Ptr is the pointer operand of a memory access instruction // I, I is known to not require scalarization, and the pointer is not also // stored. auto IsVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool { if (isa(I) && I->getOperand(0) == Ptr) return false; return getLoadStorePointerOperand(I) == Ptr && (IsUniformDecision(I, VF) || Legal->isInvariant(Ptr)); }; // Holds a list of values which are known to have at least one uniform use. // Note that there may be other uses which aren't uniform. A "uniform use" // here is something which only demands lane 0 of the unrolled iterations; // it does not imply that all lanes produce the same value (e.g. this is not // the usual meaning of uniform) SetVector HasUniformUse; // Scan the loop for instructions which are either a) known to have only // lane 0 demanded or b) are uses which demand only lane 0 of their operand. for (auto *BB : TheLoop->blocks()) for (auto &I : *BB) { if (IntrinsicInst *II = dyn_cast(&I)) { switch (II->getIntrinsicID()) { case Intrinsic::sideeffect: case Intrinsic::experimental_noalias_scope_decl: case Intrinsic::assume: case Intrinsic::lifetime_start: case Intrinsic::lifetime_end: if (TheLoop->hasLoopInvariantOperands(&I)) AddToWorklistIfAllowed(&I); break; default: break; } } if (auto *EVI = dyn_cast(&I)) { if (IsOutOfScope(EVI->getAggregateOperand())) { AddToWorklistIfAllowed(EVI); continue; } // Only ExtractValue instructions where the aggregate value comes from a // call are allowed to be non-uniform. assert(isa(EVI->getAggregateOperand()) && "Expected aggregate value to be call return value"); } // If there's no pointer operand, there's nothing to do. auto *Ptr = getLoadStorePointerOperand(&I); if (!Ptr) continue; if (IsUniformMemOpUse(&I)) AddToWorklistIfAllowed(&I); if (IsVectorizedMemAccessUse(&I, Ptr)) HasUniformUse.insert(Ptr); } // Add to the worklist any operands which have *only* uniform (e.g. lane 0 // demanding) users. Since loops are assumed to be in LCSSA form, this // disallows uses outside the loop as well. for (auto *V : HasUniformUse) { if (IsOutOfScope(V)) continue; auto *I = cast(V); bool UsersAreMemAccesses = all_of(I->users(), [&](User *U) -> bool { auto *UI = cast(U); return TheLoop->contains(UI) && IsVectorizedMemAccessUse(UI, V); }); if (UsersAreMemAccesses) AddToWorklistIfAllowed(I); } // Expand Worklist in topological order: whenever a new instruction // is added , its users should be already inside Worklist. It ensures // a uniform instruction will only be used by uniform instructions. unsigned Idx = 0; while (Idx != Worklist.size()) { Instruction *I = Worklist[Idx++]; for (auto *OV : I->operand_values()) { // isOutOfScope operands cannot be uniform instructions. if (IsOutOfScope(OV)) continue; // First order recurrence Phi's should typically be considered // non-uniform. auto *OP = dyn_cast(OV); if (OP && Legal->isFixedOrderRecurrence(OP)) continue; // If all the users of the operand are uniform, then add the // operand into the uniform worklist. auto *OI = cast(OV); if (llvm::all_of(OI->users(), [&](User *U) -> bool { auto *J = cast(U); return Worklist.count(J) || IsVectorizedMemAccessUse(J, OI); })) AddToWorklistIfAllowed(OI); } } // For an instruction to be added into Worklist above, all its users inside // the loop should also be in Worklist. However, this condition cannot be // true for phi nodes that form a cyclic dependence. We must process phi // nodes separately. An induction variable will remain uniform if all users // of the induction variable and induction variable update remain uniform. // The code below handles both pointer and non-pointer induction variables. BasicBlock *Latch = TheLoop->getLoopLatch(); for (const auto &Induction : Legal->getInductionVars()) { auto *Ind = Induction.first; auto *IndUpdate = cast(Ind->getIncomingValueForBlock(Latch)); // Determine if all users of the induction variable are uniform after // vectorization. bool UniformInd = all_of(Ind->users(), [&](User *U) -> bool { auto *I = cast(U); return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) || IsVectorizedMemAccessUse(I, Ind); }); if (!UniformInd) continue; // Determine if all users of the induction variable update instruction are // uniform after vectorization. bool UniformIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool { auto *I = cast(U); return I == Ind || Worklist.count(I) || IsVectorizedMemAccessUse(I, IndUpdate); }); if (!UniformIndUpdate) continue; // The induction variable and its update instruction will remain uniform. AddToWorklistIfAllowed(Ind); AddToWorklistIfAllowed(IndUpdate); } Uniforms[VF].insert_range(Worklist); } bool LoopVectorizationCostModel::runtimeChecksRequired() { LLVM_DEBUG(dbgs() << "LV: Performing code size checks.\n"); if (Legal->getRuntimePointerChecking()->Need) { reportVectorizationFailure("Runtime ptr check is required with -Os/-Oz", "runtime pointer checks needed. Enable vectorization of this " "loop with '#pragma clang loop vectorize(enable)' when " "compiling with -Os/-Oz", "CantVersionLoopWithOptForSize", ORE, TheLoop); return true; } if (!PSE.getPredicate().isAlwaysTrue()) { reportVectorizationFailure("Runtime SCEV check is required with -Os/-Oz", "runtime SCEV checks needed. Enable vectorization of this " "loop with '#pragma clang loop vectorize(enable)' when " "compiling with -Os/-Oz", "CantVersionLoopWithOptForSize", ORE, TheLoop); return true; } // FIXME: Avoid specializing for stride==1 instead of bailing out. if (!Legal->getLAI()->getSymbolicStrides().empty()) { reportVectorizationFailure("Runtime stride check for small trip count", "runtime stride == 1 checks needed. Enable vectorization of " "this loop without such check by compiling with -Os/-Oz", "CantVersionLoopWithOptForSize", ORE, TheLoop); return true; } return false; } bool LoopVectorizationCostModel::isScalableVectorizationAllowed() { if (IsScalableVectorizationAllowed) return *IsScalableVectorizationAllowed; IsScalableVectorizationAllowed = false; if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) return false; if (Hints->isScalableVectorizationDisabled()) { reportVectorizationInfo("Scalable vectorization is explicitly disabled", "ScalableVectorizationDisabled", ORE, TheLoop); return false; } LLVM_DEBUG(dbgs() << "LV: Scalable vectorization is available\n"); auto MaxScalableVF = ElementCount::getScalable( std::numeric_limits::max()); // Test that the loop-vectorizer can legalize all operations for this MaxVF. // FIXME: While for scalable vectors this is currently sufficient, this should // be replaced by a more detailed mechanism that filters out specific VFs, // instead of invalidating vectorization for a whole set of VFs based on the // MaxVF. // Disable scalable vectorization if the loop contains unsupported reductions. if (!canVectorizeReductions(MaxScalableVF)) { reportVectorizationInfo( "Scalable vectorization not supported for the reduction " "operations found in this loop.", "ScalableVFUnfeasible", ORE, TheLoop); return false; } // Disable scalable vectorization if the loop contains any instructions // with element types not supported for scalable vectors. if (any_of(ElementTypesInLoop, [&](Type *Ty) { return !Ty->isVoidTy() && !this->TTI.isElementTypeLegalForScalableVector(Ty); })) { reportVectorizationInfo("Scalable vectorization is not supported " "for all element types found in this loop.", "ScalableVFUnfeasible", ORE, TheLoop); return false; } if (!Legal->isSafeForAnyVectorWidth() && !getMaxVScale(*TheFunction, TTI)) { reportVectorizationInfo("The target does not provide maximum vscale value " "for safe distance analysis.", "ScalableVFUnfeasible", ORE, TheLoop); return false; } IsScalableVectorizationAllowed = true; return true; } ElementCount LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) { if (!isScalableVectorizationAllowed()) return ElementCount::getScalable(0); auto MaxScalableVF = ElementCount::getScalable( std::numeric_limits::max()); if (Legal->isSafeForAnyVectorWidth()) return MaxScalableVF; std::optional MaxVScale = getMaxVScale(*TheFunction, TTI); // Limit MaxScalableVF by the maximum safe dependence distance. MaxScalableVF = ElementCount::getScalable(MaxSafeElements / *MaxVScale); if (!MaxScalableVF) reportVectorizationInfo( "Max legal vector width too small, scalable vectorization " "unfeasible.", "ScalableVFUnfeasible", ORE, TheLoop); return MaxScalableVF; } FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF( unsigned MaxTripCount, ElementCount UserVF, bool FoldTailByMasking) { MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI); unsigned SmallestType, WidestType; std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes(); // Get the maximum safe dependence distance in bits computed by LAA. // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from // the memory accesses that is most restrictive (involved in the smallest // dependence distance). unsigned MaxSafeElementsPowerOf2 = bit_floor(Legal->getMaxSafeVectorWidthInBits() / WidestType); if (!Legal->isSafeForAnyStoreLoadForwardDistances()) { unsigned SLDist = Legal->getMaxStoreLoadForwardSafeDistanceInBits(); MaxSafeElementsPowerOf2 = std::min(MaxSafeElementsPowerOf2, SLDist / WidestType); } auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElementsPowerOf2); auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElementsPowerOf2); if (!Legal->isSafeForAnyVectorWidth()) this->MaxSafeElements = MaxSafeElementsPowerOf2; LLVM_DEBUG(dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF << ".\n"); LLVM_DEBUG(dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF << ".\n"); // First analyze the UserVF, fall back if the UserVF should be ignored. if (UserVF) { auto MaxSafeUserVF = UserVF.isScalable() ? MaxSafeScalableVF : MaxSafeFixedVF; if (ElementCount::isKnownLE(UserVF, MaxSafeUserVF)) { // If `VF=vscale x N` is safe, then so is `VF=N` if (UserVF.isScalable()) return FixedScalableVFPair( ElementCount::getFixed(UserVF.getKnownMinValue()), UserVF); return UserVF; } assert(ElementCount::isKnownGT(UserVF, MaxSafeUserVF)); // Only clamp if the UserVF is not scalable. If the UserVF is scalable, it // is better to ignore the hint and let the compiler choose a suitable VF. if (!UserVF.isScalable()) { LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF << " is unsafe, clamping to max safe VF=" << MaxSafeFixedVF << ".\n"); ORE->emit([&]() { return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor", TheLoop->getStartLoc(), TheLoop->getHeader()) << "User-specified vectorization factor " << ore::NV("UserVectorizationFactor", UserVF) << " is unsafe, clamping to maximum safe vectorization factor " << ore::NV("VectorizationFactor", MaxSafeFixedVF); }); return MaxSafeFixedVF; } if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) { LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF << " is ignored because scalable vectors are not " "available.\n"); ORE->emit([&]() { return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor", TheLoop->getStartLoc(), TheLoop->getHeader()) << "User-specified vectorization factor " << ore::NV("UserVectorizationFactor", UserVF) << " is ignored because the target does not support scalable " "vectors. The compiler will pick a more suitable value."; }); } else { LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF << " is unsafe. Ignoring scalable UserVF.\n"); ORE->emit([&]() { return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor", TheLoop->getStartLoc(), TheLoop->getHeader()) << "User-specified vectorization factor " << ore::NV("UserVectorizationFactor", UserVF) << " is unsafe. Ignoring the hint to let the compiler pick a " "more suitable value."; }); } } LLVM_DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType << " / " << WidestType << " bits.\n"); FixedScalableVFPair Result(ElementCount::getFixed(1), ElementCount::getScalable(0)); if (auto MaxVF = getMaximizedVFForTarget(MaxTripCount, SmallestType, WidestType, MaxSafeFixedVF, FoldTailByMasking)) Result.FixedVF = MaxVF; if (auto MaxVF = getMaximizedVFForTarget(MaxTripCount, SmallestType, WidestType, MaxSafeScalableVF, FoldTailByMasking)) if (MaxVF.isScalable()) { Result.ScalableVF = MaxVF; LLVM_DEBUG(dbgs() << "LV: Found feasible scalable VF = " << MaxVF << "\n"); } return Result; } FixedScalableVFPair LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) { if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) { // TODO: It may be useful to do since it's still likely to be dynamically // uniform if the target can skip. reportVectorizationFailure( "Not inserting runtime ptr check for divergent target", "runtime pointer checks needed. Not enabled for divergent target", "CantVersionLoopWithDivergentTarget", ORE, TheLoop); return FixedScalableVFPair::getNone(); } ScalarEvolution *SE = PSE.getSE(); ElementCount TC = getSmallConstantTripCount(SE, TheLoop); unsigned MaxTC = PSE.getSmallConstantMaxTripCount(); LLVM_DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n'); if (TC != ElementCount::getFixed(MaxTC)) LLVM_DEBUG(dbgs() << "LV: Found maximum trip count: " << MaxTC << '\n'); if (TC.isScalar()) { reportVectorizationFailure("Single iteration (non) loop", "loop trip count is one, irrelevant for vectorization", "SingleIterationLoop", ORE, TheLoop); return FixedScalableVFPair::getNone(); } // If BTC matches the widest induction type and is -1 then the trip count // computation will wrap to 0 and the vector trip count will be 0. Do not try // to vectorize. const SCEV *BTC = SE->getBackedgeTakenCount(TheLoop); if (!isa(BTC) && BTC->getType()->getScalarSizeInBits() >= Legal->getWidestInductionType()->getScalarSizeInBits() && SE->isKnownPredicate(CmpInst::ICMP_EQ, BTC, SE->getMinusOne(BTC->getType()))) { reportVectorizationFailure( "Trip count computation wrapped", "backedge-taken count is -1, loop trip count wrapped to 0", "TripCountWrapped", ORE, TheLoop); return FixedScalableVFPair::getNone(); } switch (ScalarEpilogueStatus) { case CM_ScalarEpilogueAllowed: return computeFeasibleMaxVF(MaxTC, UserVF, false); case CM_ScalarEpilogueNotAllowedUsePredicate: [[fallthrough]]; case CM_ScalarEpilogueNotNeededUsePredicate: LLVM_DEBUG( dbgs() << "LV: vector predicate hint/switch found.\n" << "LV: Not allowing scalar epilogue, creating predicated " << "vector loop.\n"); break; case CM_ScalarEpilogueNotAllowedLowTripLoop: // fallthrough as a special case of OptForSize case CM_ScalarEpilogueNotAllowedOptSize: if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedOptSize) LLVM_DEBUG( dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n"); else LLVM_DEBUG(dbgs() << "LV: Not allowing scalar epilogue due to low trip " << "count.\n"); // Bail if runtime checks are required, which are not good when optimising // for size. if (runtimeChecksRequired()) return FixedScalableVFPair::getNone(); break; } // Now try the tail folding // Invalidate interleave groups that require an epilogue if we can't mask // the interleave-group. if (!useMaskedInterleavedAccesses(TTI)) { assert(WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() && "No decisions should have been taken at this point"); // Note: There is no need to invalidate any cost modeling decisions here, as // none were taken so far. InterleaveInfo.invalidateGroupsRequiringScalarEpilogue(); } FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(MaxTC, UserVF, true); // Avoid tail folding if the trip count is known to be a multiple of any VF // we choose. std::optional MaxPowerOf2RuntimeVF = MaxFactors.FixedVF.getFixedValue(); if (MaxFactors.ScalableVF) { std::optional MaxVScale = getMaxVScale(*TheFunction, TTI); if (MaxVScale && TTI.isVScaleKnownToBeAPowerOfTwo()) { MaxPowerOf2RuntimeVF = std::max( *MaxPowerOf2RuntimeVF, *MaxVScale * MaxFactors.ScalableVF.getKnownMinValue()); } else MaxPowerOf2RuntimeVF = std::nullopt; // Stick with tail-folding for now. } auto NoScalarEpilogueNeeded = [this, &UserIC](unsigned MaxVF) { // Return false if the loop is neither a single-latch-exit loop nor an // early-exit loop as tail-folding is not supported in that case. if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch() && !Legal->hasUncountableEarlyExit()) return false; unsigned MaxVFtimesIC = UserIC ? MaxVF * UserIC : MaxVF; ScalarEvolution *SE = PSE.getSE(); // Calling getSymbolicMaxBackedgeTakenCount enables support for loops // with uncountable exits. For countable loops, the symbolic maximum must // remain identical to the known back-edge taken count. const SCEV *BackedgeTakenCount = PSE.getSymbolicMaxBackedgeTakenCount(); assert((Legal->hasUncountableEarlyExit() || BackedgeTakenCount == PSE.getBackedgeTakenCount()) && "Invalid loop count"); const SCEV *ExitCount = SE->getAddExpr( BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType())); const SCEV *Rem = SE->getURemExpr( SE->applyLoopGuards(ExitCount, TheLoop), SE->getConstant(BackedgeTakenCount->getType(), MaxVFtimesIC)); return Rem->isZero(); }; if (MaxPowerOf2RuntimeVF > 0u) { assert((UserVF.isNonZero() || isPowerOf2_32(*MaxPowerOf2RuntimeVF)) && "MaxFixedVF must be a power of 2"); if (NoScalarEpilogueNeeded(*MaxPowerOf2RuntimeVF)) { // Accept MaxFixedVF if we do not have a tail. LLVM_DEBUG(dbgs() << "LV: No tail will remain for any chosen VF.\n"); return MaxFactors; } } auto ExpectedTC = getSmallBestKnownTC(PSE, TheLoop); if (ExpectedTC && ExpectedTC->isFixed() && ExpectedTC->getFixedValue() <= TTI.getMinTripCountTailFoldingThreshold()) { if (MaxPowerOf2RuntimeVF > 0u) { // If we have a low-trip-count, and the fixed-width VF is known to divide // the trip count but the scalable factor does not, use the fixed-width // factor in preference to allow the generation of a non-predicated loop. if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedLowTripLoop && NoScalarEpilogueNeeded(MaxFactors.FixedVF.getFixedValue())) { LLVM_DEBUG(dbgs() << "LV: Picking a fixed-width so that no tail will " "remain for any chosen VF.\n"); MaxFactors.ScalableVF = ElementCount::getScalable(0); return MaxFactors; } } reportVectorizationFailure( "The trip count is below the minial threshold value.", "loop trip count is too low, avoiding vectorization", "LowTripCount", ORE, TheLoop); return FixedScalableVFPair::getNone(); } // If we don't know the precise trip count, or if the trip count that we // found modulo the vectorization factor is not zero, try to fold the tail // by masking. // FIXME: look for a smaller MaxVF that does divide TC rather than masking. bool ContainsScalableVF = MaxFactors.ScalableVF.isNonZero(); setTailFoldingStyles(ContainsScalableVF, UserIC); if (foldTailByMasking()) { if (getTailFoldingStyle() == TailFoldingStyle::DataWithEVL) { LLVM_DEBUG( dbgs() << "LV: tail is folded with EVL, forcing unroll factor to be 1. Will " "try to generate VP Intrinsics with scalable vector " "factors only.\n"); // Tail folded loop using VP intrinsics restricts the VF to be scalable // for now. // TODO: extend it for fixed vectors, if required. assert(ContainsScalableVF && "Expected scalable vector factor."); MaxFactors.FixedVF = ElementCount::getFixed(1); } return MaxFactors; } // If there was a tail-folding hint/switch, but we can't fold the tail by // masking, fallback to a vectorization with a scalar epilogue. if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) { LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking: vectorize with a " "scalar epilogue instead.\n"); ScalarEpilogueStatus = CM_ScalarEpilogueAllowed; return MaxFactors; } if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedUsePredicate) { LLVM_DEBUG(dbgs() << "LV: Can't fold tail by masking: don't vectorize\n"); return FixedScalableVFPair::getNone(); } if (TC.isZero()) { reportVectorizationFailure( "unable to calculate the loop count due to complex control flow", "UnknownLoopCountComplexCFG", ORE, TheLoop); return FixedScalableVFPair::getNone(); } reportVectorizationFailure( "Cannot optimize for size and vectorize at the same time.", "cannot optimize for size and vectorize at the same time. " "Enable vectorization of this loop with '#pragma clang loop " "vectorize(enable)' when compiling with -Os/-Oz", "NoTailLoopWithOptForSize", ORE, TheLoop); return FixedScalableVFPair::getNone(); } bool LoopVectorizationCostModel::shouldCalculateRegPressureForVF( ElementCount VF) { if (!useMaxBandwidth(VF.isScalable() ? TargetTransformInfo::RGK_ScalableVector : TargetTransformInfo::RGK_FixedWidthVector)) return false; // Only calculate register pressure for VFs enabled by MaxBandwidth. return ElementCount::isKnownGT( VF, VF.isScalable() ? MaxPermissibleVFWithoutMaxBW.ScalableVF : MaxPermissibleVFWithoutMaxBW.FixedVF); } bool LoopVectorizationCostModel::useMaxBandwidth( TargetTransformInfo::RegisterKind RegKind) { return MaximizeBandwidth || (MaximizeBandwidth.getNumOccurrences() == 0 && (TTI.shouldMaximizeVectorBandwidth(RegKind) || (UseWiderVFIfCallVariantsPresent && Legal->hasVectorCallVariants()))); } ElementCount LoopVectorizationCostModel::clampVFByMaxTripCount( ElementCount VF, unsigned MaxTripCount, bool FoldTailByMasking) const { unsigned EstimatedVF = VF.getKnownMinValue(); if (VF.isScalable() && TheFunction->hasFnAttribute(Attribute::VScaleRange)) { auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange); auto Min = Attr.getVScaleRangeMin(); EstimatedVF *= Min; } // When a scalar epilogue is required, at least one iteration of the scalar // loop has to execute. Adjust MaxTripCount accordingly to avoid picking a // max VF that results in a dead vector loop. if (MaxTripCount > 0 && requiresScalarEpilogue(true)) MaxTripCount -= 1; if (MaxTripCount && MaxTripCount <= EstimatedVF && (!FoldTailByMasking || isPowerOf2_32(MaxTripCount))) { // If upper bound loop trip count (TC) is known at compile time there is no // point in choosing VF greater than TC (as done in the loop below). Select // maximum power of two which doesn't exceed TC. If VF is // scalable, we only fall back on a fixed VF when the TC is less than or // equal to the known number of lanes. auto ClampedUpperTripCount = llvm::bit_floor(MaxTripCount); LLVM_DEBUG(dbgs() << "LV: Clamping the MaxVF to maximum power of two not " "exceeding the constant trip count: " << ClampedUpperTripCount << "\n"); return ElementCount::get(ClampedUpperTripCount, FoldTailByMasking ? VF.isScalable() : false); } return VF; } ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget( unsigned MaxTripCount, unsigned SmallestType, unsigned WidestType, ElementCount MaxSafeVF, bool FoldTailByMasking) { bool ComputeScalableMaxVF = MaxSafeVF.isScalable(); const TypeSize WidestRegister = TTI.getRegisterBitWidth( ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector : TargetTransformInfo::RGK_FixedWidthVector); // Convenience function to return the minimum of two ElementCounts. auto MinVF = [](const ElementCount &LHS, const ElementCount &RHS) { assert((LHS.isScalable() == RHS.isScalable()) && "Scalable flags must match"); return ElementCount::isKnownLT(LHS, RHS) ? LHS : RHS; }; // Ensure MaxVF is a power of 2; the dependence distance bound may not be. // Note that both WidestRegister and WidestType may not be a powers of 2. auto MaxVectorElementCount = ElementCount::get( llvm::bit_floor(WidestRegister.getKnownMinValue() / WidestType), ComputeScalableMaxVF); MaxVectorElementCount = MinVF(MaxVectorElementCount, MaxSafeVF); LLVM_DEBUG(dbgs() << "LV: The Widest register safe to use is: " << (MaxVectorElementCount * WidestType) << " bits.\n"); if (!MaxVectorElementCount) { LLVM_DEBUG(dbgs() << "LV: The target has no " << (ComputeScalableMaxVF ? "scalable" : "fixed") << " vector registers.\n"); return ElementCount::getFixed(1); } ElementCount MaxVF = clampVFByMaxTripCount(MaxVectorElementCount, MaxTripCount, FoldTailByMasking); // If the MaxVF was already clamped, there's no point in trying to pick a // larger one. if (MaxVF != MaxVectorElementCount) return MaxVF; TargetTransformInfo::RegisterKind RegKind = ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector : TargetTransformInfo::RGK_FixedWidthVector; if (MaxVF.isScalable()) MaxPermissibleVFWithoutMaxBW.ScalableVF = MaxVF; else MaxPermissibleVFWithoutMaxBW.FixedVF = MaxVF; if (useMaxBandwidth(RegKind)) { auto MaxVectorElementCountMaxBW = ElementCount::get( llvm::bit_floor(WidestRegister.getKnownMinValue() / SmallestType), ComputeScalableMaxVF); MaxVF = MinVF(MaxVectorElementCountMaxBW, MaxSafeVF); if (ElementCount MinVF = TTI.getMinimumVF(SmallestType, ComputeScalableMaxVF)) { if (ElementCount::isKnownLT(MaxVF, MinVF)) { LLVM_DEBUG(dbgs() << "LV: Overriding calculated MaxVF(" << MaxVF << ") with target's minimum: " << MinVF << '\n'); MaxVF = MinVF; } } MaxVF = clampVFByMaxTripCount(MaxVF, MaxTripCount, FoldTailByMasking); if (MaxVectorElementCount != MaxVF) { // Invalidate any widening decisions we might have made, in case the loop // requires prediction (decided later), but we have already made some // load/store widening decisions. invalidateCostModelingDecisions(); } } return MaxVF; } bool LoopVectorizationPlanner::isMoreProfitable(const VectorizationFactor &A, const VectorizationFactor &B, const unsigned MaxTripCount, bool HasTail) const { InstructionCost CostA = A.Cost; InstructionCost CostB = B.Cost; // Improve estimate for the vector width if it is scalable. unsigned EstimatedWidthA = A.Width.getKnownMinValue(); unsigned EstimatedWidthB = B.Width.getKnownMinValue(); if (std::optional VScale = CM.getVScaleForTuning()) { if (A.Width.isScalable()) EstimatedWidthA *= *VScale; if (B.Width.isScalable()) EstimatedWidthB *= *VScale; } // When optimizing for size choose whichever is smallest, which will be the // one with the smallest cost for the whole loop. On a tie pick the larger // vector width, on the assumption that throughput will be greater. if (CM.CostKind == TTI::TCK_CodeSize) return CostA < CostB || (CostA == CostB && EstimatedWidthA > EstimatedWidthB); // Assume vscale may be larger than 1 (or the value being tuned for), // so that scalable vectorization is slightly favorable over fixed-width // vectorization. bool PreferScalable = !TTI.preferFixedOverScalableIfEqualCost() && A.Width.isScalable() && !B.Width.isScalable(); auto CmpFn = [PreferScalable](const InstructionCost &LHS, const InstructionCost &RHS) { return PreferScalable ? LHS <= RHS : LHS < RHS; }; // To avoid the need for FP division: // (CostA / EstimatedWidthA) < (CostB / EstimatedWidthB) // <=> (CostA * EstimatedWidthB) < (CostB * EstimatedWidthA) if (!MaxTripCount) return CmpFn(CostA * EstimatedWidthB, CostB * EstimatedWidthA); auto GetCostForTC = [MaxTripCount, HasTail](unsigned VF, InstructionCost VectorCost, InstructionCost ScalarCost) { // If the trip count is a known (possibly small) constant, the trip count // will be rounded up to an integer number of iterations under // FoldTailByMasking. The total cost in that case will be // VecCost*ceil(TripCount/VF). When not folding the tail, the total // cost will be VecCost*floor(TC/VF) + ScalarCost*(TC%VF). There will be // some extra overheads, but for the purpose of comparing the costs of // different VFs we can use this to compare the total loop-body cost // expected after vectorization. if (HasTail) return VectorCost * (MaxTripCount / VF) + ScalarCost * (MaxTripCount % VF); return VectorCost * divideCeil(MaxTripCount, VF); }; auto RTCostA = GetCostForTC(EstimatedWidthA, CostA, A.ScalarCost); auto RTCostB = GetCostForTC(EstimatedWidthB, CostB, B.ScalarCost); return CmpFn(RTCostA, RTCostB); } bool LoopVectorizationPlanner::isMoreProfitable(const VectorizationFactor &A, const VectorizationFactor &B, bool HasTail) const { const unsigned MaxTripCount = PSE.getSmallConstantMaxTripCount(); return LoopVectorizationPlanner::isMoreProfitable(A, B, MaxTripCount, HasTail); } void LoopVectorizationPlanner::emitInvalidCostRemarks( OptimizationRemarkEmitter *ORE) { using RecipeVFPair = std::pair; SmallVector InvalidCosts; for (const auto &Plan : VPlans) { for (ElementCount VF : Plan->vectorFactors()) { // The VPlan-based cost model is designed for computing vector cost. // Querying VPlan-based cost model with a scarlar VF will cause some // errors because we expect the VF is vector for most of the widen // recipes. if (VF.isScalar()) continue; VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind); precomputeCosts(*Plan, VF, CostCtx); auto Iter = vp_depth_first_deep(Plan->getVectorLoopRegion()->getEntry()); for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly(Iter)) { for (auto &R : *VPBB) { if (!R.cost(VF, CostCtx).isValid()) InvalidCosts.emplace_back(&R, VF); } } } } if (InvalidCosts.empty()) return; // Emit a report of VFs with invalid costs in the loop. // Group the remarks per recipe, keeping the recipe order from InvalidCosts. DenseMap Numbering; unsigned I = 0; for (auto &Pair : InvalidCosts) if (Numbering.try_emplace(Pair.first, I).second) ++I; // Sort the list, first on recipe(number) then on VF. sort(InvalidCosts, [&Numbering](RecipeVFPair &A, RecipeVFPair &B) { unsigned NA = Numbering[A.first]; unsigned NB = Numbering[B.first]; if (NA != NB) return NA < NB; return ElementCount::isKnownLT(A.second, B.second); }); // For a list of ordered recipe-VF pairs: // [(load, VF1), (load, VF2), (store, VF1)] // group the recipes together to emit separate remarks for: // load (VF1, VF2) // store (VF1) auto Tail = ArrayRef(InvalidCosts); auto Subset = ArrayRef(); do { if (Subset.empty()) Subset = Tail.take_front(1); VPRecipeBase *R = Subset.front().first; unsigned Opcode = TypeSwitch(R) .Case( [](const auto *R) { return Instruction::PHI; }) .Case( [](const auto *R) { return Instruction::Select; }) .Case( [](const auto *R) { return Instruction::Store; }) .Case( [](const auto *R) { return Instruction::Load; }) .Case( [](const auto *R) { return Instruction::Call; }) .Case( [](const auto *R) { return R->getOpcode(); }) .Case([](const VPInterleaveRecipe *R) { return R->getStoredValues().empty() ? Instruction::Load : Instruction::Store; }); // If the next recipe is different, or if there are no other pairs, // emit a remark for the collated subset. e.g. // [(load, VF1), (load, VF2))] // to emit: // remark: invalid costs for 'load' at VF=(VF1, VF2) if (Subset == Tail || Tail[Subset.size()].first != R) { std::string OutString; raw_string_ostream OS(OutString); assert(!Subset.empty() && "Unexpected empty range"); OS << "Recipe with invalid costs prevented vectorization at VF=("; for (const auto &Pair : Subset) OS << (Pair.second == Subset.front().second ? "" : ", ") << Pair.second; OS << "):"; if (Opcode == Instruction::Call) { StringRef Name = ""; if (auto *Int = dyn_cast(R)) { Name = Int->getIntrinsicName(); } else { auto *WidenCall = dyn_cast(R); Function *CalledFn = WidenCall ? WidenCall->getCalledScalarFunction() : cast(R->getOperand(R->getNumOperands() - 1) ->getLiveInIRValue()); Name = CalledFn->getName(); } OS << " call to " << Name; } else OS << " " << Instruction::getOpcodeName(Opcode); reportVectorizationInfo(OutString, "InvalidCost", ORE, OrigLoop, nullptr, R->getDebugLoc()); Tail = Tail.drop_front(Subset.size()); Subset = {}; } else // Grow the subset by one element Subset = Tail.take_front(Subset.size() + 1); } while (!Tail.empty()); } /// Check if any recipe of \p Plan will generate a vector value, which will be /// assigned a vector register. static bool willGenerateVectors(VPlan &Plan, ElementCount VF, const TargetTransformInfo &TTI) { assert(VF.isVector() && "Checking a scalar VF?"); VPTypeAnalysis TypeInfo(Plan); DenseSet EphemeralRecipes; collectEphemeralRecipesForVPlan(Plan, EphemeralRecipes); // Set of already visited types. DenseSet Visited; for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly( vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) { for (VPRecipeBase &R : *VPBB) { if (EphemeralRecipes.contains(&R)) continue; // Continue early if the recipe is considered to not produce a vector // result. Note that this includes VPInstruction where some opcodes may // produce a vector, to preserve existing behavior as VPInstructions model // aspects not directly mapped to existing IR instructions. switch (R.getVPDefID()) { case VPDef::VPDerivedIVSC: case VPDef::VPScalarIVStepsSC: case VPDef::VPReplicateSC: case VPDef::VPInstructionSC: case VPDef::VPCanonicalIVPHISC: case VPDef::VPVectorPointerSC: case VPDef::VPVectorEndPointerSC: case VPDef::VPExpandSCEVSC: case VPDef::VPEVLBasedIVPHISC: case VPDef::VPPredInstPHISC: case VPDef::VPBranchOnMaskSC: continue; case VPDef::VPReductionSC: case VPDef::VPActiveLaneMaskPHISC: case VPDef::VPWidenCallSC: case VPDef::VPWidenCanonicalIVSC: case VPDef::VPWidenCastSC: case VPDef::VPWidenGEPSC: case VPDef::VPWidenIntrinsicSC: case VPDef::VPWidenSC: case VPDef::VPWidenSelectSC: case VPDef::VPBlendSC: case VPDef::VPFirstOrderRecurrencePHISC: case VPDef::VPHistogramSC: case VPDef::VPWidenPHISC: case VPDef::VPWidenIntOrFpInductionSC: case VPDef::VPWidenPointerInductionSC: case VPDef::VPReductionPHISC: case VPDef::VPInterleaveSC: case VPDef::VPWidenLoadEVLSC: case VPDef::VPWidenLoadSC: case VPDef::VPWidenStoreEVLSC: case VPDef::VPWidenStoreSC: break; default: llvm_unreachable("unhandled recipe"); } auto WillGenerateTargetVectors = [&TTI, VF](Type *VectorTy) { unsigned NumLegalParts = TTI.getNumberOfParts(VectorTy); if (!NumLegalParts) return false; if (VF.isScalable()) { // is assumed to be profitable over iN because // scalable registers are a distinct register class from scalar // ones. If we ever find a target which wants to lower scalable // vectors back to scalars, we'll need to update this code to // explicitly ask TTI about the register class uses for each part. return NumLegalParts <= VF.getKnownMinValue(); } // Two or more elements that share a register - are vectorized. return NumLegalParts < VF.getFixedValue(); }; // If no def nor is a store, e.g., branches, continue - no value to check. if (R.getNumDefinedValues() == 0 && !isa( &R)) continue; // For multi-def recipes, currently only interleaved loads, suffice to // check first def only. // For stores check their stored value; for interleaved stores suffice // the check first stored value only. In all cases this is the second // operand. VPValue *ToCheck = R.getNumDefinedValues() >= 1 ? R.getVPValue(0) : R.getOperand(1); Type *ScalarTy = TypeInfo.inferScalarType(ToCheck); if (!Visited.insert({ScalarTy}).second) continue; Type *WideTy = toVectorizedTy(ScalarTy, VF); if (any_of(getContainedTypes(WideTy), WillGenerateTargetVectors)) return true; } } return false; } static bool hasReplicatorRegion(VPlan &Plan) { return any_of(VPBlockUtils::blocksOnly(vp_depth_first_shallow( Plan.getVectorLoopRegion()->getEntry())), [](auto *VPRB) { return VPRB->isReplicator(); }); } #ifndef NDEBUG VectorizationFactor LoopVectorizationPlanner::selectVectorizationFactor() { InstructionCost ExpectedCost = CM.expectedCost(ElementCount::getFixed(1)); LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n"); assert(ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop"); assert( any_of(VPlans, [](std::unique_ptr &P) { return P->hasScalarVFOnly(); }) && "Expected Scalar VF to be a candidate"); const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost, ExpectedCost); VectorizationFactor ChosenFactor = ScalarCost; bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled; if (ForceVectorization && (VPlans.size() > 1 || !VPlans[0]->hasScalarVFOnly())) { // Ignore scalar width, because the user explicitly wants vectorization. // Initialize cost to max so that VF = 2 is, at least, chosen during cost // evaluation. ChosenFactor.Cost = InstructionCost::getMax(); } for (auto &P : VPlans) { ArrayRef VFs(P->vectorFactors().begin(), P->vectorFactors().end()); SmallVector RUs; if (CM.useMaxBandwidth(TargetTransformInfo::RGK_ScalableVector) || CM.useMaxBandwidth(TargetTransformInfo::RGK_FixedWidthVector)) RUs = calculateRegisterUsageForPlan(*P, VFs, TTI, CM.ValuesToIgnore); for (unsigned I = 0; I < VFs.size(); I++) { ElementCount VF = VFs[I]; // The cost for scalar VF=1 is already calculated, so ignore it. if (VF.isScalar()) continue; /// If the register pressure needs to be considered for VF, /// don't consider the VF as valid if it exceeds the number /// of registers for the target. if (CM.shouldCalculateRegPressureForVF(VF) && RUs[I].exceedsMaxNumRegs(TTI, ForceTargetNumVectorRegs)) continue; InstructionCost C = CM.expectedCost(VF); // Add on other costs that are modelled in VPlan, but not in the legacy // cost model. VPCostContext CostCtx(CM.TTI, *CM.TLI, *P, CM, CM.CostKind); VPRegionBlock *VectorRegion = P->getVectorLoopRegion(); assert(VectorRegion && "Expected to have a vector region!"); for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly( vp_depth_first_shallow(VectorRegion->getEntry()))) { for (VPRecipeBase &R : *VPBB) { auto *VPI = dyn_cast(&R); if (!VPI) continue; switch (VPI->getOpcode()) { case VPInstruction::ActiveLaneMask: case VPInstruction::ExplicitVectorLength: C += VPI->cost(VF, CostCtx); break; default: break; } } } VectorizationFactor Candidate(VF, C, ScalarCost.ScalarCost); unsigned Width = estimateElementCount(Candidate.Width, CM.getVScaleForTuning()); LLVM_DEBUG(dbgs() << "LV: Vector loop of width " << VF << " costs: " << (Candidate.Cost / Width)); if (VF.isScalable()) LLVM_DEBUG(dbgs() << " (assuming a minimum vscale of " << CM.getVScaleForTuning().value_or(1) << ")"); LLVM_DEBUG(dbgs() << ".\n"); if (!ForceVectorization && !willGenerateVectors(*P, VF, TTI)) { LLVM_DEBUG( dbgs() << "LV: Not considering vector loop of width " << VF << " because it will not generate any vector instructions.\n"); continue; } if (CM.OptForSize && !ForceVectorization && hasReplicatorRegion(*P)) { LLVM_DEBUG( dbgs() << "LV: Not considering vector loop of width " << VF << " because it would cause replicated blocks to be generated," << " which isn't allowed when optimizing for size.\n"); continue; } if (isMoreProfitable(Candidate, ChosenFactor, P->hasScalarTail())) ChosenFactor = Candidate; } } if (!EnableCondStoresVectorization && CM.hasPredStores()) { reportVectorizationFailure( "There are conditional stores.", "store that is conditionally executed prevents vectorization", "ConditionalStore", ORE, OrigLoop); ChosenFactor = ScalarCost; } LLVM_DEBUG(if (ForceVectorization && !ChosenFactor.Width.isScalar() && !isMoreProfitable(ChosenFactor, ScalarCost, !CM.foldTailByMasking())) dbgs() << "LV: Vectorization seems to be not beneficial, " << "but was forced by a user.\n"); return ChosenFactor; } #endif bool LoopVectorizationPlanner::isCandidateForEpilogueVectorization( ElementCount VF) const { // Cross iteration phis such as fixed-order recurrences and FMaxNum/FMinNum // reductions need special handling and are currently unsupported. if (any_of(OrigLoop->getHeader()->phis(), [&](PHINode &Phi) { if (!Legal->isReductionVariable(&Phi)) return Legal->isFixedOrderRecurrence(&Phi); RecurKind RK = Legal->getRecurrenceDescriptor(&Phi).getRecurrenceKind(); return RK == RecurKind::FMinNum || RK == RecurKind::FMaxNum; })) return false; // Phis with uses outside of the loop require special handling and are // currently unsupported. for (const auto &Entry : Legal->getInductionVars()) { // Look for uses of the value of the induction at the last iteration. Value *PostInc = Entry.first->getIncomingValueForBlock(OrigLoop->getLoopLatch()); for (User *U : PostInc->users()) if (!OrigLoop->contains(cast(U))) return false; // Look for uses of penultimate value of the induction. for (User *U : Entry.first->users()) if (!OrigLoop->contains(cast(U))) return false; } // Epilogue vectorization code has not been auditted to ensure it handles // non-latch exits properly. It may be fine, but it needs auditted and // tested. // TODO: Add support for loops with an early exit. if (OrigLoop->getExitingBlock() != OrigLoop->getLoopLatch()) return false; return true; } bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable( const ElementCount VF, const unsigned IC) const { // FIXME: We need a much better cost-model to take different parameters such // as register pressure, code size increase and cost of extra branches into // account. For now we apply a very crude heuristic and only consider loops // with vectorization factors larger than a certain value. // Allow the target to opt out entirely. if (!TTI.preferEpilogueVectorization()) return false; // We also consider epilogue vectorization unprofitable for targets that don't // consider interleaving beneficial (eg. MVE). if (TTI.getMaxInterleaveFactor(VF) <= 1) return false; // TODO: PR #108190 introduced a discrepancy between fixed-width and scalable // VFs when deciding profitability. // See related "TODO: extend to support scalable VFs." in // selectEpilogueVectorizationFactor. unsigned Multiplier = VF.isFixed() ? IC : 1; unsigned MinVFThreshold = EpilogueVectorizationMinVF.getNumOccurrences() > 0 ? EpilogueVectorizationMinVF : TTI.getEpilogueVectorizationMinVF(); return estimateElementCount(VF * Multiplier, VScaleForTuning) >= MinVFThreshold; } VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor( const ElementCount MainLoopVF, unsigned IC) { VectorizationFactor Result = VectorizationFactor::Disabled(); if (!EnableEpilogueVectorization) { LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is disabled.\n"); return Result; } if (!CM.isScalarEpilogueAllowed()) { LLVM_DEBUG(dbgs() << "LEV: Unable to vectorize epilogue because no " "epilogue is allowed.\n"); return Result; } // Not really a cost consideration, but check for unsupported cases here to // simplify the logic. if (!isCandidateForEpilogueVectorization(MainLoopVF)) { LLVM_DEBUG(dbgs() << "LEV: Unable to vectorize epilogue because the loop " "is not a supported candidate.\n"); return Result; } if (EpilogueVectorizationForceVF > 1) { LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization factor is forced.\n"); ElementCount ForcedEC = ElementCount::getFixed(EpilogueVectorizationForceVF); if (hasPlanWithVF(ForcedEC)) return {ForcedEC, 0, 0}; LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization forced factor is not " "viable.\n"); return Result; } if (OrigLoop->getHeader()->getParent()->hasOptSize()) { LLVM_DEBUG( dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n"); return Result; } if (!CM.isEpilogueVectorizationProfitable(MainLoopVF, IC)) { LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is not profitable for " "this loop\n"); return Result; } // If MainLoopVF = vscale x 2, and vscale is expected to be 4, then we know // the main loop handles 8 lanes per iteration. We could still benefit from // vectorizing the epilogue loop with VF=4. ElementCount EstimatedRuntimeVF = ElementCount::getFixed( estimateElementCount(MainLoopVF, CM.getVScaleForTuning())); ScalarEvolution &SE = *PSE.getSE(); Type *TCType = Legal->getWidestInductionType(); const SCEV *RemainingIterations = nullptr; unsigned MaxTripCount = 0; const SCEV *TC = vputils::getSCEVExprForVPValue(getPlanFor(MainLoopVF).getTripCount(), SE); assert(!isa(TC) && "Trip count SCEV must be computable"); RemainingIterations = SE.getURemExpr(TC, SE.getElementCount(TCType, MainLoopVF * IC)); // No iterations left to process in the epilogue. if (RemainingIterations->isZero()) return Result; if (MainLoopVF.isFixed()) { MaxTripCount = MainLoopVF.getFixedValue() * IC - 1; if (SE.isKnownPredicate(CmpInst::ICMP_ULT, RemainingIterations, SE.getConstant(TCType, MaxTripCount))) { MaxTripCount = SE.getUnsignedRangeMax(RemainingIterations).getZExtValue(); } LLVM_DEBUG(dbgs() << "LEV: Maximum Trip Count for Epilogue: " << MaxTripCount << "\n"); } for (auto &NextVF : ProfitableVFs) { // Skip candidate VFs without a corresponding VPlan. if (!hasPlanWithVF(NextVF.Width)) continue; // Skip candidate VFs with widths >= the (estimated) runtime VF (scalable // vectors) or > the VF of the main loop (fixed vectors). if ((!NextVF.Width.isScalable() && MainLoopVF.isScalable() && ElementCount::isKnownGE(NextVF.Width, EstimatedRuntimeVF)) || (NextVF.Width.isScalable() && ElementCount::isKnownGE(NextVF.Width, MainLoopVF)) || (!NextVF.Width.isScalable() && !MainLoopVF.isScalable() && ElementCount::isKnownGT(NextVF.Width, MainLoopVF))) continue; // If NextVF is greater than the number of remaining iterations, the // epilogue loop would be dead. Skip such factors. if (RemainingIterations && !NextVF.Width.isScalable()) { if (SE.isKnownPredicate( CmpInst::ICMP_UGT, SE.getConstant(TCType, NextVF.Width.getFixedValue()), RemainingIterations)) continue; } if (Result.Width.isScalar() || isMoreProfitable(NextVF, Result, MaxTripCount, !CM.foldTailByMasking())) Result = NextVF; } if (Result != VectorizationFactor::Disabled()) LLVM_DEBUG(dbgs() << "LEV: Vectorizing epilogue loop with VF = " << Result.Width << "\n"); return Result; } std::pair LoopVectorizationCostModel::getSmallestAndWidestTypes() { unsigned MinWidth = -1U; unsigned MaxWidth = 8; const DataLayout &DL = TheFunction->getDataLayout(); // For in-loop reductions, no element types are added to ElementTypesInLoop // if there are no loads/stores in the loop. In this case, check through the // reduction variables to determine the maximum width. if (ElementTypesInLoop.empty() && !Legal->getReductionVars().empty()) { for (const auto &PhiDescriptorPair : Legal->getReductionVars()) { const RecurrenceDescriptor &RdxDesc = PhiDescriptorPair.second; // When finding the min width used by the recurrence we need to account // for casts on the input operands of the recurrence. MinWidth = std::min( MinWidth, std::min(RdxDesc.getMinWidthCastToRecurrenceTypeInBits(), RdxDesc.getRecurrenceType()->getScalarSizeInBits())); MaxWidth = std::max(MaxWidth, RdxDesc.getRecurrenceType()->getScalarSizeInBits()); } } else { for (Type *T : ElementTypesInLoop) { MinWidth = std::min( MinWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedValue()); MaxWidth = std::max( MaxWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedValue()); } } return {MinWidth, MaxWidth}; } void LoopVectorizationCostModel::collectElementTypesForWidening() { ElementTypesInLoop.clear(); // For each block. for (BasicBlock *BB : TheLoop->blocks()) { // For each instruction in the loop. for (Instruction &I : BB->instructionsWithoutDebug()) { Type *T = I.getType(); // Skip ignored values. if (ValuesToIgnore.count(&I)) continue; // Only examine Loads, Stores and PHINodes. if (!isa(I) && !isa(I) && !isa(I)) continue; // Examine PHI nodes that are reduction variables. Update the type to // account for the recurrence type. if (auto *PN = dyn_cast(&I)) { if (!Legal->isReductionVariable(PN)) continue; const RecurrenceDescriptor &RdxDesc = Legal->getRecurrenceDescriptor(PN); if (PreferInLoopReductions || useOrderedReductions(RdxDesc) || TTI.preferInLoopReduction(RdxDesc.getRecurrenceKind(), RdxDesc.getRecurrenceType())) continue; T = RdxDesc.getRecurrenceType(); } // Examine the stored values. if (auto *ST = dyn_cast(&I)) T = ST->getValueOperand()->getType(); assert(T->isSized() && "Expected the load/store/recurrence type to be sized"); ElementTypesInLoop.insert(T); } } } unsigned LoopVectorizationPlanner::selectInterleaveCount(VPlan &Plan, ElementCount VF, InstructionCost LoopCost) { // -- The interleave heuristics -- // We interleave the loop in order to expose ILP and reduce the loop overhead. // There are many micro-architectural considerations that we can't predict // at this level. For example, frontend pressure (on decode or fetch) due to // code size, or the number and capabilities of the execution ports. // // We use the following heuristics to select the interleave count: // 1. If the code has reductions, then we interleave to break the cross // iteration dependency. // 2. If the loop is really small, then we interleave to reduce the loop // overhead. // 3. We don't interleave if we think that we will spill registers to memory // due to the increased register pressure. if (!CM.isScalarEpilogueAllowed()) return 1; if (any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(), IsaPred)) { LLVM_DEBUG(dbgs() << "LV: Preference for VP intrinsics indicated. " "Unroll factor forced to be 1.\n"); return 1; } // We used the distance for the interleave count. if (!Legal->isSafeForAnyVectorWidth()) return 1; // We don't attempt to perform interleaving for loops with uncountable early // exits because the VPInstruction::AnyOf code cannot currently handle // multiple parts. if (Plan.hasEarlyExit()) return 1; const bool HasReductions = any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(), IsaPred); // If we did not calculate the cost for VF (because the user selected the VF) // then we calculate the cost of VF here. if (LoopCost == 0) { if (VF.isScalar()) LoopCost = CM.expectedCost(VF); else LoopCost = cost(Plan, VF); assert(LoopCost.isValid() && "Expected to have chosen a VF with valid cost"); // Loop body is free and there is no need for interleaving. if (LoopCost == 0) return 1; } VPRegisterUsage R = calculateRegisterUsageForPlan(Plan, {VF}, TTI, CM.ValuesToIgnore)[0]; // We divide by these constants so assume that we have at least one // instruction that uses at least one register. for (auto &Pair : R.MaxLocalUsers) { Pair.second = std::max(Pair.second, 1U); } // We calculate the interleave count using the following formula. // Subtract the number of loop invariants from the number of available // registers. These registers are used by all of the interleaved instances. // Next, divide the remaining registers by the number of registers that is // required by the loop, in order to estimate how many parallel instances // fit without causing spills. All of this is rounded down if necessary to be // a power of two. We want power of two interleave count to simplify any // addressing operations or alignment considerations. // We also want power of two interleave counts to ensure that the induction // variable of the vector loop wraps to zero, when tail is folded by masking; // this currently happens when OptForSize, in which case IC is set to 1 above. unsigned IC = UINT_MAX; for (const auto &Pair : R.MaxLocalUsers) { unsigned TargetNumRegisters = TTI.getNumberOfRegisters(Pair.first); LLVM_DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters << " registers of " << TTI.getRegisterClassName(Pair.first) << " register class\n"); if (VF.isScalar()) { if (ForceTargetNumScalarRegs.getNumOccurrences() > 0) TargetNumRegisters = ForceTargetNumScalarRegs; } else { if (ForceTargetNumVectorRegs.getNumOccurrences() > 0) TargetNumRegisters = ForceTargetNumVectorRegs; } unsigned MaxLocalUsers = Pair.second; unsigned LoopInvariantRegs = 0; if (R.LoopInvariantRegs.contains(Pair.first)) LoopInvariantRegs = R.LoopInvariantRegs[Pair.first]; unsigned TmpIC = llvm::bit_floor((TargetNumRegisters - LoopInvariantRegs) / MaxLocalUsers); // Don't count the induction variable as interleaved. if (EnableIndVarRegisterHeur) { TmpIC = llvm::bit_floor((TargetNumRegisters - LoopInvariantRegs - 1) / std::max(1U, (MaxLocalUsers - 1))); } IC = std::min(IC, TmpIC); } // Clamp the interleave ranges to reasonable counts. unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF); // Check if the user has overridden the max. if (VF.isScalar()) { if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0) MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor; } else { if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0) MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor; } // Try to get the exact trip count, or an estimate based on profiling data or // ConstantMax from PSE, failing that. auto BestKnownTC = getSmallBestKnownTC(PSE, OrigLoop); // For fixed length VFs treat a scalable trip count as unknown. if (BestKnownTC && (BestKnownTC->isFixed() || VF.isScalable())) { // Re-evaluate trip counts and VFs to be in the same numerical space. unsigned AvailableTC = estimateElementCount(*BestKnownTC, CM.getVScaleForTuning()); unsigned EstimatedVF = estimateElementCount(VF, CM.getVScaleForTuning()); // At least one iteration must be scalar when this constraint holds. So the // maximum available iterations for interleaving is one less. if (CM.requiresScalarEpilogue(VF.isVector())) --AvailableTC; unsigned InterleaveCountLB = bit_floor(std::max( 1u, std::min(AvailableTC / (EstimatedVF * 2), MaxInterleaveCount))); if (getSmallConstantTripCount(PSE.getSE(), OrigLoop).isNonZero()) { // If the best known trip count is exact, we select between two // prospective ICs, where // // 1) the aggressive IC is capped by the trip count divided by VF // 2) the conservative IC is capped by the trip count divided by (VF * 2) // // The final IC is selected in a way that the epilogue loop trip count is // minimized while maximizing the IC itself, so that we either run the // vector loop at least once if it generates a small epilogue loop, or // else we run the vector loop at least twice. unsigned InterleaveCountUB = bit_floor(std::max( 1u, std::min(AvailableTC / EstimatedVF, MaxInterleaveCount))); MaxInterleaveCount = InterleaveCountLB; if (InterleaveCountUB != InterleaveCountLB) { unsigned TailTripCountUB = (AvailableTC % (EstimatedVF * InterleaveCountUB)); unsigned TailTripCountLB = (AvailableTC % (EstimatedVF * InterleaveCountLB)); // If both produce same scalar tail, maximize the IC to do the same work // in fewer vector loop iterations if (TailTripCountUB == TailTripCountLB) MaxInterleaveCount = InterleaveCountUB; } } else { // If trip count is an estimated compile time constant, limit the // IC to be capped by the trip count divided by VF * 2, such that the // vector loop runs at least twice to make interleaving seem profitable // when there is an epilogue loop present. Since exact Trip count is not // known we choose to be conservative in our IC estimate. MaxInterleaveCount = InterleaveCountLB; } } assert(MaxInterleaveCount > 0 && "Maximum interleave count must be greater than 0"); // Clamp the calculated IC to be between the 1 and the max interleave count // that the target and trip count allows. if (IC > MaxInterleaveCount) IC = MaxInterleaveCount; else // Make sure IC is greater than 0. IC = std::max(1u, IC); assert(IC > 0 && "Interleave count must be greater than 0."); // Interleave if we vectorized this loop and there is a reduction that could // benefit from interleaving. if (VF.isVector() && HasReductions) { LLVM_DEBUG(dbgs() << "LV: Interleaving because of reductions.\n"); return IC; } // For any scalar loop that either requires runtime checks or predication we // are better off leaving this to the unroller. Note that if we've already // vectorized the loop we will have done the runtime check and so interleaving // won't require further checks. bool ScalarInterleavingRequiresPredication = (VF.isScalar() && any_of(OrigLoop->blocks(), [this](BasicBlock *BB) { return Legal->blockNeedsPredication(BB); })); bool ScalarInterleavingRequiresRuntimePointerCheck = (VF.isScalar() && Legal->getRuntimePointerChecking()->Need); // We want to interleave small loops in order to reduce the loop overhead and // potentially expose ILP opportunities. LLVM_DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n' << "LV: IC is " << IC << '\n' << "LV: VF is " << VF << '\n'); const bool AggressivelyInterleaveReductions = TTI.enableAggressiveInterleaving(HasReductions); if (!ScalarInterleavingRequiresRuntimePointerCheck && !ScalarInterleavingRequiresPredication && LoopCost < SmallLoopCost) { // We assume that the cost overhead is 1 and we use the cost model // to estimate the cost of the loop and interleave until the cost of the // loop overhead is about 5% of the cost of the loop. unsigned SmallIC = std::min(IC, (unsigned)llvm::bit_floor( SmallLoopCost / LoopCost.getValue())); // Interleave until store/load ports (estimated by max interleave count) are // saturated. unsigned NumStores = 0; unsigned NumLoads = 0; for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly( vp_depth_first_deep(Plan.getVectorLoopRegion()->getEntry()))) { for (VPRecipeBase &R : *VPBB) { if (isa(&R)) { NumLoads++; continue; } if (isa(&R)) { NumStores++; continue; } if (auto *InterleaveR = dyn_cast(&R)) { if (unsigned StoreOps = InterleaveR->getNumStoreOperands()) NumStores += StoreOps; else NumLoads += InterleaveR->getNumDefinedValues(); continue; } if (auto *RepR = dyn_cast(&R)) { NumLoads += isa(RepR->getUnderlyingInstr()); NumStores += isa(RepR->getUnderlyingInstr()); continue; } if (isa(&R)) { NumLoads++; NumStores++; continue; } } } unsigned StoresIC = IC / (NumStores ? NumStores : 1); unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1); // There is little point in interleaving for reductions containing selects // and compares when VF=1 since it may just create more overhead than it's // worth for loops with small trip counts. This is because we still have to // do the final reduction after the loop. bool HasSelectCmpReductions = HasReductions && any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(), [](VPRecipeBase &R) { auto *RedR = dyn_cast(&R); return RedR && (RecurrenceDescriptor::isAnyOfRecurrenceKind( RedR->getRecurrenceKind()) || RecurrenceDescriptor::isFindIVRecurrenceKind( RedR->getRecurrenceKind())); }); if (HasSelectCmpReductions) { LLVM_DEBUG(dbgs() << "LV: Not interleaving select-cmp reductions.\n"); return 1; } // If we have a scalar reduction (vector reductions are already dealt with // by this point), we can increase the critical path length if the loop // we're interleaving is inside another loop. For tree-wise reductions // set the limit to 2, and for ordered reductions it's best to disable // interleaving entirely. if (HasReductions && OrigLoop->getLoopDepth() > 1) { bool HasOrderedReductions = any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(), [](VPRecipeBase &R) { auto *RedR = dyn_cast(&R); return RedR && RedR->isOrdered(); }); if (HasOrderedReductions) { LLVM_DEBUG( dbgs() << "LV: Not interleaving scalar ordered reductions.\n"); return 1; } unsigned F = MaxNestedScalarReductionIC; SmallIC = std::min(SmallIC, F); StoresIC = std::min(StoresIC, F); LoadsIC = std::min(LoadsIC, F); } if (EnableLoadStoreRuntimeInterleave && std::max(StoresIC, LoadsIC) > SmallIC) { LLVM_DEBUG( dbgs() << "LV: Interleaving to saturate store or load ports.\n"); return std::max(StoresIC, LoadsIC); } // If there are scalar reductions and TTI has enabled aggressive // interleaving for reductions, we will interleave to expose ILP. if (VF.isScalar() && AggressivelyInterleaveReductions) { LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n"); // Interleave no less than SmallIC but not as aggressive as the normal IC // to satisfy the rare situation when resources are too limited. return std::max(IC / 2, SmallIC); } LLVM_DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n"); return SmallIC; } // Interleave if this is a large loop (small loops are already dealt with by // this point) that could benefit from interleaving. if (AggressivelyInterleaveReductions) { LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n"); return IC; } LLVM_DEBUG(dbgs() << "LV: Not Interleaving.\n"); return 1; } bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I, ElementCount VF) { // TODO: Cost model for emulated masked load/store is completely // broken. This hack guides the cost model to use an artificially // high enough value to practically disable vectorization with such // operations, except where previously deployed legality hack allowed // using very low cost values. This is to avoid regressions coming simply // from moving "masked load/store" check from legality to cost model. // Masked Load/Gather emulation was previously never allowed. // Limited number of Masked Store/Scatter emulation was allowed. assert((isPredicatedInst(I)) && "Expecting a scalar emulated instruction"); return isa(I) || (isa(I) && NumPredStores > NumberOfStoresToPredicate); } void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) { assert(VF.isVector() && "Expected VF >= 2"); // If we've already collected the instructions to scalarize or the predicated // BBs after vectorization, there's nothing to do. Collection may already have // occurred if we have a user-selected VF and are now computing the expected // cost for interleaving. if (InstsToScalarize.contains(VF) || PredicatedBBsAfterVectorization.contains(VF)) return; // Initialize a mapping for VF in InstsToScalalarize. If we find that it's // not profitable to scalarize any instructions, the presence of VF in the // map will indicate that we've analyzed it already. ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF]; // Find all the instructions that are scalar with predication in the loop and // determine if it would be better to not if-convert the blocks they are in. // If so, we also record the instructions to scalarize. for (BasicBlock *BB : TheLoop->blocks()) { if (!blockNeedsPredicationForAnyReason(BB)) continue; for (Instruction &I : *BB) if (isScalarWithPredication(&I, VF)) { ScalarCostsTy ScalarCosts; // Do not apply discount logic for: // 1. Scalars after vectorization, as there will only be a single copy // of the instruction. // 2. Scalable VF, as that would lead to invalid scalarization costs. // 3. Emulated masked memrefs, if a hacked cost is needed. if (!isScalarAfterVectorization(&I, VF) && !VF.isScalable() && !useEmulatedMaskMemRefHack(&I, VF) && computePredInstDiscount(&I, ScalarCosts, VF) >= 0) { for (const auto &[I, IC] : ScalarCosts) ScalarCostsVF.insert({I, IC}); // Check if we decided to scalarize a call. If so, update the widening // decision of the call to CM_Scalarize with the computed scalar cost. for (const auto &[I, Cost] : ScalarCosts) { auto *CI = dyn_cast(I); if (!CI || !CallWideningDecisions.contains({CI, VF})) continue; CallWideningDecisions[{CI, VF}].Kind = CM_Scalarize; CallWideningDecisions[{CI, VF}].Cost = Cost; } } // Remember that BB will remain after vectorization. PredicatedBBsAfterVectorization[VF].insert(BB); for (auto *Pred : predecessors(BB)) { if (Pred->getSingleSuccessor() == BB) PredicatedBBsAfterVectorization[VF].insert(Pred); } } } } InstructionCost LoopVectorizationCostModel::computePredInstDiscount( Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) { assert(!isUniformAfterVectorization(PredInst, VF) && "Instruction marked uniform-after-vectorization will be predicated"); // Initialize the discount to zero, meaning that the scalar version and the // vector version cost the same. InstructionCost Discount = 0; // Holds instructions to analyze. The instructions we visit are mapped in // ScalarCosts. Those instructions are the ones that would be scalarized if // we find that the scalar version costs less. SmallVector Worklist; // Returns true if the given instruction can be scalarized. auto CanBeScalarized = [&](Instruction *I) -> bool { // We only attempt to scalarize instructions forming a single-use chain // from the original predicated block that would otherwise be vectorized. // Although not strictly necessary, we give up on instructions we know will // already be scalar to avoid traversing chains that are unlikely to be // beneficial. if (!I->hasOneUse() || PredInst->getParent() != I->getParent() || isScalarAfterVectorization(I, VF)) return false; // If the instruction is scalar with predication, it will be analyzed // separately. We ignore it within the context of PredInst. if (isScalarWithPredication(I, VF)) return false; // If any of the instruction's operands are uniform after vectorization, // the instruction cannot be scalarized. This prevents, for example, a // masked load from being scalarized. // // We assume we will only emit a value for lane zero of an instruction // marked uniform after vectorization, rather than VF identical values. // Thus, if we scalarize an instruction that uses a uniform, we would // create uses of values corresponding to the lanes we aren't emitting code // for. This behavior can be changed by allowing getScalarValue to clone // the lane zero values for uniforms rather than asserting. for (Use &U : I->operands()) if (auto *J = dyn_cast(U.get())) if (isUniformAfterVectorization(J, VF)) return false; // Otherwise, we can scalarize the instruction. return true; }; // Compute the expected cost discount from scalarizing the entire expression // feeding the predicated instruction. We currently only consider expressions // that are single-use instruction chains. Worklist.push_back(PredInst); while (!Worklist.empty()) { Instruction *I = Worklist.pop_back_val(); // If we've already analyzed the instruction, there's nothing to do. if (ScalarCosts.contains(I)) continue; // Cannot scalarize fixed-order recurrence phis at the moment. if (isa(I) && Legal->isFixedOrderRecurrence(cast(I))) continue; // Compute the cost of the vector instruction. Note that this cost already // includes the scalarization overhead of the predicated instruction. InstructionCost VectorCost = getInstructionCost(I, VF); // Compute the cost of the scalarized instruction. This cost is the cost of // the instruction as if it wasn't if-converted and instead remained in the // predicated block. We will scale this cost by block probability after // computing the scalarization overhead. InstructionCost ScalarCost = VF.getFixedValue() * getInstructionCost(I, ElementCount::getFixed(1)); // Compute the scalarization overhead of needed insertelement instructions // and phi nodes. if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) { Type *WideTy = toVectorizedTy(I->getType(), VF); for (Type *VectorTy : getContainedTypes(WideTy)) { ScalarCost += TTI.getScalarizationOverhead( cast(VectorTy), APInt::getAllOnes(VF.getFixedValue()), /*Insert=*/true, /*Extract=*/false, CostKind); } ScalarCost += VF.getFixedValue() * TTI.getCFInstrCost(Instruction::PHI, CostKind); } // Compute the scalarization overhead of needed extractelement // instructions. For each of the instruction's operands, if the operand can // be scalarized, add it to the worklist; otherwise, account for the // overhead. for (Use &U : I->operands()) if (auto *J = dyn_cast(U.get())) { assert(canVectorizeTy(J->getType()) && "Instruction has non-scalar type"); if (CanBeScalarized(J)) Worklist.push_back(J); else if (needsExtract(J, VF)) { Type *WideTy = toVectorizedTy(J->getType(), VF); for (Type *VectorTy : getContainedTypes(WideTy)) { ScalarCost += TTI.getScalarizationOverhead( cast(VectorTy), APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ false, /*Extract*/ true, CostKind); } } } // Scale the total scalar cost by block probability. ScalarCost /= getPredBlockCostDivisor(CostKind); // Compute the discount. A non-negative discount means the vector version // of the instruction costs more, and scalarizing would be beneficial. Discount += VectorCost - ScalarCost; ScalarCosts[I] = ScalarCost; } return Discount; } InstructionCost LoopVectorizationCostModel::expectedCost(ElementCount VF) { InstructionCost Cost; // If the vector loop gets executed exactly once with the given VF, ignore the // costs of comparison and induction instructions, as they'll get simplified // away. SmallPtrSet ValuesToIgnoreForVF; auto TC = getSmallConstantTripCount(PSE.getSE(), TheLoop); if (TC == VF && !foldTailByMasking()) addFullyUnrolledInstructionsToIgnore(TheLoop, Legal->getInductionVars(), ValuesToIgnoreForVF); // For each block. for (BasicBlock *BB : TheLoop->blocks()) { InstructionCost BlockCost; // For each instruction in the old loop. for (Instruction &I : BB->instructionsWithoutDebug()) { // Skip ignored values. if (ValuesToIgnore.count(&I) || ValuesToIgnoreForVF.count(&I) || (VF.isVector() && VecValuesToIgnore.count(&I))) continue; InstructionCost C = getInstructionCost(&I, VF); // Check if we should override the cost. if (C.isValid() && ForceTargetInstructionCost.getNumOccurrences() > 0) C = InstructionCost(ForceTargetInstructionCost); BlockCost += C; LLVM_DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF " << VF << " For instruction: " << I << '\n'); } // If we are vectorizing a predicated block, it will have been // if-converted. This means that the block's instructions (aside from // stores and instructions that may divide by zero) will now be // unconditionally executed. For the scalar case, we may not always execute // the predicated block, if it is an if-else block. Thus, scale the block's // cost by the probability of executing it. blockNeedsPredication from // Legal is used so as to not include all blocks in tail folded loops. if (VF.isScalar() && Legal->blockNeedsPredication(BB)) BlockCost /= getPredBlockCostDivisor(CostKind); Cost += BlockCost; } return Cost; } /// Gets Address Access SCEV after verifying that the access pattern /// is loop invariant except the induction variable dependence. /// /// This SCEV can be sent to the Target in order to estimate the address /// calculation cost. static const SCEV *getAddressAccessSCEV( Value *Ptr, LoopVectorizationLegality *Legal, PredicatedScalarEvolution &PSE, const Loop *TheLoop) { auto *Gep = dyn_cast(Ptr); if (!Gep) return nullptr; // We are looking for a gep with all loop invariant indices except for one // which should be an induction variable. auto *SE = PSE.getSE(); unsigned NumOperands = Gep->getNumOperands(); for (unsigned Idx = 1; Idx < NumOperands; ++Idx) { Value *Opd = Gep->getOperand(Idx); if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) && !Legal->isInductionVariable(Opd)) return nullptr; } // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV. return PSE.getSCEV(Ptr); } InstructionCost LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I, ElementCount VF) { assert(VF.isVector() && "Scalarization cost of instruction implies vectorization."); if (VF.isScalable()) return InstructionCost::getInvalid(); Type *ValTy = getLoadStoreType(I); auto *SE = PSE.getSE(); unsigned AS = getLoadStoreAddressSpace(I); Value *Ptr = getLoadStorePointerOperand(I); Type *PtrTy = toVectorTy(Ptr->getType(), VF); // NOTE: PtrTy is a vector to signal `TTI::getAddressComputationCost` // that it is being called from this specific place. // Figure out whether the access is strided and get the stride value // if it's known in compile time const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, PSE, TheLoop); // Get the cost of the scalar memory instruction and address computation. InstructionCost Cost = VF.getFixedValue() * TTI.getAddressComputationCost( PtrTy, SE, PtrSCEV, CostKind); // Don't pass *I here, since it is scalar but will actually be part of a // vectorized loop where the user of it is a vectorized instruction. const Align Alignment = getLoadStoreAlignment(I); Cost += VF.getFixedValue() * TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment, AS, CostKind); // Get the overhead of the extractelement and insertelement instructions // we might create due to scalarization. Cost += getScalarizationOverhead(I, VF); // If we have a predicated load/store, it will need extra i1 extracts and // conditional branches, but may not be executed for each vector lane. Scale // the cost by the probability of executing the predicated block. if (isPredicatedInst(I)) { Cost /= getPredBlockCostDivisor(CostKind); // Add the cost of an i1 extract and a branch auto *VecI1Ty = VectorType::get(IntegerType::getInt1Ty(ValTy->getContext()), VF); Cost += TTI.getScalarizationOverhead( VecI1Ty, APInt::getAllOnes(VF.getFixedValue()), /*Insert=*/false, /*Extract=*/true, CostKind); Cost += TTI.getCFInstrCost(Instruction::Br, CostKind); if (useEmulatedMaskMemRefHack(I, VF)) // Artificially setting to a high enough value to practically disable // vectorization with such operations. Cost = 3000000; } return Cost; } InstructionCost LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I, ElementCount VF) { Type *ValTy = getLoadStoreType(I); auto *VectorTy = cast(toVectorTy(ValTy, VF)); Value *Ptr = getLoadStorePointerOperand(I); unsigned AS = getLoadStoreAddressSpace(I); int ConsecutiveStride = Legal->isConsecutivePtr(ValTy, Ptr); assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Stride should be 1 or -1 for consecutive memory access"); const Align Alignment = getLoadStoreAlignment(I); InstructionCost Cost = 0; if (Legal->isMaskRequired(I)) { Cost += TTI.getMaskedMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, CostKind); } else { TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0)); Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS, CostKind, OpInfo, I); } bool Reverse = ConsecutiveStride < 0; if (Reverse) Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, VectorTy, {}, CostKind, 0); return Cost; } InstructionCost LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I, ElementCount VF) { assert(Legal->isUniformMemOp(*I, VF)); Type *ValTy = getLoadStoreType(I); Type *PtrTy = getLoadStorePointerOperand(I)->getType(); auto *VectorTy = cast(toVectorTy(ValTy, VF)); const Align Alignment = getLoadStoreAlignment(I); unsigned AS = getLoadStoreAddressSpace(I); if (isa(I)) { return TTI.getAddressComputationCost(PtrTy, nullptr, nullptr, CostKind) + TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS, CostKind) + TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy, VectorTy, {}, CostKind); } StoreInst *SI = cast(I); bool IsLoopInvariantStoreValue = Legal->isInvariant(SI->getValueOperand()); // TODO: We have existing tests that request the cost of extracting element // VF.getKnownMinValue() - 1 from a scalable vector. This does not represent // the actual generated code, which involves extracting the last element of // a scalable vector where the lane to extract is unknown at compile time. return TTI.getAddressComputationCost(PtrTy, nullptr, nullptr, CostKind) + TTI.getMemoryOpCost(Instruction::Store, ValTy, Alignment, AS, CostKind) + (IsLoopInvariantStoreValue ? 0 : TTI.getVectorInstrCost(Instruction::ExtractElement, VectorTy, CostKind, VF.getKnownMinValue() - 1)); } InstructionCost LoopVectorizationCostModel::getGatherScatterCost(Instruction *I, ElementCount VF) { Type *ValTy = getLoadStoreType(I); auto *VectorTy = cast(toVectorTy(ValTy, VF)); const Align Alignment = getLoadStoreAlignment(I); const Value *Ptr = getLoadStorePointerOperand(I); Type *PtrTy = toVectorTy(Ptr->getType(), VF); return TTI.getAddressComputationCost(PtrTy, nullptr, nullptr, CostKind) + TTI.getGatherScatterOpCost(I->getOpcode(), VectorTy, Ptr, Legal->isMaskRequired(I), Alignment, CostKind, I); } InstructionCost LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I, ElementCount VF) { const auto *Group = getInterleavedAccessGroup(I); assert(Group && "Fail to get an interleaved access group."); Instruction *InsertPos = Group->getInsertPos(); Type *ValTy = getLoadStoreType(InsertPos); auto *VectorTy = cast(toVectorTy(ValTy, VF)); unsigned AS = getLoadStoreAddressSpace(InsertPos); unsigned InterleaveFactor = Group->getFactor(); auto *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor); // Holds the indices of existing members in the interleaved group. SmallVector Indices; for (unsigned IF = 0; IF < InterleaveFactor; IF++) if (Group->getMember(IF)) Indices.push_back(IF); // Calculate the cost of the whole interleaved group. bool UseMaskForGaps = (Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed()) || (isa(I) && !Group->isFull()); InstructionCost Cost = TTI.getInterleavedMemoryOpCost( InsertPos->getOpcode(), WideVecTy, Group->getFactor(), Indices, Group->getAlign(), AS, CostKind, Legal->isMaskRequired(I), UseMaskForGaps); if (Group->isReverse()) { // TODO: Add support for reversed masked interleaved access. assert(!Legal->isMaskRequired(I) && "Reverse masked interleaved access not supported."); Cost += Group->getNumMembers() * TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy, VectorTy, {}, CostKind, 0); } return Cost; } std::optional LoopVectorizationCostModel::getReductionPatternCost(Instruction *I, ElementCount VF, Type *Ty) const { using namespace llvm::PatternMatch; // Early exit for no inloop reductions if (InLoopReductions.empty() || VF.isScalar() || !isa(Ty)) return std::nullopt; auto *VectorTy = cast(Ty); // We are looking for a pattern of, and finding the minimal acceptable cost: // reduce(mul(ext(A), ext(B))) or // reduce(mul(A, B)) or // reduce(ext(A)) or // reduce(A). // The basic idea is that we walk down the tree to do that, finding the root // reduction instruction in InLoopReductionImmediateChains. From there we find // the pattern of mul/ext and test the cost of the entire pattern vs the cost // of the components. If the reduction cost is lower then we return it for the // reduction instruction and 0 for the other instructions in the pattern. If // it is not we return an invalid cost specifying the orignal cost method // should be used. Instruction *RetI = I; if (match(RetI, m_ZExtOrSExt(m_Value()))) { if (!RetI->hasOneUser()) return std::nullopt; RetI = RetI->user_back(); } if (match(RetI, m_OneUse(m_Mul(m_Value(), m_Value()))) && RetI->user_back()->getOpcode() == Instruction::Add) { RetI = RetI->user_back(); } // Test if the found instruction is a reduction, and if not return an invalid // cost specifying the parent to use the original cost modelling. Instruction *LastChain = InLoopReductionImmediateChains.lookup(RetI); if (!LastChain) return std::nullopt; // Find the reduction this chain is a part of and calculate the basic cost of // the reduction on its own. Instruction *ReductionPhi = LastChain; while (!isa(ReductionPhi)) ReductionPhi = InLoopReductionImmediateChains.at(ReductionPhi); const RecurrenceDescriptor &RdxDesc = Legal->getRecurrenceDescriptor(cast(ReductionPhi)); InstructionCost BaseCost; RecurKind RK = RdxDesc.getRecurrenceKind(); if (RecurrenceDescriptor::isMinMaxRecurrenceKind(RK)) { Intrinsic::ID MinMaxID = getMinMaxReductionIntrinsicOp(RK); BaseCost = TTI.getMinMaxReductionCost(MinMaxID, VectorTy, RdxDesc.getFastMathFlags(), CostKind); } else { BaseCost = TTI.getArithmeticReductionCost( RdxDesc.getOpcode(), VectorTy, RdxDesc.getFastMathFlags(), CostKind); } // For a call to the llvm.fmuladd intrinsic we need to add the cost of a // normal fmul instruction to the cost of the fadd reduction. if (RK == RecurKind::FMulAdd) BaseCost += TTI.getArithmeticInstrCost(Instruction::FMul, VectorTy, CostKind); // If we're using ordered reductions then we can just return the base cost // here, since getArithmeticReductionCost calculates the full ordered // reduction cost when FP reassociation is not allowed. if (useOrderedReductions(RdxDesc)) return BaseCost; // Get the operand that was not the reduction chain and match it to one of the // patterns, returning the better cost if it is found. Instruction *RedOp = RetI->getOperand(1) == LastChain ? dyn_cast(RetI->getOperand(0)) : dyn_cast(RetI->getOperand(1)); VectorTy = VectorType::get(I->getOperand(0)->getType(), VectorTy); Instruction *Op0, *Op1; if (RedOp && RdxDesc.getOpcode() == Instruction::Add && match(RedOp, m_ZExtOrSExt(m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) && match(Op0, m_ZExtOrSExt(m_Value())) && Op0->getOpcode() == Op1->getOpcode() && Op0->getOperand(0)->getType() == Op1->getOperand(0)->getType() && !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1) && (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) { // Matched reduce.add(ext(mul(ext(A), ext(B))) // Note that the extend opcodes need to all match, or if A==B they will have // been converted to zext(mul(sext(A), sext(A))) as it is known positive, // which is equally fine. bool IsUnsigned = isa(Op0); auto *ExtType = VectorType::get(Op0->getOperand(0)->getType(), VectorTy); auto *MulType = VectorType::get(Op0->getType(), VectorTy); InstructionCost ExtCost = TTI.getCastInstrCost(Op0->getOpcode(), MulType, ExtType, TTI::CastContextHint::None, CostKind, Op0); InstructionCost MulCost = TTI.getArithmeticInstrCost(Instruction::Mul, MulType, CostKind); InstructionCost Ext2Cost = TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, MulType, TTI::CastContextHint::None, CostKind, RedOp); InstructionCost RedCost = TTI.getMulAccReductionCost( IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, CostKind); if (RedCost.isValid() && RedCost < ExtCost * 2 + MulCost + Ext2Cost + BaseCost) return I == RetI ? RedCost : 0; } else if (RedOp && match(RedOp, m_ZExtOrSExt(m_Value())) && !TheLoop->isLoopInvariant(RedOp)) { // Matched reduce(ext(A)) bool IsUnsigned = isa(RedOp); auto *ExtType = VectorType::get(RedOp->getOperand(0)->getType(), VectorTy); InstructionCost RedCost = TTI.getExtendedReductionCost( RdxDesc.getOpcode(), IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, RdxDesc.getFastMathFlags(), CostKind); InstructionCost ExtCost = TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, ExtType, TTI::CastContextHint::None, CostKind, RedOp); if (RedCost.isValid() && RedCost < BaseCost + ExtCost) return I == RetI ? RedCost : 0; } else if (RedOp && RdxDesc.getOpcode() == Instruction::Add && match(RedOp, m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) { if (match(Op0, m_ZExtOrSExt(m_Value())) && Op0->getOpcode() == Op1->getOpcode() && !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1)) { bool IsUnsigned = isa(Op0); Type *Op0Ty = Op0->getOperand(0)->getType(); Type *Op1Ty = Op1->getOperand(0)->getType(); Type *LargestOpTy = Op0Ty->getIntegerBitWidth() < Op1Ty->getIntegerBitWidth() ? Op1Ty : Op0Ty; auto *ExtType = VectorType::get(LargestOpTy, VectorTy); // Matched reduce.add(mul(ext(A), ext(B))), where the two ext may be of // different sizes. We take the largest type as the ext to reduce, and add // the remaining cost as, for example reduce(mul(ext(ext(A)), ext(B))). InstructionCost ExtCost0 = TTI.getCastInstrCost( Op0->getOpcode(), VectorTy, VectorType::get(Op0Ty, VectorTy), TTI::CastContextHint::None, CostKind, Op0); InstructionCost ExtCost1 = TTI.getCastInstrCost( Op1->getOpcode(), VectorTy, VectorType::get(Op1Ty, VectorTy), TTI::CastContextHint::None, CostKind, Op1); InstructionCost MulCost = TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); InstructionCost RedCost = TTI.getMulAccReductionCost( IsUnsigned, RdxDesc.getRecurrenceType(), ExtType, CostKind); InstructionCost ExtraExtCost = 0; if (Op0Ty != LargestOpTy || Op1Ty != LargestOpTy) { Instruction *ExtraExtOp = (Op0Ty != LargestOpTy) ? Op0 : Op1; ExtraExtCost = TTI.getCastInstrCost( ExtraExtOp->getOpcode(), ExtType, VectorType::get(ExtraExtOp->getOperand(0)->getType(), VectorTy), TTI::CastContextHint::None, CostKind, ExtraExtOp); } if (RedCost.isValid() && (RedCost + ExtraExtCost) < (ExtCost0 + ExtCost1 + MulCost + BaseCost)) return I == RetI ? RedCost : 0; } else if (!match(I, m_ZExtOrSExt(m_Value()))) { // Matched reduce.add(mul()) InstructionCost MulCost = TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); InstructionCost RedCost = TTI.getMulAccReductionCost( true, RdxDesc.getRecurrenceType(), VectorTy, CostKind); if (RedCost.isValid() && RedCost < MulCost + BaseCost) return I == RetI ? RedCost : 0; } } return I == RetI ? std::optional(BaseCost) : std::nullopt; } InstructionCost LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I, ElementCount VF) { // Calculate scalar cost only. Vectorization cost should be ready at this // moment. if (VF.isScalar()) { Type *ValTy = getLoadStoreType(I); Type *PtrTy = getLoadStorePointerOperand(I)->getType(); const Align Alignment = getLoadStoreAlignment(I); unsigned AS = getLoadStoreAddressSpace(I); TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0)); return TTI.getAddressComputationCost(PtrTy, nullptr, nullptr, CostKind) + TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, CostKind, OpInfo, I); } return getWideningCost(I, VF); } InstructionCost LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I, ElementCount VF) const { // There is no mechanism yet to create a scalable scalarization loop, // so this is currently Invalid. if (VF.isScalable()) return InstructionCost::getInvalid(); if (VF.isScalar()) return 0; InstructionCost Cost = 0; Type *RetTy = toVectorizedTy(I->getType(), VF); if (!RetTy->isVoidTy() && (!isa(I) || !TTI.supportsEfficientVectorElementLoadStore())) { for (Type *VectorTy : getContainedTypes(RetTy)) { Cost += TTI.getScalarizationOverhead( cast(VectorTy), APInt::getAllOnes(VF.getFixedValue()), /*Insert=*/true, /*Extract=*/false, CostKind); } } // Some targets keep addresses scalar. if (isa(I) && !TTI.prefersVectorizedAddressing()) return Cost; // Some targets support efficient element stores. if (isa(I) && TTI.supportsEfficientVectorElementLoadStore()) return Cost; // Collect operands to consider. CallInst *CI = dyn_cast(I); Instruction::op_range Ops = CI ? CI->args() : I->operands(); // Skip operands that do not require extraction/scalarization and do not incur // any overhead. SmallVector Tys; for (auto *V : filterExtractingOperands(Ops, VF)) Tys.push_back(maybeVectorizeType(V->getType(), VF)); return Cost + TTI.getOperandsScalarizationOverhead( filterExtractingOperands(Ops, VF), Tys, CostKind); } void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) { if (VF.isScalar()) return; NumPredStores = 0; for (BasicBlock *BB : TheLoop->blocks()) { // For each instruction in the old loop. for (Instruction &I : *BB) { Value *Ptr = getLoadStorePointerOperand(&I); if (!Ptr) continue; // TODO: We should generate better code and update the cost model for // predicated uniform stores. Today they are treated as any other // predicated store (see added test cases in // invariant-store-vectorization.ll). if (isa(&I) && isScalarWithPredication(&I, VF)) NumPredStores++; if (Legal->isUniformMemOp(I, VF)) { auto IsLegalToScalarize = [&]() { if (!VF.isScalable()) // Scalarization of fixed length vectors "just works". return true; // We have dedicated lowering for unpredicated uniform loads and // stores. Note that even with tail folding we know that at least // one lane is active (i.e. generalized predication is not possible // here), and the logic below depends on this fact. if (!foldTailByMasking()) return true; // For scalable vectors, a uniform memop load is always // uniform-by-parts and we know how to scalarize that. if (isa(I)) return true; // A uniform store isn't neccessarily uniform-by-part // and we can't assume scalarization. auto &SI = cast(I); return TheLoop->isLoopInvariant(SI.getValueOperand()); }; const InstructionCost GatherScatterCost = isLegalGatherOrScatter(&I, VF) ? getGatherScatterCost(&I, VF) : InstructionCost::getInvalid(); // Load: Scalar load + broadcast // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract // FIXME: This cost is a significant under-estimate for tail folded // memory ops. const InstructionCost ScalarizationCost = IsLegalToScalarize() ? getUniformMemOpCost(&I, VF) : InstructionCost::getInvalid(); // Choose better solution for the current VF, Note that Invalid // costs compare as maximumal large. If both are invalid, we get // scalable invalid which signals a failure and a vectorization abort. if (GatherScatterCost < ScalarizationCost) setWideningDecision(&I, VF, CM_GatherScatter, GatherScatterCost); else setWideningDecision(&I, VF, CM_Scalarize, ScalarizationCost); continue; } // We assume that widening is the best solution when possible. if (memoryInstructionCanBeWidened(&I, VF)) { InstructionCost Cost = getConsecutiveMemOpCost(&I, VF); int ConsecutiveStride = Legal->isConsecutivePtr( getLoadStoreType(&I), getLoadStorePointerOperand(&I)); assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) && "Expected consecutive stride."); InstWidening Decision = ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse; setWideningDecision(&I, VF, Decision, Cost); continue; } // Choose between Interleaving, Gather/Scatter or Scalarization. InstructionCost InterleaveCost = InstructionCost::getInvalid(); unsigned NumAccesses = 1; if (isAccessInterleaved(&I)) { const auto *Group = getInterleavedAccessGroup(&I); assert(Group && "Fail to get an interleaved access group."); // Make one decision for the whole group. if (getWideningDecision(&I, VF) != CM_Unknown) continue; NumAccesses = Group->getNumMembers(); if (interleavedAccessCanBeWidened(&I, VF)) InterleaveCost = getInterleaveGroupCost(&I, VF); } InstructionCost GatherScatterCost = isLegalGatherOrScatter(&I, VF) ? getGatherScatterCost(&I, VF) * NumAccesses : InstructionCost::getInvalid(); InstructionCost ScalarizationCost = getMemInstScalarizationCost(&I, VF) * NumAccesses; // Choose better solution for the current VF, // write down this decision and use it during vectorization. InstructionCost Cost; InstWidening Decision; if (InterleaveCost <= GatherScatterCost && InterleaveCost < ScalarizationCost) { Decision = CM_Interleave; Cost = InterleaveCost; } else if (GatherScatterCost < ScalarizationCost) { Decision = CM_GatherScatter; Cost = GatherScatterCost; } else { Decision = CM_Scalarize; Cost = ScalarizationCost; } // If the instructions belongs to an interleave group, the whole group // receives the same decision. The whole group receives the cost, but // the cost will actually be assigned to one instruction. if (const auto *Group = getInterleavedAccessGroup(&I)) setWideningDecision(Group, VF, Decision, Cost); else setWideningDecision(&I, VF, Decision, Cost); } } // Make sure that any load of address and any other address computation // remains scalar unless there is gather/scatter support. This avoids // inevitable extracts into address registers, and also has the benefit of // activating LSR more, since that pass can't optimize vectorized // addresses. if (TTI.prefersVectorizedAddressing()) return; // Start with all scalar pointer uses. SmallPtrSet AddrDefs; for (BasicBlock *BB : TheLoop->blocks()) for (Instruction &I : *BB) { Instruction *PtrDef = dyn_cast_or_null(getLoadStorePointerOperand(&I)); if (PtrDef && TheLoop->contains(PtrDef) && getWideningDecision(&I, VF) != CM_GatherScatter) AddrDefs.insert(PtrDef); } // Add all instructions used to generate the addresses. SmallVector Worklist; append_range(Worklist, AddrDefs); while (!Worklist.empty()) { Instruction *I = Worklist.pop_back_val(); for (auto &Op : I->operands()) if (auto *InstOp = dyn_cast(Op)) if ((InstOp->getParent() == I->getParent()) && !isa(InstOp) && AddrDefs.insert(InstOp).second) Worklist.push_back(InstOp); } for (auto *I : AddrDefs) { if (isa(I)) { // Setting the desired widening decision should ideally be handled in // by cost functions, but since this involves the task of finding out // if the loaded register is involved in an address computation, it is // instead changed here when we know this is the case. InstWidening Decision = getWideningDecision(I, VF); if (Decision == CM_Widen || Decision == CM_Widen_Reverse) // Scalarize a widened load of address. setWideningDecision( I, VF, CM_Scalarize, (VF.getKnownMinValue() * getMemoryInstructionCost(I, ElementCount::getFixed(1)))); else if (const auto *Group = getInterleavedAccessGroup(I)) { // Scalarize an interleave group of address loads. for (unsigned I = 0; I < Group->getFactor(); ++I) { if (Instruction *Member = Group->getMember(I)) setWideningDecision( Member, VF, CM_Scalarize, (VF.getKnownMinValue() * getMemoryInstructionCost(Member, ElementCount::getFixed(1)))); } } } else { // Cannot scalarize fixed-order recurrence phis at the moment. if (isa(I) && Legal->isFixedOrderRecurrence(cast(I))) continue; // Make sure I gets scalarized and a cost estimate without // scalarization overhead. ForcedScalars[VF].insert(I); } } } void LoopVectorizationCostModel::setVectorizedCallDecision(ElementCount VF) { assert(!VF.isScalar() && "Trying to set a vectorization decision for a scalar VF"); auto ForcedScalar = ForcedScalars.find(VF); for (BasicBlock *BB : TheLoop->blocks()) { // For each instruction in the old loop. for (Instruction &I : *BB) { CallInst *CI = dyn_cast(&I); if (!CI) continue; InstructionCost ScalarCost = InstructionCost::getInvalid(); InstructionCost VectorCost = InstructionCost::getInvalid(); InstructionCost IntrinsicCost = InstructionCost::getInvalid(); Function *ScalarFunc = CI->getCalledFunction(); Type *ScalarRetTy = CI->getType(); SmallVector Tys, ScalarTys; for (auto &ArgOp : CI->args()) ScalarTys.push_back(ArgOp->getType()); // Estimate cost of scalarized vector call. The source operands are // assumed to be vectors, so we need to extract individual elements from // there, execute VF scalar calls, and then gather the result into the // vector return value. InstructionCost ScalarCallCost = TTI.getCallInstrCost(ScalarFunc, ScalarRetTy, ScalarTys, CostKind); // Compute costs of unpacking argument values for the scalar calls and // packing the return values to a vector. InstructionCost ScalarizationCost = getScalarizationOverhead(CI, VF); ScalarCost = ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost; // Honor ForcedScalars and UniformAfterVectorization decisions. // TODO: For calls, it might still be more profitable to widen. Use // VPlan-based cost model to compare different options. if (VF.isVector() && ((ForcedScalar != ForcedScalars.end() && ForcedScalar->second.contains(CI)) || isUniformAfterVectorization(CI, VF))) { setCallWideningDecision(CI, VF, CM_Scalarize, nullptr, Intrinsic::not_intrinsic, std::nullopt, ScalarCost); continue; } bool MaskRequired = Legal->isMaskRequired(CI); // Compute corresponding vector type for return value and arguments. Type *RetTy = toVectorizedTy(ScalarRetTy, VF); for (Type *ScalarTy : ScalarTys) Tys.push_back(toVectorizedTy(ScalarTy, VF)); // An in-loop reduction using an fmuladd intrinsic is a special case; // we don't want the normal cost for that intrinsic. if (RecurrenceDescriptor::isFMulAddIntrinsic(CI)) if (auto RedCost = getReductionPatternCost(CI, VF, RetTy)) { setCallWideningDecision(CI, VF, CM_IntrinsicCall, nullptr, getVectorIntrinsicIDForCall(CI, TLI), std::nullopt, *RedCost); continue; } // Find the cost of vectorizing the call, if we can find a suitable // vector variant of the function. VFInfo FuncInfo; Function *VecFunc = nullptr; // Search through any available variants for one we can use at this VF. for (VFInfo &Info : VFDatabase::getMappings(*CI)) { // Must match requested VF. if (Info.Shape.VF != VF) continue; // Must take a mask argument if one is required if (MaskRequired && !Info.isMasked()) continue; // Check that all parameter kinds are supported bool ParamsOk = true; for (VFParameter Param : Info.Shape.Parameters) { switch (Param.ParamKind) { case VFParamKind::Vector: break; case VFParamKind::OMP_Uniform: { Value *ScalarParam = CI->getArgOperand(Param.ParamPos); // Make sure the scalar parameter in the loop is invariant. if (!PSE.getSE()->isLoopInvariant(PSE.getSCEV(ScalarParam), TheLoop)) ParamsOk = false; break; } case VFParamKind::OMP_Linear: { Value *ScalarParam = CI->getArgOperand(Param.ParamPos); // Find the stride for the scalar parameter in this loop and see if // it matches the stride for the variant. // TODO: do we need to figure out the cost of an extract to get the // first lane? Or do we hope that it will be folded away? ScalarEvolution *SE = PSE.getSE(); const auto *SAR = dyn_cast(SE->getSCEV(ScalarParam)); if (!SAR || SAR->getLoop() != TheLoop) { ParamsOk = false; break; } const SCEVConstant *Step = dyn_cast(SAR->getStepRecurrence(*SE)); if (!Step || Step->getAPInt().getSExtValue() != Param.LinearStepOrPos) ParamsOk = false; break; } case VFParamKind::GlobalPredicate: break; default: ParamsOk = false; break; } } if (!ParamsOk) continue; // Found a suitable candidate, stop here. VecFunc = CI->getModule()->getFunction(Info.VectorName); FuncInfo = Info; break; } if (TLI && VecFunc && !CI->isNoBuiltin()) VectorCost = TTI.getCallInstrCost(nullptr, RetTy, Tys, CostKind); // Find the cost of an intrinsic; some targets may have instructions that // perform the operation without needing an actual call. Intrinsic::ID IID = getVectorIntrinsicIDForCall(CI, TLI); if (IID != Intrinsic::not_intrinsic) IntrinsicCost = getVectorIntrinsicCost(CI, VF); InstructionCost Cost = ScalarCost; InstWidening Decision = CM_Scalarize; if (VectorCost <= Cost) { Cost = VectorCost; Decision = CM_VectorCall; } if (IntrinsicCost <= Cost) { Cost = IntrinsicCost; Decision = CM_IntrinsicCall; } setCallWideningDecision(CI, VF, Decision, VecFunc, IID, FuncInfo.getParamIndexForOptionalMask(), Cost); } } } bool LoopVectorizationCostModel::shouldConsiderInvariant(Value *Op) { if (!Legal->isInvariant(Op)) return false; // Consider Op invariant, if it or its operands aren't predicated // instruction in the loop. In that case, it is not trivially hoistable. auto *OpI = dyn_cast(Op); return !OpI || !TheLoop->contains(OpI) || (!isPredicatedInst(OpI) && (!isa(OpI) || OpI->getParent() != TheLoop->getHeader()) && all_of(OpI->operands(), [this](Value *Op) { return shouldConsiderInvariant(Op); })); } InstructionCost LoopVectorizationCostModel::getInstructionCost(Instruction *I, ElementCount VF) { // If we know that this instruction will remain uniform, check the cost of // the scalar version. if (isUniformAfterVectorization(I, VF)) VF = ElementCount::getFixed(1); if (VF.isVector() && isProfitableToScalarize(I, VF)) return InstsToScalarize[VF][I]; // Forced scalars do not have any scalarization overhead. auto ForcedScalar = ForcedScalars.find(VF); if (VF.isVector() && ForcedScalar != ForcedScalars.end()) { auto InstSet = ForcedScalar->second; if (InstSet.count(I)) return getInstructionCost(I, ElementCount::getFixed(1)) * VF.getKnownMinValue(); } Type *RetTy = I->getType(); if (canTruncateToMinimalBitwidth(I, VF)) RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]); auto *SE = PSE.getSE(); Type *VectorTy; if (isScalarAfterVectorization(I, VF)) { [[maybe_unused]] auto HasSingleCopyAfterVectorization = [this](Instruction *I, ElementCount VF) -> bool { if (VF.isScalar()) return true; auto Scalarized = InstsToScalarize.find(VF); assert(Scalarized != InstsToScalarize.end() && "VF not yet analyzed for scalarization profitability"); return !Scalarized->second.count(I) && llvm::all_of(I->users(), [&](User *U) { auto *UI = cast(U); return !Scalarized->second.count(UI); }); }; // With the exception of GEPs and PHIs, after scalarization there should // only be one copy of the instruction generated in the loop. This is // because the VF is either 1, or any instructions that need scalarizing // have already been dealt with by the time we get here. As a result, // it means we don't have to multiply the instruction cost by VF. assert(I->getOpcode() == Instruction::GetElementPtr || I->getOpcode() == Instruction::PHI || (I->getOpcode() == Instruction::BitCast && I->getType()->isPointerTy()) || HasSingleCopyAfterVectorization(I, VF)); VectorTy = RetTy; } else VectorTy = toVectorizedTy(RetTy, VF); if (VF.isVector() && VectorTy->isVectorTy() && !TTI.getNumberOfParts(VectorTy)) return InstructionCost::getInvalid(); // TODO: We need to estimate the cost of intrinsic calls. switch (I->getOpcode()) { case Instruction::GetElementPtr: // We mark this instruction as zero-cost because the cost of GEPs in // vectorized code depends on whether the corresponding memory instruction // is scalarized or not. Therefore, we handle GEPs with the memory // instruction cost. return 0; case Instruction::Br: { // In cases of scalarized and predicated instructions, there will be VF // predicated blocks in the vectorized loop. Each branch around these // blocks requires also an extract of its vector compare i1 element. // Note that the conditional branch from the loop latch will be replaced by // a single branch controlling the loop, so there is no extra overhead from // scalarization. bool ScalarPredicatedBB = false; BranchInst *BI = cast(I); if (VF.isVector() && BI->isConditional() && (PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(0)) || PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(1))) && BI->getParent() != TheLoop->getLoopLatch()) ScalarPredicatedBB = true; if (ScalarPredicatedBB) { // Not possible to scalarize scalable vector with predicated instructions. if (VF.isScalable()) return InstructionCost::getInvalid(); // Return cost for branches around scalarized and predicated blocks. auto *VecI1Ty = VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF); return ( TTI.getScalarizationOverhead( VecI1Ty, APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ false, /*Extract*/ true, CostKind) + (TTI.getCFInstrCost(Instruction::Br, CostKind) * VF.getFixedValue())); } if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar()) // The back-edge branch will remain, as will all scalar branches. return TTI.getCFInstrCost(Instruction::Br, CostKind); // This branch will be eliminated by if-conversion. return 0; // Note: We currently assume zero cost for an unconditional branch inside // a predicated block since it will become a fall-through, although we // may decide in the future to call TTI for all branches. } case Instruction::Switch: { if (VF.isScalar()) return TTI.getCFInstrCost(Instruction::Switch, CostKind); auto *Switch = cast(I); return Switch->getNumCases() * TTI.getCmpSelInstrCost( Instruction::ICmp, toVectorTy(Switch->getCondition()->getType(), VF), toVectorTy(Type::getInt1Ty(I->getContext()), VF), CmpInst::ICMP_EQ, CostKind); } case Instruction::PHI: { auto *Phi = cast(I); // First-order recurrences are replaced by vector shuffles inside the loop. if (VF.isVector() && Legal->isFixedOrderRecurrence(Phi)) { SmallVector Mask(VF.getKnownMinValue()); std::iota(Mask.begin(), Mask.end(), VF.getKnownMinValue() - 1); return TTI.getShuffleCost(TargetTransformInfo::SK_Splice, cast(VectorTy), cast(VectorTy), Mask, CostKind, VF.getKnownMinValue() - 1); } // Phi nodes in non-header blocks (not inductions, reductions, etc.) are // converted into select instructions. We require N - 1 selects per phi // node, where N is the number of incoming values. if (VF.isVector() && Phi->getParent() != TheLoop->getHeader()) { Type *ResultTy = Phi->getType(); // All instructions in an Any-of reduction chain are narrowed to bool. // Check if that is the case for this phi node. auto *HeaderUser = cast_if_present( find_singleton(Phi->users(), [this](User *U, bool) -> User * { auto *Phi = dyn_cast(U); if (Phi && Phi->getParent() == TheLoop->getHeader()) return Phi; return nullptr; })); if (HeaderUser) { auto &ReductionVars = Legal->getReductionVars(); auto Iter = ReductionVars.find(HeaderUser); if (Iter != ReductionVars.end() && RecurrenceDescriptor::isAnyOfRecurrenceKind( Iter->second.getRecurrenceKind())) ResultTy = Type::getInt1Ty(Phi->getContext()); } return (Phi->getNumIncomingValues() - 1) * TTI.getCmpSelInstrCost( Instruction::Select, toVectorTy(ResultTy, VF), toVectorTy(Type::getInt1Ty(Phi->getContext()), VF), CmpInst::BAD_ICMP_PREDICATE, CostKind); } // When tail folding with EVL, if the phi is part of an out of loop // reduction then it will be transformed into a wide vp_merge. if (VF.isVector() && foldTailWithEVL() && Legal->getReductionVars().contains(Phi) && !isInLoopReduction(Phi)) { IntrinsicCostAttributes ICA( Intrinsic::vp_merge, toVectorTy(Phi->getType(), VF), {toVectorTy(Type::getInt1Ty(Phi->getContext()), VF)}); return TTI.getIntrinsicInstrCost(ICA, CostKind); } return TTI.getCFInstrCost(Instruction::PHI, CostKind); } case Instruction::UDiv: case Instruction::SDiv: case Instruction::URem: case Instruction::SRem: if (VF.isVector() && isPredicatedInst(I)) { const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF); return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost) ? ScalarCost : SafeDivisorCost; } // We've proven all lanes safe to speculate, fall through. [[fallthrough]]; case Instruction::Add: case Instruction::Sub: { auto Info = Legal->getHistogramInfo(I); if (Info && VF.isVector()) { const HistogramInfo *HGram = Info.value(); // Assume that a non-constant update value (or a constant != 1) requires // a multiply, and add that into the cost. InstructionCost MulCost = TTI::TCC_Free; ConstantInt *RHS = dyn_cast(I->getOperand(1)); if (!RHS || RHS->getZExtValue() != 1) MulCost = TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); // Find the cost of the histogram operation itself. Type *PtrTy = VectorType::get(HGram->Load->getPointerOperandType(), VF); Type *ScalarTy = I->getType(); Type *MaskTy = VectorType::get(Type::getInt1Ty(I->getContext()), VF); IntrinsicCostAttributes ICA(Intrinsic::experimental_vector_histogram_add, Type::getVoidTy(I->getContext()), {PtrTy, ScalarTy, MaskTy}); // Add the costs together with the add/sub operation. return TTI.getIntrinsicInstrCost(ICA, CostKind) + MulCost + TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, CostKind); } [[fallthrough]]; } case Instruction::FAdd: case Instruction::FSub: case Instruction::Mul: case Instruction::FMul: case Instruction::FDiv: case Instruction::FRem: case Instruction::Shl: case Instruction::LShr: case Instruction::AShr: case Instruction::And: case Instruction::Or: case Instruction::Xor: { // If we're speculating on the stride being 1, the multiplication may // fold away. We can generalize this for all operations using the notion // of neutral elements. (TODO) if (I->getOpcode() == Instruction::Mul && ((TheLoop->isLoopInvariant(I->getOperand(0)) && PSE.getSCEV(I->getOperand(0))->isOne()) || (TheLoop->isLoopInvariant(I->getOperand(1)) && PSE.getSCEV(I->getOperand(1))->isOne()))) return 0; // Detect reduction patterns if (auto RedCost = getReductionPatternCost(I, VF, VectorTy)) return *RedCost; // Certain instructions can be cheaper to vectorize if they have a constant // second vector operand. One example of this are shifts on x86. Value *Op2 = I->getOperand(1); if (!isa(Op2) && TheLoop->isLoopInvariant(Op2) && PSE.getSE()->isSCEVable(Op2->getType()) && isa(PSE.getSCEV(Op2))) { Op2 = cast(PSE.getSCEV(Op2))->getValue(); } auto Op2Info = TTI.getOperandInfo(Op2); if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue && shouldConsiderInvariant(Op2)) Op2Info.Kind = TargetTransformInfo::OK_UniformValue; SmallVector Operands(I->operand_values()); return TTI.getArithmeticInstrCost( I->getOpcode(), VectorTy, CostKind, {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, Op2Info, Operands, I, TLI); } case Instruction::FNeg: { return TTI.getArithmeticInstrCost( I->getOpcode(), VectorTy, CostKind, {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None}, I->getOperand(0), I); } case Instruction::Select: { SelectInst *SI = cast(I); const SCEV *CondSCEV = SE->getSCEV(SI->getCondition()); bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop)); const Value *Op0, *Op1; using namespace llvm::PatternMatch; if (!ScalarCond && (match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1))) || match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))))) { // select x, y, false --> x & y // select x, true, y --> x | y const auto [Op1VK, Op1VP] = TTI::getOperandInfo(Op0); const auto [Op2VK, Op2VP] = TTI::getOperandInfo(Op1); assert(Op0->getType()->getScalarSizeInBits() == 1 && Op1->getType()->getScalarSizeInBits() == 1); SmallVector Operands{Op0, Op1}; return TTI.getArithmeticInstrCost( match(I, m_LogicalOr()) ? Instruction::Or : Instruction::And, VectorTy, CostKind, {Op1VK, Op1VP}, {Op2VK, Op2VP}, Operands, I); } Type *CondTy = SI->getCondition()->getType(); if (!ScalarCond) CondTy = VectorType::get(CondTy, VF); CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE; if (auto *Cmp = dyn_cast(SI->getCondition())) Pred = Cmp->getPredicate(); return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy, Pred, CostKind, {TTI::OK_AnyValue, TTI::OP_None}, {TTI::OK_AnyValue, TTI::OP_None}, I); } case Instruction::ICmp: case Instruction::FCmp: { Type *ValTy = I->getOperand(0)->getType(); if (canTruncateToMinimalBitwidth(I, VF)) { [[maybe_unused]] Instruction *Op0AsInstruction = dyn_cast(I->getOperand(0)); assert((!canTruncateToMinimalBitwidth(Op0AsInstruction, VF) || MinBWs[I] == MinBWs[Op0AsInstruction]) && "if both the operand and the compare are marked for " "truncation, they must have the same bitwidth"); ValTy = IntegerType::get(ValTy->getContext(), MinBWs[I]); } VectorTy = toVectorTy(ValTy, VF); return TTI.getCmpSelInstrCost( I->getOpcode(), VectorTy, CmpInst::makeCmpResultType(VectorTy), cast(I)->getPredicate(), CostKind, {TTI::OK_AnyValue, TTI::OP_None}, {TTI::OK_AnyValue, TTI::OP_None}, I); } case Instruction::Store: case Instruction::Load: { ElementCount Width = VF; if (Width.isVector()) { InstWidening Decision = getWideningDecision(I, Width); assert(Decision != CM_Unknown && "CM decision should be taken at this point"); if (getWideningCost(I, VF) == InstructionCost::getInvalid()) return InstructionCost::getInvalid(); if (Decision == CM_Scalarize) Width = ElementCount::getFixed(1); } VectorTy = toVectorTy(getLoadStoreType(I), Width); return getMemoryInstructionCost(I, VF); } case Instruction::BitCast: if (I->getType()->isPointerTy()) return 0; [[fallthrough]]; case Instruction::ZExt: case Instruction::SExt: case Instruction::FPToUI: case Instruction::FPToSI: case Instruction::FPExt: case Instruction::PtrToInt: case Instruction::IntToPtr: case Instruction::SIToFP: case Instruction::UIToFP: case Instruction::Trunc: case Instruction::FPTrunc: { // Computes the CastContextHint from a Load/Store instruction. auto ComputeCCH = [&](Instruction *I) -> TTI::CastContextHint { assert((isa(I) || isa(I)) && "Expected a load or a store!"); if (VF.isScalar() || !TheLoop->contains(I)) return TTI::CastContextHint::Normal; switch (getWideningDecision(I, VF)) { case LoopVectorizationCostModel::CM_GatherScatter: return TTI::CastContextHint::GatherScatter; case LoopVectorizationCostModel::CM_Interleave: return TTI::CastContextHint::Interleave; case LoopVectorizationCostModel::CM_Scalarize: case LoopVectorizationCostModel::CM_Widen: return isPredicatedInst(I) ? TTI::CastContextHint::Masked : TTI::CastContextHint::Normal; case LoopVectorizationCostModel::CM_Widen_Reverse: return TTI::CastContextHint::Reversed; case LoopVectorizationCostModel::CM_Unknown: llvm_unreachable("Instr did not go through cost modelling?"); case LoopVectorizationCostModel::CM_VectorCall: case LoopVectorizationCostModel::CM_IntrinsicCall: llvm_unreachable_internal("Instr has invalid widening decision"); } llvm_unreachable("Unhandled case!"); }; unsigned Opcode = I->getOpcode(); TTI::CastContextHint CCH = TTI::CastContextHint::None; // For Trunc, the context is the only user, which must be a StoreInst. if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) { if (I->hasOneUse()) if (StoreInst *Store = dyn_cast(*I->user_begin())) CCH = ComputeCCH(Store); } // For Z/Sext, the context is the operand, which must be a LoadInst. else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt || Opcode == Instruction::FPExt) { if (LoadInst *Load = dyn_cast(I->getOperand(0))) CCH = ComputeCCH(Load); } // We optimize the truncation of induction variables having constant // integer steps. The cost of these truncations is the same as the scalar // operation. if (isOptimizableIVTruncate(I, VF)) { auto *Trunc = cast(I); return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(), Trunc->getSrcTy(), CCH, CostKind, Trunc); } // Detect reduction patterns if (auto RedCost = getReductionPatternCost(I, VF, VectorTy)) return *RedCost; Type *SrcScalarTy = I->getOperand(0)->getType(); Instruction *Op0AsInstruction = dyn_cast(I->getOperand(0)); if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF)) SrcScalarTy = IntegerType::get(SrcScalarTy->getContext(), MinBWs[Op0AsInstruction]); Type *SrcVecTy = VectorTy->isVectorTy() ? toVectorTy(SrcScalarTy, VF) : SrcScalarTy; if (canTruncateToMinimalBitwidth(I, VF)) { // If the result type is <= the source type, there will be no extend // after truncating the users to the minimal required bitwidth. if (VectorTy->getScalarSizeInBits() <= SrcVecTy->getScalarSizeInBits() && (I->getOpcode() == Instruction::ZExt || I->getOpcode() == Instruction::SExt)) return 0; } return TTI.getCastInstrCost(Opcode, VectorTy, SrcVecTy, CCH, CostKind, I); } case Instruction::Call: return getVectorCallCost(cast(I), VF); case Instruction::ExtractValue: return TTI.getInstructionCost(I, CostKind); case Instruction::Alloca: // We cannot easily widen alloca to a scalable alloca, as // the result would need to be a vector of pointers. if (VF.isScalable()) return InstructionCost::getInvalid(); [[fallthrough]]; default: // This opcode is unknown. Assume that it is the same as 'mul'. return TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind); } // end of switch. } void LoopVectorizationCostModel::collectValuesToIgnore() { // Ignore ephemeral values. CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore); SmallVector DeadInterleavePointerOps; SmallVector DeadOps; // If a scalar epilogue is required, users outside the loop won't use // live-outs from the vector loop but from the scalar epilogue. Ignore them if // that is the case. bool RequiresScalarEpilogue = requiresScalarEpilogue(true); auto IsLiveOutDead = [this, RequiresScalarEpilogue](User *U) { return RequiresScalarEpilogue && !TheLoop->contains(cast(U)->getParent()); }; LoopBlocksDFS DFS(TheLoop); DFS.perform(LI); MapVector> DeadInvariantStoreOps; for (BasicBlock *BB : reverse(make_range(DFS.beginRPO(), DFS.endRPO()))) for (Instruction &I : reverse(*BB)) { // Find all stores to invariant variables. Since they are going to sink // outside the loop we do not need calculate cost for them. StoreInst *SI; if ((SI = dyn_cast(&I)) && Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) { ValuesToIgnore.insert(&I); DeadInvariantStoreOps[SI->getPointerOperand()].push_back( SI->getValueOperand()); } if (VecValuesToIgnore.contains(&I) || ValuesToIgnore.contains(&I)) continue; // Add instructions that would be trivially dead and are only used by // values already ignored to DeadOps to seed worklist. if (wouldInstructionBeTriviallyDead(&I, TLI) && all_of(I.users(), [this, IsLiveOutDead](User *U) { return VecValuesToIgnore.contains(U) || ValuesToIgnore.contains(U) || IsLiveOutDead(U); })) DeadOps.push_back(&I); // For interleave groups, we only create a pointer for the start of the // interleave group. Queue up addresses of group members except the insert // position for further processing. if (isAccessInterleaved(&I)) { auto *Group = getInterleavedAccessGroup(&I); if (Group->getInsertPos() == &I) continue; Value *PointerOp = getLoadStorePointerOperand(&I); DeadInterleavePointerOps.push_back(PointerOp); } // Queue branches for analysis. They are dead, if their successors only // contain dead instructions. if (auto *Br = dyn_cast(&I)) { if (Br->isConditional()) DeadOps.push_back(&I); } } // Mark ops feeding interleave group members as free, if they are only used // by other dead computations. for (unsigned I = 0; I != DeadInterleavePointerOps.size(); ++I) { auto *Op = dyn_cast(DeadInterleavePointerOps[I]); if (!Op || !TheLoop->contains(Op) || any_of(Op->users(), [this](User *U) { Instruction *UI = cast(U); return !VecValuesToIgnore.contains(U) && (!isAccessInterleaved(UI) || getInterleavedAccessGroup(UI)->getInsertPos() == UI); })) continue; VecValuesToIgnore.insert(Op); DeadInterleavePointerOps.append(Op->op_begin(), Op->op_end()); } for (const auto &[_, Ops] : DeadInvariantStoreOps) llvm::append_range(DeadOps, drop_end(Ops)); // Mark ops that would be trivially dead and are only used by ignored // instructions as free. BasicBlock *Header = TheLoop->getHeader(); // Returns true if the block contains only dead instructions. Such blocks will // be removed by VPlan-to-VPlan transforms and won't be considered by the // VPlan-based cost model, so skip them in the legacy cost-model as well. auto IsEmptyBlock = [this](BasicBlock *BB) { return all_of(*BB, [this](Instruction &I) { return ValuesToIgnore.contains(&I) || VecValuesToIgnore.contains(&I) || (isa(&I) && !cast(&I)->isConditional()); }); }; for (unsigned I = 0; I != DeadOps.size(); ++I) { auto *Op = dyn_cast(DeadOps[I]); // Check if the branch should be considered dead. if (auto *Br = dyn_cast_or_null(Op)) { BasicBlock *ThenBB = Br->getSuccessor(0); BasicBlock *ElseBB = Br->getSuccessor(1); // Don't considers branches leaving the loop for simplification. if (!TheLoop->contains(ThenBB) || !TheLoop->contains(ElseBB)) continue; bool ThenEmpty = IsEmptyBlock(ThenBB); bool ElseEmpty = IsEmptyBlock(ElseBB); if ((ThenEmpty && ElseEmpty) || (ThenEmpty && ThenBB->getSingleSuccessor() == ElseBB && ElseBB->phis().empty()) || (ElseEmpty && ElseBB->getSingleSuccessor() == ThenBB && ThenBB->phis().empty())) { VecValuesToIgnore.insert(Br); DeadOps.push_back(Br->getCondition()); } continue; } // Skip any op that shouldn't be considered dead. if (!Op || !TheLoop->contains(Op) || (isa(Op) && Op->getParent() == Header) || !wouldInstructionBeTriviallyDead(Op, TLI) || any_of(Op->users(), [this, IsLiveOutDead](User *U) { return !VecValuesToIgnore.contains(U) && !ValuesToIgnore.contains(U) && !IsLiveOutDead(U); })) continue; // If all of Op's users are in ValuesToIgnore, add it to ValuesToIgnore // which applies for both scalar and vector versions. Otherwise it is only // dead in vector versions, so only add it to VecValuesToIgnore. if (all_of(Op->users(), [this](User *U) { return ValuesToIgnore.contains(U); })) ValuesToIgnore.insert(Op); VecValuesToIgnore.insert(Op); DeadOps.append(Op->op_begin(), Op->op_end()); } // Ignore type-promoting instructions we identified during reduction // detection. for (const auto &Reduction : Legal->getReductionVars()) { const RecurrenceDescriptor &RedDes = Reduction.second; const SmallPtrSetImpl &Casts = RedDes.getCastInsts(); VecValuesToIgnore.insert_range(Casts); } // Ignore type-casting instructions we identified during induction // detection. for (const auto &Induction : Legal->getInductionVars()) { const InductionDescriptor &IndDes = Induction.second; const SmallVectorImpl &Casts = IndDes.getCastInsts(); VecValuesToIgnore.insert_range(Casts); } } void LoopVectorizationCostModel::collectInLoopReductions() { // Avoid duplicating work finding in-loop reductions. if (!InLoopReductions.empty()) return; for (const auto &Reduction : Legal->getReductionVars()) { PHINode *Phi = Reduction.first; const RecurrenceDescriptor &RdxDesc = Reduction.second; // We don't collect reductions that are type promoted (yet). if (RdxDesc.getRecurrenceType() != Phi->getType()) continue; // If the target would prefer this reduction to happen "in-loop", then we // want to record it as such. RecurKind Kind = RdxDesc.getRecurrenceKind(); if (!PreferInLoopReductions && !useOrderedReductions(RdxDesc) && !TTI.preferInLoopReduction(Kind, Phi->getType())) continue; // Check that we can correctly put the reductions into the loop, by // finding the chain of operations that leads from the phi to the loop // exit value. SmallVector ReductionOperations = RdxDesc.getReductionOpChain(Phi, TheLoop); bool InLoop = !ReductionOperations.empty(); if (InLoop) { InLoopReductions.insert(Phi); // Add the elements to InLoopReductionImmediateChains for cost modelling. Instruction *LastChain = Phi; for (auto *I : ReductionOperations) { InLoopReductionImmediateChains[I] = LastChain; LastChain = I; } } LLVM_DEBUG(dbgs() << "LV: Using " << (InLoop ? "inloop" : "out of loop") << " reduction for phi: " << *Phi << "\n"); } } // This function will select a scalable VF if the target supports scalable // vectors and a fixed one otherwise. // TODO: we could return a pair of values that specify the max VF and // min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of // `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment // doesn't have a cost model that can choose which plan to execute if // more than one is generated. static ElementCount determineVPlanVF(const TargetTransformInfo &TTI, LoopVectorizationCostModel &CM) { unsigned WidestType; std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes(); TargetTransformInfo::RegisterKind RegKind = TTI.enableScalableVectorization() ? TargetTransformInfo::RGK_ScalableVector : TargetTransformInfo::RGK_FixedWidthVector; TypeSize RegSize = TTI.getRegisterBitWidth(RegKind); unsigned N = RegSize.getKnownMinValue() / WidestType; return ElementCount::get(N, RegSize.isScalable()); } VectorizationFactor LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) { ElementCount VF = UserVF; // Outer loop handling: They may require CFG and instruction level // transformations before even evaluating whether vectorization is profitable. // Since we cannot modify the incoming IR, we need to build VPlan upfront in // the vectorization pipeline. if (!OrigLoop->isInnermost()) { // If the user doesn't provide a vectorization factor, determine a // reasonable one. if (UserVF.isZero()) { VF = determineVPlanVF(TTI, CM); LLVM_DEBUG(dbgs() << "LV: VPlan computed VF " << VF << ".\n"); // Make sure we have a VF > 1 for stress testing. if (VPlanBuildStressTest && (VF.isScalar() || VF.isZero())) { LLVM_DEBUG(dbgs() << "LV: VPlan stress testing: " << "overriding computed VF.\n"); VF = ElementCount::getFixed(4); } } else if (UserVF.isScalable() && !TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) { LLVM_DEBUG(dbgs() << "LV: Not vectorizing. Scalable VF requested, but " << "not supported by the target.\n"); reportVectorizationFailure( "Scalable vectorization requested but not supported by the target", "the scalable user-specified vectorization width for outer-loop " "vectorization cannot be used because the target does not support " "scalable vectors.", "ScalableVFUnfeasible", ORE, OrigLoop); return VectorizationFactor::Disabled(); } assert(EnableVPlanNativePath && "VPlan-native path is not enabled."); assert(isPowerOf2_32(VF.getKnownMinValue()) && "VF needs to be a power of two"); LLVM_DEBUG(dbgs() << "LV: Using " << (!UserVF.isZero() ? "user " : "") << "VF " << VF << " to build VPlans.\n"); buildVPlans(VF, VF); if (VPlans.empty()) return VectorizationFactor::Disabled(); // For VPlan build stress testing, we bail out after VPlan construction. if (VPlanBuildStressTest) return VectorizationFactor::Disabled(); return {VF, 0 /*Cost*/, 0 /* ScalarCost */}; } LLVM_DEBUG( dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the " "VPlan-native path.\n"); return VectorizationFactor::Disabled(); } void LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) { assert(OrigLoop->isInnermost() && "Inner loop expected."); CM.collectValuesToIgnore(); CM.collectElementTypesForWidening(); FixedScalableVFPair MaxFactors = CM.computeMaxVF(UserVF, UserIC); if (!MaxFactors) // Cases that should not to be vectorized nor interleaved. return; // Invalidate interleave groups if all blocks of loop will be predicated. if (CM.blockNeedsPredicationForAnyReason(OrigLoop->getHeader()) && !useMaskedInterleavedAccesses(TTI)) { LLVM_DEBUG( dbgs() << "LV: Invalidate all interleaved groups due to fold-tail by masking " "which requires masked-interleaved support.\n"); if (CM.InterleaveInfo.invalidateGroups()) // Invalidating interleave groups also requires invalidating all decisions // based on them, which includes widening decisions and uniform and scalar // values. CM.invalidateCostModelingDecisions(); } if (CM.foldTailByMasking()) Legal->prepareToFoldTailByMasking(); ElementCount MaxUserVF = UserVF.isScalable() ? MaxFactors.ScalableVF : MaxFactors.FixedVF; if (UserVF) { if (!ElementCount::isKnownLE(UserVF, MaxUserVF)) { reportVectorizationInfo( "UserVF ignored because it may be larger than the maximal safe VF", "InvalidUserVF", ORE, OrigLoop); } else { assert(isPowerOf2_32(UserVF.getKnownMinValue()) && "VF needs to be a power of two"); // Collect the instructions (and their associated costs) that will be more // profitable to scalarize. CM.collectInLoopReductions(); if (CM.selectUserVectorizationFactor(UserVF)) { LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n"); buildVPlansWithVPRecipes(UserVF, UserVF); LLVM_DEBUG(printPlans(dbgs())); return; } reportVectorizationInfo("UserVF ignored because of invalid costs.", "InvalidCost", ORE, OrigLoop); } } // Collect the Vectorization Factor Candidates. SmallVector VFCandidates; for (auto VF = ElementCount::getFixed(1); ElementCount::isKnownLE(VF, MaxFactors.FixedVF); VF *= 2) VFCandidates.push_back(VF); for (auto VF = ElementCount::getScalable(1); ElementCount::isKnownLE(VF, MaxFactors.ScalableVF); VF *= 2) VFCandidates.push_back(VF); CM.collectInLoopReductions(); for (const auto &VF : VFCandidates) { // Collect Uniform and Scalar instructions after vectorization with VF. CM.collectNonVectorizedAndSetWideningDecisions(VF); } buildVPlansWithVPRecipes(ElementCount::getFixed(1), MaxFactors.FixedVF); buildVPlansWithVPRecipes(ElementCount::getScalable(1), MaxFactors.ScalableVF); LLVM_DEBUG(printPlans(dbgs())); } InstructionCost VPCostContext::getLegacyCost(Instruction *UI, ElementCount VF) const { if (ForceTargetInstructionCost.getNumOccurrences()) return InstructionCost(ForceTargetInstructionCost.getNumOccurrences()); return CM.getInstructionCost(UI, VF); } bool VPCostContext::isLegacyUniformAfterVectorization(Instruction *I, ElementCount VF) const { return CM.isUniformAfterVectorization(I, VF); } bool VPCostContext::skipCostComputation(Instruction *UI, bool IsVector) const { return CM.ValuesToIgnore.contains(UI) || (IsVector && CM.VecValuesToIgnore.contains(UI)) || SkipCostComputation.contains(UI); } InstructionCost LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF, VPCostContext &CostCtx) const { InstructionCost Cost; // Cost modeling for inductions is inaccurate in the legacy cost model // compared to the recipes that are generated. To match here initially during // VPlan cost model bring up directly use the induction costs from the legacy // cost model. Note that we do this as pre-processing; the VPlan may not have // any recipes associated with the original induction increment instruction // and may replace truncates with VPWidenIntOrFpInductionRecipe. We precompute // the cost of induction phis and increments (both that are represented by // recipes and those that are not), to avoid distinguishing between them here, // and skip all recipes that represent induction phis and increments (the // former case) later on, if they exist, to avoid counting them twice. // Similarly we pre-compute the cost of any optimized truncates. // TODO: Switch to more accurate costing based on VPlan. for (const auto &[IV, IndDesc] : Legal->getInductionVars()) { Instruction *IVInc = cast( IV->getIncomingValueForBlock(OrigLoop->getLoopLatch())); SmallVector IVInsts = {IVInc}; for (unsigned I = 0; I != IVInsts.size(); I++) { for (Value *Op : IVInsts[I]->operands()) { auto *OpI = dyn_cast(Op); if (Op == IV || !OpI || !OrigLoop->contains(OpI) || !Op->hasOneUse()) continue; IVInsts.push_back(OpI); } } IVInsts.push_back(IV); for (User *U : IV->users()) { auto *CI = cast(U); if (!CostCtx.CM.isOptimizableIVTruncate(CI, VF)) continue; IVInsts.push_back(CI); } // If the vector loop gets executed exactly once with the given VF, ignore // the costs of comparison and induction instructions, as they'll get // simplified away. // TODO: Remove this code after stepping away from the legacy cost model and // adding code to simplify VPlans before calculating their costs. auto TC = getSmallConstantTripCount(PSE.getSE(), OrigLoop); if (TC == VF && !CM.foldTailByMasking()) addFullyUnrolledInstructionsToIgnore(OrigLoop, Legal->getInductionVars(), CostCtx.SkipCostComputation); for (Instruction *IVInst : IVInsts) { if (CostCtx.skipCostComputation(IVInst, VF.isVector())) continue; InstructionCost InductionCost = CostCtx.getLegacyCost(IVInst, VF); LLVM_DEBUG({ dbgs() << "Cost of " << InductionCost << " for VF " << VF << ": induction instruction " << *IVInst << "\n"; }); Cost += InductionCost; CostCtx.SkipCostComputation.insert(IVInst); } } /// Compute the cost of all exiting conditions of the loop using the legacy /// cost model. This is to match the legacy behavior, which adds the cost of /// all exit conditions. Note that this over-estimates the cost, as there will /// be a single condition to control the vector loop. SmallVector Exiting; CM.TheLoop->getExitingBlocks(Exiting); SetVector ExitInstrs; // Collect all exit conditions. for (BasicBlock *EB : Exiting) { auto *Term = dyn_cast(EB->getTerminator()); if (!Term || CostCtx.skipCostComputation(Term, VF.isVector())) continue; if (auto *CondI = dyn_cast(Term->getOperand(0))) { ExitInstrs.insert(CondI); } } // Compute the cost of all instructions only feeding the exit conditions. for (unsigned I = 0; I != ExitInstrs.size(); ++I) { Instruction *CondI = ExitInstrs[I]; if (!OrigLoop->contains(CondI) || !CostCtx.SkipCostComputation.insert(CondI).second) continue; InstructionCost CondICost = CostCtx.getLegacyCost(CondI, VF); LLVM_DEBUG({ dbgs() << "Cost of " << CondICost << " for VF " << VF << ": exit condition instruction " << *CondI << "\n"; }); Cost += CondICost; for (Value *Op : CondI->operands()) { auto *OpI = dyn_cast(Op); if (!OpI || CostCtx.skipCostComputation(OpI, VF.isVector()) || any_of(OpI->users(), [&ExitInstrs, this](User *U) { return OrigLoop->contains(cast(U)->getParent()) && !ExitInstrs.contains(cast(U)); })) continue; ExitInstrs.insert(OpI); } } // Pre-compute the costs for branches except for the backedge, as the number // of replicate regions in a VPlan may not directly match the number of // branches, which would lead to different decisions. // TODO: Compute cost of branches for each replicate region in the VPlan, // which is more accurate than the legacy cost model. for (BasicBlock *BB : OrigLoop->blocks()) { if (CostCtx.skipCostComputation(BB->getTerminator(), VF.isVector())) continue; CostCtx.SkipCostComputation.insert(BB->getTerminator()); if (BB == OrigLoop->getLoopLatch()) continue; auto BranchCost = CostCtx.getLegacyCost(BB->getTerminator(), VF); Cost += BranchCost; } // Pre-compute costs for instructions that are forced-scalar or profitable to // scalarize. Their costs will be computed separately in the legacy cost // model. for (Instruction *ForcedScalar : CM.ForcedScalars[VF]) { if (CostCtx.skipCostComputation(ForcedScalar, VF.isVector())) continue; CostCtx.SkipCostComputation.insert(ForcedScalar); InstructionCost ForcedCost = CostCtx.getLegacyCost(ForcedScalar, VF); LLVM_DEBUG({ dbgs() << "Cost of " << ForcedCost << " for VF " << VF << ": forced scalar " << *ForcedScalar << "\n"; }); Cost += ForcedCost; } for (const auto &[Scalarized, ScalarCost] : CM.InstsToScalarize[VF]) { if (CostCtx.skipCostComputation(Scalarized, VF.isVector())) continue; CostCtx.SkipCostComputation.insert(Scalarized); LLVM_DEBUG({ dbgs() << "Cost of " << ScalarCost << " for VF " << VF << ": profitable to scalarize " << *Scalarized << "\n"; }); Cost += ScalarCost; } return Cost; } InstructionCost LoopVectorizationPlanner::cost(VPlan &Plan, ElementCount VF) const { VPCostContext CostCtx(CM.TTI, *CM.TLI, Plan, CM, CM.CostKind); InstructionCost Cost = precomputeCosts(Plan, VF, CostCtx); // Now compute and add the VPlan-based cost. Cost += Plan.cost(VF, CostCtx); #ifndef NDEBUG unsigned EstimatedWidth = estimateElementCount(VF, CM.getVScaleForTuning()); LLVM_DEBUG(dbgs() << "Cost for VF " << VF << ": " << Cost << " (Estimated cost per lane: "); if (Cost.isValid()) { double CostPerLane = double(Cost.getValue()) / EstimatedWidth; LLVM_DEBUG(dbgs() << format("%.1f", CostPerLane)); } else /* No point dividing an invalid cost - it will still be invalid */ LLVM_DEBUG(dbgs() << "Invalid"); LLVM_DEBUG(dbgs() << ")\n"); #endif return Cost; } #ifndef NDEBUG /// Return true if the original loop \ TheLoop contains any instructions that do /// not have corresponding recipes in \p Plan and are not marked to be ignored /// in \p CostCtx. This means the VPlan contains simplification that the legacy /// cost-model did not account for. static bool planContainsAdditionalSimplifications(VPlan &Plan, VPCostContext &CostCtx, Loop *TheLoop, ElementCount VF) { // First collect all instructions for the recipes in Plan. auto GetInstructionForCost = [](const VPRecipeBase *R) -> Instruction * { if (auto *S = dyn_cast(R)) return dyn_cast_or_null(S->getUnderlyingValue()); if (auto *WidenMem = dyn_cast(R)) return &WidenMem->getIngredient(); return nullptr; }; DenseSet SeenInstrs; auto Iter = vp_depth_first_deep(Plan.getVectorLoopRegion()->getEntry()); for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly(Iter)) { for (VPRecipeBase &R : *VPBB) { if (auto *IR = dyn_cast(&R)) { auto *IG = IR->getInterleaveGroup(); unsigned NumMembers = IG->getNumMembers(); for (unsigned I = 0; I != NumMembers; ++I) { if (Instruction *M = IG->getMember(I)) SeenInstrs.insert(M); } continue; } // Unused FOR splices are removed by VPlan transforms, so the VPlan-based // cost model won't cost it whilst the legacy will. if (auto *FOR = dyn_cast(&R)) { if (none_of(FOR->users(), [](VPUser *U) { auto *VPI = dyn_cast(U); return VPI && VPI->getOpcode() == VPInstruction::FirstOrderRecurrenceSplice; })) return true; } // The VPlan-based cost model is more accurate for partial reduction and // comparing against the legacy cost isn't desirable. if (isa(&R)) return true; /// If a VPlan transform folded a recipe to one producing a single-scalar, /// but the original instruction wasn't uniform-after-vectorization in the /// legacy cost model, the legacy cost overestimates the actual cost. if (auto *RepR = dyn_cast(&R)) { if (RepR->isSingleScalar() && !CostCtx.isLegacyUniformAfterVectorization( RepR->getUnderlyingInstr(), VF)) return true; } if (Instruction *UI = GetInstructionForCost(&R)) { // If we adjusted the predicate of the recipe, the cost in the legacy // cost model may be different. if (auto *WidenCmp = dyn_cast(&R)) { if ((WidenCmp->getOpcode() == Instruction::ICmp || WidenCmp->getOpcode() == Instruction::FCmp) && WidenCmp->getPredicate() != cast(UI)->getPredicate()) return true; } SeenInstrs.insert(UI); } } } // Return true if the loop contains any instructions that are not also part of // the VPlan or are skipped for VPlan-based cost computations. This indicates // that the VPlan contains extra simplifications. return any_of(TheLoop->blocks(), [&SeenInstrs, &CostCtx, TheLoop](BasicBlock *BB) { return any_of(*BB, [&SeenInstrs, &CostCtx, TheLoop, BB](Instruction &I) { // Skip induction phis when checking for simplifications, as they may not // be lowered directly be lowered to a corresponding PHI recipe. if (isa(&I) && BB == TheLoop->getHeader() && CostCtx.CM.Legal->isInductionPhi(cast(&I))) return false; return !SeenInstrs.contains(&I) && !CostCtx.skipCostComputation(&I, true); }); }); } #endif VectorizationFactor LoopVectorizationPlanner::computeBestVF() { if (VPlans.empty()) return VectorizationFactor::Disabled(); // If there is a single VPlan with a single VF, return it directly. VPlan &FirstPlan = *VPlans[0]; if (VPlans.size() == 1 && size(FirstPlan.vectorFactors()) == 1) return {*FirstPlan.vectorFactors().begin(), 0, 0}; LLVM_DEBUG(dbgs() << "LV: Computing best VF using cost kind: " << (CM.CostKind == TTI::TCK_RecipThroughput ? "Reciprocal Throughput\n" : CM.CostKind == TTI::TCK_Latency ? "Instruction Latency\n" : CM.CostKind == TTI::TCK_CodeSize ? "Code Size\n" : CM.CostKind == TTI::TCK_SizeAndLatency ? "Code Size and Latency\n" : "Unknown\n")); ElementCount ScalarVF = ElementCount::getFixed(1); assert(hasPlanWithVF(ScalarVF) && "More than a single plan/VF w/o any plan having scalar VF"); // TODO: Compute scalar cost using VPlan-based cost model. InstructionCost ScalarCost = CM.expectedCost(ScalarVF); LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ScalarCost << ".\n"); VectorizationFactor ScalarFactor(ScalarVF, ScalarCost, ScalarCost); VectorizationFactor BestFactor = ScalarFactor; bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled; if (ForceVectorization) { // Ignore scalar width, because the user explicitly wants vectorization. // Initialize cost to max so that VF = 2 is, at least, chosen during cost // evaluation. BestFactor.Cost = InstructionCost::getMax(); } for (auto &P : VPlans) { ArrayRef VFs(P->vectorFactors().begin(), P->vectorFactors().end()); SmallVector RUs; if (CM.useMaxBandwidth(TargetTransformInfo::RGK_ScalableVector) || CM.useMaxBandwidth(TargetTransformInfo::RGK_FixedWidthVector)) RUs = calculateRegisterUsageForPlan(*P, VFs, TTI, CM.ValuesToIgnore); for (unsigned I = 0; I < VFs.size(); I++) { ElementCount VF = VFs[I]; if (VF.isScalar()) continue; if (!ForceVectorization && !willGenerateVectors(*P, VF, TTI)) { LLVM_DEBUG( dbgs() << "LV: Not considering vector loop of width " << VF << " because it will not generate any vector instructions.\n"); continue; } if (CM.OptForSize && !ForceVectorization && hasReplicatorRegion(*P)) { LLVM_DEBUG( dbgs() << "LV: Not considering vector loop of width " << VF << " because it would cause replicated blocks to be generated," << " which isn't allowed when optimizing for size.\n"); continue; } InstructionCost Cost = cost(*P, VF); VectorizationFactor CurrentFactor(VF, Cost, ScalarCost); if (CM.shouldCalculateRegPressureForVF(VF) && RUs[I].exceedsMaxNumRegs(TTI, ForceTargetNumVectorRegs)) { LLVM_DEBUG(dbgs() << "LV(REG): Not considering vector loop of width " << VF << " because it uses too many registers\n"); continue; } if (isMoreProfitable(CurrentFactor, BestFactor, P->hasScalarTail())) BestFactor = CurrentFactor; // If profitable add it to ProfitableVF list. if (isMoreProfitable(CurrentFactor, ScalarFactor, P->hasScalarTail())) ProfitableVFs.push_back(CurrentFactor); } } #ifndef NDEBUG // Select the optimal vectorization factor according to the legacy cost-model. // This is now only used to verify the decisions by the new VPlan-based // cost-model and will be retired once the VPlan-based cost-model is // stabilized. VectorizationFactor LegacyVF = selectVectorizationFactor(); VPlan &BestPlan = getPlanFor(BestFactor.Width); // Pre-compute the cost and use it to check if BestPlan contains any // simplifications not accounted for in the legacy cost model. If that's the // case, don't trigger the assertion, as the extra simplifications may cause a // different VF to be picked by the VPlan-based cost model. VPCostContext CostCtx(CM.TTI, *CM.TLI, BestPlan, CM, CM.CostKind); precomputeCosts(BestPlan, BestFactor.Width, CostCtx); // Verify that the VPlan-based and legacy cost models agree, except for VPlans // with early exits and plans with additional VPlan simplifications. The // legacy cost model doesn't properly model costs for such loops. assert((BestFactor.Width == LegacyVF.Width || BestPlan.hasEarlyExit() || planContainsAdditionalSimplifications(getPlanFor(BestFactor.Width), CostCtx, OrigLoop, BestFactor.Width) || planContainsAdditionalSimplifications( getPlanFor(LegacyVF.Width), CostCtx, OrigLoop, LegacyVF.Width)) && " VPlan cost model and legacy cost model disagreed"); assert((BestFactor.Width.isScalar() || BestFactor.ScalarCost > 0) && "when vectorizing, the scalar cost must be computed."); #endif LLVM_DEBUG(dbgs() << "LV: Selecting VF: " << BestFactor.Width << ".\n"); return BestFactor; } static void addRuntimeUnrollDisableMetaData(Loop *L) { SmallVector MDs; // Reserve first location for self reference to the LoopID metadata node. MDs.push_back(nullptr); bool IsUnrollMetadata = false; MDNode *LoopID = L->getLoopID(); if (LoopID) { // First find existing loop unrolling disable metadata. for (unsigned I = 1, IE = LoopID->getNumOperands(); I < IE; ++I) { auto *MD = dyn_cast(LoopID->getOperand(I)); if (MD) { const auto *S = dyn_cast(MD->getOperand(0)); IsUnrollMetadata = S && S->getString().starts_with("llvm.loop.unroll.disable"); } MDs.push_back(LoopID->getOperand(I)); } } if (!IsUnrollMetadata) { // Add runtime unroll disable metadata. LLVMContext &Context = L->getHeader()->getContext(); SmallVector DisableOperands; DisableOperands.push_back( MDString::get(Context, "llvm.loop.unroll.runtime.disable")); MDNode *DisableNode = MDNode::get(Context, DisableOperands); MDs.push_back(DisableNode); MDNode *NewLoopID = MDNode::get(Context, MDs); // Set operand 0 to refer to the loop id itself. NewLoopID->replaceOperandWith(0, NewLoopID); L->setLoopID(NewLoopID); } } static Value *getStartValueFromReductionResult(VPInstruction *RdxResult) { using namespace VPlanPatternMatch; assert(RdxResult->getOpcode() == VPInstruction::ComputeFindIVResult && "RdxResult must be ComputeFindIVResult"); VPValue *StartVPV = RdxResult->getOperand(1); match(StartVPV, m_Freeze(m_VPValue(StartVPV))); return StartVPV->getLiveInIRValue(); } // If \p EpiResumePhiR is resume VPPhi for a reduction when vectorizing the // epilog loop, fix the reduction's scalar PHI node by adding the incoming value // from the main vector loop. static void fixReductionScalarResumeWhenVectorizingEpilog( VPPhi *EpiResumePhiR, VPTransformState &State, BasicBlock *BypassBlock) { // Get the VPInstruction computing the reduction result in the middle block. // The first operand may not be from the middle block if it is not connected // to the scalar preheader. In that case, there's nothing to fix. VPValue *Incoming = EpiResumePhiR->getOperand(0); match(Incoming, VPlanPatternMatch::m_ZExtOrSExt( VPlanPatternMatch::m_VPValue(Incoming))); auto *EpiRedResult = dyn_cast(Incoming); if (!EpiRedResult || (EpiRedResult->getOpcode() != VPInstruction::ComputeAnyOfResult && EpiRedResult->getOpcode() != VPInstruction::ComputeReductionResult && EpiRedResult->getOpcode() != VPInstruction::ComputeFindIVResult)) return; auto *EpiRedHeaderPhi = cast(EpiRedResult->getOperand(0)); RecurKind Kind = EpiRedHeaderPhi->getRecurrenceKind(); Value *MainResumeValue; if (auto *VPI = dyn_cast(EpiRedHeaderPhi->getStartValue())) { assert((VPI->getOpcode() == VPInstruction::Broadcast || VPI->getOpcode() == VPInstruction::ReductionStartVector) && "unexpected start recipe"); MainResumeValue = VPI->getOperand(0)->getUnderlyingValue(); } else MainResumeValue = EpiRedHeaderPhi->getStartValue()->getUnderlyingValue(); if (RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind)) { [[maybe_unused]] Value *StartV = EpiRedResult->getOperand(1)->getLiveInIRValue(); auto *Cmp = cast(MainResumeValue); assert(Cmp->getPredicate() == CmpInst::ICMP_NE && "AnyOf expected to start with ICMP_NE"); assert(Cmp->getOperand(1) == StartV && "AnyOf expected to start by comparing main resume value to original " "start value"); MainResumeValue = Cmp->getOperand(0); } else if (RecurrenceDescriptor::isFindIVRecurrenceKind(Kind)) { Value *StartV = getStartValueFromReductionResult(EpiRedResult); Value *SentinelV = EpiRedResult->getOperand(2)->getLiveInIRValue(); using namespace llvm::PatternMatch; Value *Cmp, *OrigResumeV, *CmpOp; [[maybe_unused]] bool IsExpectedPattern = match(MainResumeValue, m_Select(m_OneUse(m_Value(Cmp)), m_Specific(SentinelV), m_Value(OrigResumeV))) && (match(Cmp, m_SpecificICmp(ICmpInst::ICMP_EQ, m_Specific(OrigResumeV), m_Value(CmpOp))) && ((CmpOp == StartV && isGuaranteedNotToBeUndefOrPoison(CmpOp)))); assert(IsExpectedPattern && "Unexpected reduction resume pattern"); MainResumeValue = OrigResumeV; } PHINode *MainResumePhi = cast(MainResumeValue); // When fixing reductions in the epilogue loop we should already have // created a bc.merge.rdx Phi after the main vector body. Ensure that we carry // over the incoming values correctly. auto *EpiResumePhi = cast(State.get(EpiResumePhiR, true)); EpiResumePhi->setIncomingValueForBlock( BypassBlock, MainResumePhi->getIncomingValueForBlock(BypassBlock)); } DenseMap LoopVectorizationPlanner::executePlan( ElementCount BestVF, unsigned BestUF, VPlan &BestVPlan, InnerLoopVectorizer &ILV, DominatorTree *DT, bool VectorizingEpilogue) { assert(BestVPlan.hasVF(BestVF) && "Trying to execute plan with unsupported VF"); assert(BestVPlan.hasUF(BestUF) && "Trying to execute plan with unsupported UF"); if (BestVPlan.hasEarlyExit()) ++LoopsEarlyExitVectorized; // TODO: Move to VPlan transform stage once the transition to the VPlan-based // cost model is complete for better cost estimates. VPlanTransforms::runPass(VPlanTransforms::unrollByUF, BestVPlan, BestUF); VPlanTransforms::runPass(VPlanTransforms::replicateByVF, BestVPlan, BestVF); VPlanTransforms::runPass(VPlanTransforms::materializeBroadcasts, BestVPlan); bool HasBranchWeights = hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator()); if (HasBranchWeights) { std::optional VScale = CM.getVScaleForTuning(); VPlanTransforms::runPass(VPlanTransforms::addBranchWeightToMiddleTerminator, BestVPlan, BestVF, VScale); } if (!VectorizingEpilogue) { // Checks are the same for all VPlans, added to BestVPlan only for // compactness. attachRuntimeChecks(BestVPlan, ILV.RTChecks, HasBranchWeights); } // Retrieving VectorPH now when it's easier while VPlan still has Regions. VPBasicBlock *VectorPH = cast(BestVPlan.getVectorPreheader()); VPlanTransforms::optimizeForVFAndUF(BestVPlan, BestVF, BestUF, PSE); VPlanTransforms::simplifyRecipes(BestVPlan); VPlanTransforms::removeBranchOnConst(BestVPlan); VPlanTransforms::narrowInterleaveGroups( BestVPlan, BestVF, TTI.getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)); VPlanTransforms::removeDeadRecipes(BestVPlan); VPlanTransforms::convertToConcreteRecipes(BestVPlan); // Regions are dissolved after optimizing for VF and UF, which completely // removes unneeded loop regions first. VPlanTransforms::dissolveLoopRegions(BestVPlan); // Canonicalize EVL loops after regions are dissolved. VPlanTransforms::canonicalizeEVLLoops(BestVPlan); VPlanTransforms::materializeBackedgeTakenCount(BestVPlan, VectorPH); VPlanTransforms::materializeVectorTripCount( BestVPlan, VectorPH, CM.foldTailByMasking(), CM.requiresScalarEpilogue(BestVF.isVector())); VPlanTransforms::materializeVFAndVFxUF(BestVPlan, VectorPH, BestVF); // Perform the actual loop transformation. VPTransformState State(&TTI, BestVF, LI, DT, ILV.AC, ILV.Builder, &BestVPlan, OrigLoop->getParentLoop(), Legal->getWidestInductionType()); #ifdef EXPENSIVE_CHECKS assert(DT->verify(DominatorTree::VerificationLevel::Fast)); #endif // 0. Generate SCEV-dependent code in the entry, including TripCount, before // making any changes to the CFG. DenseMap ExpandedSCEVs; auto *Entry = cast(BestVPlan.getEntry()); State.Builder.SetInsertPoint(Entry->getIRBasicBlock()->getTerminator()); for (VPRecipeBase &R : make_early_inc_range(*Entry)) { auto *ExpSCEV = dyn_cast(&R); if (!ExpSCEV) continue; ExpSCEV->execute(State); ExpandedSCEVs[ExpSCEV->getSCEV()] = State.get(ExpSCEV, VPLane(0)); VPValue *Exp = BestVPlan.getOrAddLiveIn(ExpandedSCEVs[ExpSCEV->getSCEV()]); ExpSCEV->replaceAllUsesWith(Exp); if (BestVPlan.getTripCount() == ExpSCEV) BestVPlan.resetTripCount(Exp); ExpSCEV->eraseFromParent(); } if (!ILV.getTripCount()) ILV.setTripCount(State.get(BestVPlan.getTripCount(), VPLane(0))); else assert(VectorizingEpilogue && "should only re-use the existing trip " "count during epilogue vectorization"); // 1. Set up the skeleton for vectorization, including vector pre-header and // middle block. The vector loop is created during VPlan execution. BasicBlock *EntryBB = cast(BestVPlan.getEntry())->getIRBasicBlock(); State.CFG.PrevBB = ILV.createVectorizedLoopSkeleton(); VPlanTransforms::removeDeadRecipes(BestVPlan); assert(verifyVPlanIsValid(BestVPlan, true /*VerifyLate*/) && "final VPlan is invalid"); ILV.printDebugTracesAtStart(); //===------------------------------------------------===// // // Notice: any optimization or new instruction that go // into the code below should also be implemented in // the cost-model. // //===------------------------------------------------===// // Move check blocks to their final position. // TODO: Move as part of VPIRBB execute and update impacted tests. if (BasicBlock *MemCheckBlock = ILV.RTChecks.getMemRuntimeChecks().second) MemCheckBlock->moveAfter(EntryBB); if (BasicBlock *SCEVCheckBlock = ILV.RTChecks.getSCEVChecks().second) SCEVCheckBlock->moveAfter(EntryBB); BestVPlan.execute(&State); // 2.5 When vectorizing the epilogue, fix reduction resume values from the // additional bypass block. if (VectorizingEpilogue) { assert(!BestVPlan.hasEarlyExit() && "Epilogue vectorisation not yet supported with early exits"); BasicBlock *PH = OrigLoop->getLoopPreheader(); BasicBlock *BypassBlock = ILV.getAdditionalBypassBlock(); for (auto *Pred : predecessors(PH)) { for (PHINode &Phi : PH->phis()) { if (Phi.getBasicBlockIndex(Pred) != -1) continue; Phi.addIncoming(Phi.getIncomingValueForBlock(BypassBlock), Pred); } } VPBasicBlock *ScalarPH = BestVPlan.getScalarPreheader(); if (ScalarPH->getNumPredecessors() > 0) { // If ScalarPH has predecessors, we may need to update its reduction // resume values. for (VPRecipeBase &R : ScalarPH->phis()) { fixReductionScalarResumeWhenVectorizingEpilog(cast(&R), State, BypassBlock); } } } // 2.6. Maintain Loop Hints // Keep all loop hints from the original loop on the vector loop (we'll // replace the vectorizer-specific hints below). VPBasicBlock *HeaderVPBB = vputils::getFirstLoopHeader(BestVPlan, State.VPDT); if (HeaderVPBB) { MDNode *OrigLoopID = OrigLoop->getLoopID(); std::optional VectorizedLoopID = makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, LLVMLoopVectorizeFollowupVectorized}); Loop *L = LI->getLoopFor(State.CFG.VPBB2IRBB[HeaderVPBB]); if (VectorizedLoopID) { L->setLoopID(*VectorizedLoopID); } else { // Keep all loop hints from the original loop on the vector loop (we'll // replace the vectorizer-specific hints below). if (MDNode *LID = OrigLoop->getLoopID()) L->setLoopID(LID); LoopVectorizeHints Hints(L, true, *ORE); Hints.setAlreadyVectorized(); // Check if it's EVL-vectorized and mark the corresponding metadata. bool IsEVLVectorized = llvm::any_of(*HeaderVPBB, [](const VPRecipeBase &Recipe) { // Looking for the ExplictVectorLength VPInstruction. if (const auto *VI = dyn_cast(&Recipe)) return VI->getOpcode() == VPInstruction::ExplicitVectorLength; return false; }); if (IsEVLVectorized) { LLVMContext &Context = L->getHeader()->getContext(); MDNode *LoopID = L->getLoopID(); auto *IsEVLVectorizedMD = MDNode::get( Context, {MDString::get(Context, "llvm.loop.isvectorized.tailfoldingstyle"), MDString::get(Context, "evl")}); MDNode *NewLoopID = makePostTransformationMetadata(Context, LoopID, {}, {IsEVLVectorizedMD}); L->setLoopID(NewLoopID); } } TargetTransformInfo::UnrollingPreferences UP; TTI.getUnrollingPreferences(L, *PSE.getSE(), UP, ORE); if (!UP.UnrollVectorizedLoop || VectorizingEpilogue) addRuntimeUnrollDisableMetaData(L); } // 3. Fix the vectorized code: take care of header phi's, live-outs, // predication, updating analyses. ILV.fixVectorizedLoop(State); ILV.printDebugTracesAtEnd(); return ExpandedSCEVs; } //===--------------------------------------------------------------------===// // EpilogueVectorizerMainLoop //===--------------------------------------------------------------------===// /// This function is partially responsible for generating the control flow /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. BasicBlock *EpilogueVectorizerMainLoop::createEpilogueVectorizedLoopSkeleton() { createVectorLoopSkeleton(""); // Generate the code to check the minimum iteration count of the vector // epilogue (see below). EPI.EpilogueIterationCountCheck = emitIterationCountCheck(LoopScalarPreHeader, true); EPI.EpilogueIterationCountCheck->setName("iter.check"); // Generate the iteration count check for the main loop, *after* the check // for the epilogue loop, so that the path-length is shorter for the case // that goes directly through the vector epilogue. The longer-path length for // the main loop is compensated for, by the gain from vectorizing the larger // trip count. Note: the branch will get updated later on when we vectorize // the epilogue. EPI.MainLoopIterationCountCheck = emitIterationCountCheck(LoopScalarPreHeader, false); replaceVPBBWithIRVPBB(Plan.getScalarPreheader(), LoopScalarPreHeader); return LoopVectorPreHeader; } void EpilogueVectorizerMainLoop::printDebugTracesAtStart() { LLVM_DEBUG({ dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n" << "Main Loop VF:" << EPI.MainLoopVF << ", Main Loop UF:" << EPI.MainLoopUF << ", Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }); } void EpilogueVectorizerMainLoop::printDebugTracesAtEnd() { DEBUG_WITH_TYPE(VerboseDebug, { dbgs() << "intermediate fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }); } BasicBlock * EpilogueVectorizerMainLoop::emitIterationCountCheck(BasicBlock *Bypass, bool ForEpilogue) { assert(Bypass && "Expected valid bypass basic block."); Value *Count = getTripCount(); MinProfitableTripCount = ElementCount::getFixed(0); Value *CheckMinIters = createIterationCountCheck(ForEpilogue ? EPI.EpilogueVF : EPI.MainLoopVF, ForEpilogue ? EPI.EpilogueUF : EPI.MainLoopUF); BasicBlock *const TCCheckBlock = LoopVectorPreHeader; if (!ForEpilogue) TCCheckBlock->setName("vector.main.loop.iter.check"); // Create new preheader for vector loop. LoopVectorPreHeader = SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(), static_cast(nullptr), LI, nullptr, "vector.ph"); if (ForEpilogue) { // Save the trip count so we don't have to regenerate it in the // vec.epilog.iter.check. This is safe to do because the trip count // generated here dominates the vector epilog iter check. EPI.TripCount = Count; } else { VectorPHVPBB = replaceVPBBWithIRVPBB(VectorPHVPBB, LoopVectorPreHeader); } BranchInst &BI = *BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters); if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator())) setBranchWeights(BI, MinItersBypassWeights, /*IsExpected=*/false); ReplaceInstWithInst(TCCheckBlock->getTerminator(), &BI); // When vectorizing the main loop, its trip-count check is placed in a new // block, whereas the overall trip-count check is placed in the VPlan entry // block. When vectorizing the epilogue loop, its trip-count check is placed // in the VPlan entry block. if (!ForEpilogue) introduceCheckBlockInVPlan(TCCheckBlock); return TCCheckBlock; } //===--------------------------------------------------------------------===// // EpilogueVectorizerEpilogueLoop //===--------------------------------------------------------------------===// /// This function is partially responsible for generating the control flow /// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization. BasicBlock * EpilogueVectorizerEpilogueLoop::createEpilogueVectorizedLoopSkeleton() { createVectorLoopSkeleton("vec.epilog."); // Now, compare the remaining count and if there aren't enough iterations to // execute the vectorized epilogue skip to the scalar part. LoopVectorPreHeader->setName("vec.epilog.ph"); BasicBlock *VecEpilogueIterationCountCheck = SplitBlock(LoopVectorPreHeader, LoopVectorPreHeader->begin(), DT, LI, nullptr, "vec.epilog.iter.check", true); VectorPHVPBB = replaceVPBBWithIRVPBB(VectorPHVPBB, LoopVectorPreHeader); emitMinimumVectorEpilogueIterCountCheck(LoopScalarPreHeader, VecEpilogueIterationCountCheck); AdditionalBypassBlock = VecEpilogueIterationCountCheck; // Adjust the control flow taking the state info from the main loop // vectorization into account. assert(EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck && "expected this to be saved from the previous pass."); EPI.MainLoopIterationCountCheck->getTerminator()->replaceUsesOfWith( VecEpilogueIterationCountCheck, LoopVectorPreHeader); EPI.EpilogueIterationCountCheck->getTerminator()->replaceUsesOfWith( VecEpilogueIterationCountCheck, LoopScalarPreHeader); // Adjust the terminators of runtime check blocks and phis using them. BasicBlock *SCEVCheckBlock = RTChecks.getSCEVChecks().second; BasicBlock *MemCheckBlock = RTChecks.getMemRuntimeChecks().second; if (SCEVCheckBlock) SCEVCheckBlock->getTerminator()->replaceUsesOfWith( VecEpilogueIterationCountCheck, LoopScalarPreHeader); if (MemCheckBlock) MemCheckBlock->getTerminator()->replaceUsesOfWith( VecEpilogueIterationCountCheck, LoopScalarPreHeader); DT->changeImmediateDominator(LoopScalarPreHeader, EPI.EpilogueIterationCountCheck); // The vec.epilog.iter.check block may contain Phi nodes from inductions or // reductions which merge control-flow from the latch block and the middle // block. Update the incoming values here and move the Phi into the preheader. SmallVector PhisInBlock( llvm::make_pointer_range(VecEpilogueIterationCountCheck->phis())); for (PHINode *Phi : PhisInBlock) { Phi->moveBefore(LoopVectorPreHeader->getFirstNonPHIIt()); Phi->replaceIncomingBlockWith( VecEpilogueIterationCountCheck->getSinglePredecessor(), VecEpilogueIterationCountCheck); // If the phi doesn't have an incoming value from the // EpilogueIterationCountCheck, we are done. Otherwise remove the incoming // value and also those from other check blocks. This is needed for // reduction phis only. if (none_of(Phi->blocks(), [&](BasicBlock *IncB) { return EPI.EpilogueIterationCountCheck == IncB; })) continue; Phi->removeIncomingValue(EPI.EpilogueIterationCountCheck); if (SCEVCheckBlock) Phi->removeIncomingValue(SCEVCheckBlock); if (MemCheckBlock) Phi->removeIncomingValue(MemCheckBlock); } replaceVPBBWithIRVPBB(Plan.getScalarPreheader(), LoopScalarPreHeader); return LoopVectorPreHeader; } BasicBlock * EpilogueVectorizerEpilogueLoop::emitMinimumVectorEpilogueIterCountCheck( BasicBlock *Bypass, BasicBlock *Insert) { assert(EPI.TripCount && "Expected trip count to have been saved in the first pass."); Value *TC = EPI.TripCount; IRBuilder<> Builder(Insert->getTerminator()); Value *Count = Builder.CreateSub(TC, EPI.VectorTripCount, "n.vec.remaining"); // Generate code to check if the loop's trip count is less than VF * UF of the // vector epilogue loop. auto P = Cost->requiresScalarEpilogue(EPI.EpilogueVF.isVector()) ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_ULT; Value *CheckMinIters = Builder.CreateICmp(P, Count, createStepForVF(Builder, Count->getType(), EPI.EpilogueVF, EPI.EpilogueUF), "min.epilog.iters.check"); BranchInst &BI = *BranchInst::Create(Bypass, LoopVectorPreHeader, CheckMinIters); if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator())) { auto VScale = Cost->getVScaleForTuning(); unsigned MainLoopStep = estimateElementCount(EPI.MainLoopVF * EPI.MainLoopUF, VScale); unsigned EpilogueLoopStep = estimateElementCount(EPI.EpilogueVF * EPI.EpilogueUF, VScale); // We assume the remaining `Count` is equally distributed in // [0, MainLoopStep) // So the probability for `Count < EpilogueLoopStep` should be // min(MainLoopStep, EpilogueLoopStep) / MainLoopStep unsigned EstimatedSkipCount = std::min(MainLoopStep, EpilogueLoopStep); const uint32_t Weights[] = {EstimatedSkipCount, MainLoopStep - EstimatedSkipCount}; setBranchWeights(BI, Weights, /*IsExpected=*/false); } ReplaceInstWithInst(Insert->getTerminator(), &BI); // A new entry block has been created for the epilogue VPlan. Hook it in, as // otherwise we would try to modify the entry to the main vector loop. VPIRBasicBlock *NewEntry = Plan.createVPIRBasicBlock(Insert); VPBasicBlock *OldEntry = Plan.getEntry(); VPBlockUtils::reassociateBlocks(OldEntry, NewEntry); Plan.setEntry(NewEntry); // OldEntry is now dead and will be cleaned up when the plan gets destroyed. return Insert; } void EpilogueVectorizerEpilogueLoop::printDebugTracesAtStart() { LLVM_DEBUG({ dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n" << "Epilogue Loop VF:" << EPI.EpilogueVF << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n"; }); } void EpilogueVectorizerEpilogueLoop::printDebugTracesAtEnd() { DEBUG_WITH_TYPE(VerboseDebug, { dbgs() << "final fn:\n" << *OrigLoop->getHeader()->getParent() << "\n"; }); } VPWidenMemoryRecipe * VPRecipeBuilder::tryToWidenMemory(Instruction *I, ArrayRef Operands, VFRange &Range) { assert((isa(I) || isa(I)) && "Must be called with either a load or store"); auto WillWiden = [&](ElementCount VF) -> bool { LoopVectorizationCostModel::InstWidening Decision = CM.getWideningDecision(I, VF); assert(Decision != LoopVectorizationCostModel::CM_Unknown && "CM decision should be taken at this point."); if (Decision == LoopVectorizationCostModel::CM_Interleave) return true; if (CM.isScalarAfterVectorization(I, VF) || CM.isProfitableToScalarize(I, VF)) return false; return Decision != LoopVectorizationCostModel::CM_Scalarize; }; if (!LoopVectorizationPlanner::getDecisionAndClampRange(WillWiden, Range)) return nullptr; VPValue *Mask = nullptr; if (Legal->isMaskRequired(I)) Mask = getBlockInMask(Builder.getInsertBlock()); // Determine if the pointer operand of the access is either consecutive or // reverse consecutive. LoopVectorizationCostModel::InstWidening Decision = CM.getWideningDecision(I, Range.Start); bool Reverse = Decision == LoopVectorizationCostModel::CM_Widen_Reverse; bool Consecutive = Reverse || Decision == LoopVectorizationCostModel::CM_Widen; VPValue *Ptr = isa(I) ? Operands[0] : Operands[1]; if (Consecutive) { auto *GEP = dyn_cast( Ptr->getUnderlyingValue()->stripPointerCasts()); VPSingleDefRecipe *VectorPtr; if (Reverse) { // When folding the tail, we may compute an address that we don't in the // original scalar loop and it may not be inbounds. Drop Inbounds in that // case. GEPNoWrapFlags Flags = (CM.foldTailByMasking() || !GEP || !GEP->isInBounds()) ? GEPNoWrapFlags::none() : GEPNoWrapFlags::inBounds(); VectorPtr = new VPVectorEndPointerRecipe(Ptr, &Plan.getVF(), getLoadStoreType(I), /*Stride*/ -1, Flags, I->getDebugLoc()); } else { VectorPtr = new VPVectorPointerRecipe(Ptr, getLoadStoreType(I), GEP ? GEP->getNoWrapFlags() : GEPNoWrapFlags::none(), I->getDebugLoc()); } Builder.insert(VectorPtr); Ptr = VectorPtr; } if (LoadInst *Load = dyn_cast(I)) return new VPWidenLoadRecipe(*Load, Ptr, Mask, Consecutive, Reverse, VPIRMetadata(*Load, LVer), I->getDebugLoc()); StoreInst *Store = cast(I); return new VPWidenStoreRecipe(*Store, Ptr, Operands[0], Mask, Consecutive, Reverse, VPIRMetadata(*Store, LVer), I->getDebugLoc()); } /// Creates a VPWidenIntOrFpInductionRecpipe for \p Phi. If needed, it will also /// insert a recipe to expand the step for the induction recipe. static VPWidenIntOrFpInductionRecipe * createWidenInductionRecipes(PHINode *Phi, Instruction *PhiOrTrunc, VPValue *Start, const InductionDescriptor &IndDesc, VPlan &Plan, ScalarEvolution &SE, Loop &OrigLoop) { assert(IndDesc.getStartValue() == Phi->getIncomingValueForBlock(OrigLoop.getLoopPreheader())); assert(SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) && "step must be loop invariant"); VPValue *Step = vputils::getOrCreateVPValueForSCEVExpr(Plan, IndDesc.getStep(), SE); if (auto *TruncI = dyn_cast(PhiOrTrunc)) { return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, &Plan.getVF(), IndDesc, TruncI, TruncI->getDebugLoc()); } assert(isa(PhiOrTrunc) && "must be a phi node here"); return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, &Plan.getVF(), IndDesc, Phi->getDebugLoc()); } VPHeaderPHIRecipe *VPRecipeBuilder::tryToOptimizeInductionPHI( PHINode *Phi, ArrayRef Operands, VFRange &Range) { // Check if this is an integer or fp induction. If so, build the recipe that // produces its scalar and vector values. if (auto *II = Legal->getIntOrFpInductionDescriptor(Phi)) return createWidenInductionRecipes(Phi, Phi, Operands[0], *II, Plan, *PSE.getSE(), *OrigLoop); // Check if this is pointer induction. If so, build the recipe for it. if (auto *II = Legal->getPointerInductionDescriptor(Phi)) { VPValue *Step = vputils::getOrCreateVPValueForSCEVExpr(Plan, II->getStep(), *PSE.getSE()); return new VPWidenPointerInductionRecipe( Phi, Operands[0], Step, &Plan.getVFxUF(), *II, LoopVectorizationPlanner::getDecisionAndClampRange( [&](ElementCount VF) { return CM.isScalarAfterVectorization(Phi, VF); }, Range), Phi->getDebugLoc()); } return nullptr; } VPWidenIntOrFpInductionRecipe *VPRecipeBuilder::tryToOptimizeInductionTruncate( TruncInst *I, ArrayRef Operands, VFRange &Range) { // Optimize the special case where the source is a constant integer // induction variable. Notice that we can only optimize the 'trunc' case // because (a) FP conversions lose precision, (b) sext/zext may wrap, and // (c) other casts depend on pointer size. // Determine whether \p K is a truncation based on an induction variable that // can be optimized. auto IsOptimizableIVTruncate = [&](Instruction *K) -> std::function { return [=](ElementCount VF) -> bool { return CM.isOptimizableIVTruncate(K, VF); }; }; if (LoopVectorizationPlanner::getDecisionAndClampRange( IsOptimizableIVTruncate(I), Range)) { auto *Phi = cast(I->getOperand(0)); const InductionDescriptor &II = *Legal->getIntOrFpInductionDescriptor(Phi); VPValue *Start = Plan.getOrAddLiveIn(II.getStartValue()); return createWidenInductionRecipes(Phi, I, Start, II, Plan, *PSE.getSE(), *OrigLoop); } return nullptr; } VPSingleDefRecipe *VPRecipeBuilder::tryToWidenCall(CallInst *CI, ArrayRef Operands, VFRange &Range) { bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange( [this, CI](ElementCount VF) { return CM.isScalarWithPredication(CI, VF); }, Range); if (IsPredicated) return nullptr; Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI); if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end || ID == Intrinsic::lifetime_start || ID == Intrinsic::sideeffect || ID == Intrinsic::pseudoprobe || ID == Intrinsic::experimental_noalias_scope_decl)) return nullptr; SmallVector Ops(Operands.take_front(CI->arg_size())); // Is it beneficial to perform intrinsic call compared to lib call? bool ShouldUseVectorIntrinsic = ID && LoopVectorizationPlanner::getDecisionAndClampRange( [&](ElementCount VF) -> bool { return CM.getCallWideningDecision(CI, VF).Kind == LoopVectorizationCostModel::CM_IntrinsicCall; }, Range); if (ShouldUseVectorIntrinsic) return new VPWidenIntrinsicRecipe(*CI, ID, Ops, CI->getType(), CI->getDebugLoc()); Function *Variant = nullptr; std::optional MaskPos; // Is better to call a vectorized version of the function than to to scalarize // the call? auto ShouldUseVectorCall = LoopVectorizationPlanner::getDecisionAndClampRange( [&](ElementCount VF) -> bool { // The following case may be scalarized depending on the VF. // The flag shows whether we can use a usual Call for vectorized // version of the instruction. // If we've found a variant at a previous VF, then stop looking. A // vectorized variant of a function expects input in a certain shape // -- basically the number of input registers, the number of lanes // per register, and whether there's a mask required. // We store a pointer to the variant in the VPWidenCallRecipe, so // once we have an appropriate variant it's only valid for that VF. // This will force a different vplan to be generated for each VF that // finds a valid variant. if (Variant) return false; LoopVectorizationCostModel::CallWideningDecision Decision = CM.getCallWideningDecision(CI, VF); if (Decision.Kind == LoopVectorizationCostModel::CM_VectorCall) { Variant = Decision.Variant; MaskPos = Decision.MaskPos; return true; } return false; }, Range); if (ShouldUseVectorCall) { if (MaskPos.has_value()) { // We have 2 cases that would require a mask: // 1) The block needs to be predicated, either due to a conditional // in the scalar loop or use of an active lane mask with // tail-folding, and we use the appropriate mask for the block. // 2) No mask is required for the block, but the only available // vector variant at this VF requires a mask, so we synthesize an // all-true mask. VPValue *Mask = nullptr; if (Legal->isMaskRequired(CI)) Mask = getBlockInMask(Builder.getInsertBlock()); else Mask = Plan.getOrAddLiveIn( ConstantInt::getTrue(IntegerType::getInt1Ty(CI->getContext()))); Ops.insert(Ops.begin() + *MaskPos, Mask); } Ops.push_back(Operands.back()); return new VPWidenCallRecipe(CI, Variant, Ops, CI->getDebugLoc()); } return nullptr; } bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const { assert(!isa(I) && !isa(I) && !isa(I) && !isa(I) && "Instruction should have been handled earlier"); // Instruction should be widened, unless it is scalar after vectorization, // scalarization is profitable or it is predicated. auto WillScalarize = [this, I](ElementCount VF) -> bool { return CM.isScalarAfterVectorization(I, VF) || CM.isProfitableToScalarize(I, VF) || CM.isScalarWithPredication(I, VF); }; return !LoopVectorizationPlanner::getDecisionAndClampRange(WillScalarize, Range); } VPWidenRecipe *VPRecipeBuilder::tryToWiden(Instruction *I, ArrayRef Operands) { switch (I->getOpcode()) { default: return nullptr; case Instruction::SDiv: case Instruction::UDiv: case Instruction::SRem: case Instruction::URem: { // If not provably safe, use a select to form a safe divisor before widening the // div/rem operation itself. Otherwise fall through to general handling below. if (CM.isPredicatedInst(I)) { SmallVector Ops(Operands); VPValue *Mask = getBlockInMask(Builder.getInsertBlock()); VPValue *One = Plan.getOrAddLiveIn(ConstantInt::get(I->getType(), 1u, false)); auto *SafeRHS = Builder.createSelect(Mask, Ops[1], One, I->getDebugLoc()); Ops[1] = SafeRHS; return new VPWidenRecipe(*I, Ops); } [[fallthrough]]; } case Instruction::Add: case Instruction::And: case Instruction::AShr: case Instruction::FAdd: case Instruction::FCmp: case Instruction::FDiv: case Instruction::FMul: case Instruction::FNeg: case Instruction::FRem: case Instruction::FSub: case Instruction::ICmp: case Instruction::LShr: case Instruction::Mul: case Instruction::Or: case Instruction::Select: case Instruction::Shl: case Instruction::Sub: case Instruction::Xor: case Instruction::Freeze: { SmallVector NewOps(Operands); if (Instruction::isBinaryOp(I->getOpcode())) { // The legacy cost model uses SCEV to check if some of the operands are // constants. To match the legacy cost model's behavior, use SCEV to try // to replace operands with constants. ScalarEvolution &SE = *PSE.getSE(); auto GetConstantViaSCEV = [this, &SE](VPValue *Op) { if (!Op->isLiveIn()) return Op; Value *V = Op->getUnderlyingValue(); if (isa(V) || !SE.isSCEVable(V->getType())) return Op; auto *C = dyn_cast(SE.getSCEV(V)); if (!C) return Op; return Plan.getOrAddLiveIn(C->getValue()); }; // For Mul, the legacy cost model checks both operands. if (I->getOpcode() == Instruction::Mul) NewOps[0] = GetConstantViaSCEV(NewOps[0]); // For other binops, the legacy cost model only checks the second operand. NewOps[1] = GetConstantViaSCEV(NewOps[1]); } return new VPWidenRecipe(*I, NewOps); } case Instruction::ExtractValue: { SmallVector NewOps(Operands); Type *I32Ty = IntegerType::getInt32Ty(I->getContext()); auto *EVI = cast(I); assert(EVI->getNumIndices() == 1 && "Expected one extractvalue index"); unsigned Idx = EVI->getIndices()[0]; NewOps.push_back(Plan.getOrAddLiveIn(ConstantInt::get(I32Ty, Idx, false))); return new VPWidenRecipe(*I, NewOps); } }; } VPHistogramRecipe * VPRecipeBuilder::tryToWidenHistogram(const HistogramInfo *HI, ArrayRef Operands) { // FIXME: Support other operations. unsigned Opcode = HI->Update->getOpcode(); assert((Opcode == Instruction::Add || Opcode == Instruction::Sub) && "Histogram update operation must be an Add or Sub"); SmallVector HGramOps; // Bucket address. HGramOps.push_back(Operands[1]); // Increment value. HGramOps.push_back(getVPValueOrAddLiveIn(HI->Update->getOperand(1))); // In case of predicated execution (due to tail-folding, or conditional // execution, or both), pass the relevant mask. if (Legal->isMaskRequired(HI->Store)) HGramOps.push_back(getBlockInMask(Builder.getInsertBlock())); return new VPHistogramRecipe(Opcode, HGramOps, HI->Store->getDebugLoc()); } VPReplicateRecipe * VPRecipeBuilder::handleReplication(Instruction *I, ArrayRef Operands, VFRange &Range) { bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange( [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); }, Range); bool IsPredicated = CM.isPredicatedInst(I); // Even if the instruction is not marked as uniform, there are certain // intrinsic calls that can be effectively treated as such, so we check for // them here. Conservatively, we only do this for scalable vectors, since // for fixed-width VFs we can always fall back on full scalarization. if (!IsUniform && Range.Start.isScalable() && isa(I)) { switch (cast(I)->getIntrinsicID()) { case Intrinsic::assume: case Intrinsic::lifetime_start: case Intrinsic::lifetime_end: // For scalable vectors if one of the operands is variant then we still // want to mark as uniform, which will generate one instruction for just // the first lane of the vector. We can't scalarize the call in the same // way as for fixed-width vectors because we don't know how many lanes // there are. // // The reasons for doing it this way for scalable vectors are: // 1. For the assume intrinsic generating the instruction for the first // lane is still be better than not generating any at all. For // example, the input may be a splat across all lanes. // 2. For the lifetime start/end intrinsics the pointer operand only // does anything useful when the input comes from a stack object, // which suggests it should always be uniform. For non-stack objects // the effect is to poison the object, which still allows us to // remove the call. IsUniform = true; break; default: break; } } VPValue *BlockInMask = nullptr; if (!IsPredicated) { // Finalize the recipe for Instr, first if it is not predicated. LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n"); } else { LLVM_DEBUG(dbgs() << "LV: Scalarizing and predicating:" << *I << "\n"); // Instructions marked for predication are replicated and a mask operand is // added initially. Masked replicate recipes will later be placed under an // if-then construct to prevent side-effects. Generate recipes to compute // the block mask for this region. BlockInMask = getBlockInMask(Builder.getInsertBlock()); } // Note that there is some custom logic to mark some intrinsics as uniform // manually above for scalable vectors, which this assert needs to account for // as well. assert((Range.Start.isScalar() || !IsUniform || !IsPredicated || (Range.Start.isScalable() && isa(I))) && "Should not predicate a uniform recipe"); auto *Recipe = new VPReplicateRecipe(I, Operands, IsUniform, BlockInMask, VPIRMetadata(*I, LVer)); return Recipe; } /// Find all possible partial reductions in the loop and track all of those that /// are valid so recipes can be formed later. void VPRecipeBuilder::collectScaledReductions(VFRange &Range) { // Find all possible partial reductions. SmallVector> PartialReductionChains; for (const auto &[Phi, RdxDesc] : Legal->getReductionVars()) { getScaledReductions(Phi, RdxDesc.getLoopExitInstr(), Range, PartialReductionChains); } // A partial reduction is invalid if any of its extends are used by // something that isn't another partial reduction. This is because the // extends are intended to be lowered along with the reduction itself. // Build up a set of partial reduction ops for efficient use checking. SmallSet PartialReductionOps; for (const auto &[PartialRdx, _] : PartialReductionChains) PartialReductionOps.insert(PartialRdx.ExtendUser); auto ExtendIsOnlyUsedByPartialReductions = [&PartialReductionOps](Instruction *Extend) { return all_of(Extend->users(), [&](const User *U) { return PartialReductionOps.contains(U); }); }; // Check if each use of a chain's two extends is a partial reduction // and only add those that don't have non-partial reduction users. for (auto Pair : PartialReductionChains) { PartialReductionChain Chain = Pair.first; if (ExtendIsOnlyUsedByPartialReductions(Chain.ExtendA) && (!Chain.ExtendB || ExtendIsOnlyUsedByPartialReductions(Chain.ExtendB))) ScaledReductionMap.try_emplace(Chain.Reduction, Pair.second); } } bool VPRecipeBuilder::getScaledReductions( Instruction *PHI, Instruction *RdxExitInstr, VFRange &Range, SmallVectorImpl> &Chains) { if (!CM.TheLoop->contains(RdxExitInstr)) return false; auto *Update = dyn_cast(RdxExitInstr); if (!Update) return false; Value *Op = Update->getOperand(0); Value *PhiOp = Update->getOperand(1); if (Op == PHI) std::swap(Op, PhiOp); // Try and get a scaled reduction from the first non-phi operand. // If one is found, we use the discovered reduction instruction in // place of the accumulator for costing. if (auto *OpInst = dyn_cast(Op)) { if (getScaledReductions(PHI, OpInst, Range, Chains)) { PHI = Chains.rbegin()->first.Reduction; Op = Update->getOperand(0); PhiOp = Update->getOperand(1); if (Op == PHI) std::swap(Op, PhiOp); } } if (PhiOp != PHI) return false; using namespace llvm::PatternMatch; // If the update is a binary operator, check both of its operands to see if // they are extends. Otherwise, see if the update comes directly from an // extend. Instruction *Exts[2] = {nullptr}; BinaryOperator *ExtendUser = dyn_cast(Op); std::optional BinOpc; Type *ExtOpTypes[2] = {nullptr}; auto CollectExtInfo = [this, &Exts, &ExtOpTypes](SmallVectorImpl &Ops) -> bool { unsigned I = 0; for (Value *OpI : Ops) { Value *ExtOp; if (!match(OpI, m_ZExtOrSExt(m_Value(ExtOp)))) return false; Exts[I] = cast(OpI); // TODO: We should be able to support live-ins. if (!CM.TheLoop->contains(Exts[I])) return false; ExtOpTypes[I] = ExtOp->getType(); I++; } return true; }; if (ExtendUser) { if (!ExtendUser->hasOneUse()) return false; // Use the side-effect of match to replace BinOp only if the pattern is // matched, we don't care at this point whether it actually matched. match(ExtendUser, m_Neg(m_BinOp(ExtendUser))); SmallVector Ops(ExtendUser->operands()); if (!CollectExtInfo(Ops)) return false; BinOpc = std::make_optional(ExtendUser->getOpcode()); } else if (match(Update, m_Add(m_Value(), m_Value()))) { // We already know the operands for Update are Op and PhiOp. SmallVector Ops({Op}); if (!CollectExtInfo(Ops)) return false; ExtendUser = Update; BinOpc = std::nullopt; } else return false; TTI::PartialReductionExtendKind OpAExtend = TTI::getPartialReductionExtendKind(Exts[0]); TTI::PartialReductionExtendKind OpBExtend = Exts[1] ? TTI::getPartialReductionExtendKind(Exts[1]) : TTI::PR_None; PartialReductionChain Chain(RdxExitInstr, Exts[0], Exts[1], ExtendUser); TypeSize PHISize = PHI->getType()->getPrimitiveSizeInBits(); TypeSize ASize = ExtOpTypes[0]->getPrimitiveSizeInBits(); if (!PHISize.hasKnownScalarFactor(ASize)) return false; unsigned TargetScaleFactor = PHISize.getKnownScalarFactor(ASize); if (LoopVectorizationPlanner::getDecisionAndClampRange( [&](ElementCount VF) { InstructionCost Cost = TTI->getPartialReductionCost( Update->getOpcode(), ExtOpTypes[0], ExtOpTypes[1], PHI->getType(), VF, OpAExtend, OpBExtend, BinOpc, CM.CostKind); return Cost.isValid(); }, Range)) { Chains.emplace_back(Chain, TargetScaleFactor); return true; } return false; } VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(VPSingleDefRecipe *R, VFRange &Range) { // First, check for specific widening recipes that deal with inductions, Phi // nodes, calls and memory operations. VPRecipeBase *Recipe; Instruction *Instr = R->getUnderlyingInstr(); SmallVector Operands(R->operands()); if (auto *PhiR = dyn_cast(R)) { VPBasicBlock *Parent = PhiR->getParent(); [[maybe_unused]] VPRegionBlock *LoopRegionOf = Parent->getEnclosingLoopRegion(); assert(LoopRegionOf && LoopRegionOf->getEntry() == Parent && "Non-header phis should have been handled during predication"); auto *Phi = cast(R->getUnderlyingInstr()); assert(Operands.size() == 2 && "Must have 2 operands for header phis"); if ((Recipe = tryToOptimizeInductionPHI(Phi, Operands, Range))) return Recipe; VPHeaderPHIRecipe *PhiRecipe = nullptr; assert((Legal->isReductionVariable(Phi) || Legal->isFixedOrderRecurrence(Phi)) && "can only widen reductions and fixed-order recurrences here"); VPValue *StartV = Operands[0]; if (Legal->isReductionVariable(Phi)) { const RecurrenceDescriptor &RdxDesc = Legal->getRecurrenceDescriptor(Phi); assert(RdxDesc.getRecurrenceStartValue() == Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader())); // If the PHI is used by a partial reduction, set the scale factor. unsigned ScaleFactor = getScalingForReduction(RdxDesc.getLoopExitInstr()).value_or(1); PhiRecipe = new VPReductionPHIRecipe( Phi, RdxDesc.getRecurrenceKind(), *StartV, CM.isInLoopReduction(Phi), CM.useOrderedReductions(RdxDesc), ScaleFactor); } else { // TODO: Currently fixed-order recurrences are modeled as chains of // first-order recurrences. If there are no users of the intermediate // recurrences in the chain, the fixed order recurrence should be modeled // directly, enabling more efficient codegen. PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV); } // Add backedge value. PhiRecipe->addOperand(Operands[1]); return PhiRecipe; } assert(!R->isPhi() && "only VPPhi nodes expected at this point"); if (isa(Instr) && (Recipe = tryToOptimizeInductionTruncate( cast(Instr), Operands, Range))) return Recipe; // All widen recipes below deal only with VF > 1. if (LoopVectorizationPlanner::getDecisionAndClampRange( [&](ElementCount VF) { return VF.isScalar(); }, Range)) return nullptr; if (auto *CI = dyn_cast(Instr)) return tryToWidenCall(CI, Operands, Range); if (StoreInst *SI = dyn_cast(Instr)) if (auto HistInfo = Legal->getHistogramInfo(SI)) return tryToWidenHistogram(*HistInfo, Operands); if (isa(Instr) || isa(Instr)) return tryToWidenMemory(Instr, Operands, Range); if (std::optional ScaleFactor = getScalingForReduction(Instr)) return tryToCreatePartialReduction(Instr, Operands, ScaleFactor.value()); if (!shouldWiden(Instr, Range)) return nullptr; if (auto *GEP = dyn_cast(Instr)) return new VPWidenGEPRecipe(GEP, Operands); if (auto *SI = dyn_cast(Instr)) { return new VPWidenSelectRecipe(*SI, Operands); } if (auto *CI = dyn_cast(Instr)) { return new VPWidenCastRecipe(CI->getOpcode(), Operands[0], CI->getType(), *CI); } return tryToWiden(Instr, Operands); } VPRecipeBase * VPRecipeBuilder::tryToCreatePartialReduction(Instruction *Reduction, ArrayRef Operands, unsigned ScaleFactor) { assert(Operands.size() == 2 && "Unexpected number of operands for partial reduction"); VPValue *BinOp = Operands[0]; VPValue *Accumulator = Operands[1]; VPRecipeBase *BinOpRecipe = BinOp->getDefiningRecipe(); if (isa(BinOpRecipe) || isa(BinOpRecipe)) std::swap(BinOp, Accumulator); unsigned ReductionOpcode = Reduction->getOpcode(); if (ReductionOpcode == Instruction::Sub) { auto *const Zero = ConstantInt::get(Reduction->getType(), 0); SmallVector Ops; Ops.push_back(Plan.getOrAddLiveIn(Zero)); Ops.push_back(BinOp); BinOp = new VPWidenRecipe(*Reduction, Ops); Builder.insert(BinOp->getDefiningRecipe()); ReductionOpcode = Instruction::Add; } VPValue *Cond = nullptr; if (CM.blockNeedsPredicationForAnyReason(Reduction->getParent())) { assert((ReductionOpcode == Instruction::Add || ReductionOpcode == Instruction::Sub) && "Expected an ADD or SUB operation for predicated partial " "reductions (because the neutral element in the mask is zero)!"); Cond = getBlockInMask(Builder.getInsertBlock()); VPValue *Zero = Plan.getOrAddLiveIn(ConstantInt::get(Reduction->getType(), 0)); BinOp = Builder.createSelect(Cond, BinOp, Zero, Reduction->getDebugLoc()); } return new VPPartialReductionRecipe(ReductionOpcode, Accumulator, BinOp, Cond, ScaleFactor, Reduction); } void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF, ElementCount MaxVF) { if (ElementCount::isKnownGT(MinVF, MaxVF)) return; assert(OrigLoop->isInnermost() && "Inner loop expected."); const LoopAccessInfo *LAI = Legal->getLAI(); LoopVersioning LVer(*LAI, LAI->getRuntimePointerChecking()->getChecks(), OrigLoop, LI, DT, PSE.getSE()); if (!LAI->getRuntimePointerChecking()->getChecks().empty() && !LAI->getRuntimePointerChecking()->getDiffChecks()) { // Only use noalias metadata when using memory checks guaranteeing no // overlap across all iterations. LVer.prepareNoAliasMetadata(); } // Create initial base VPlan0, to serve as common starting point for all // candidates built later for specific VF ranges. auto VPlan0 = VPlanTransforms::buildVPlan0( OrigLoop, *LI, Legal->getWidestInductionType(), getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()), PSE); auto MaxVFTimes2 = MaxVF * 2; for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFTimes2);) { VFRange SubRange = {VF, MaxVFTimes2}; if (auto Plan = tryToBuildVPlanWithVPRecipes( std::unique_ptr(VPlan0->duplicate()), SubRange, &LVer)) { bool HasScalarVF = Plan->hasScalarVFOnly(); // Now optimize the initial VPlan. if (!HasScalarVF) VPlanTransforms::runPass(VPlanTransforms::truncateToMinimalBitwidths, *Plan, CM.getMinimalBitwidths()); VPlanTransforms::runPass(VPlanTransforms::optimize, *Plan); // TODO: try to put it close to addActiveLaneMask(). if (CM.foldTailWithEVL() && !HasScalarVF) VPlanTransforms::runPass(VPlanTransforms::addExplicitVectorLength, *Plan, CM.getMaxSafeElements()); assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid"); VPlans.push_back(std::move(Plan)); } VF = SubRange.End; } } /// Create and return a ResumePhi for \p WideIV, unless it is truncated. If the /// induction recipe is not canonical, creates a VPDerivedIVRecipe to compute /// the end value of the induction. static VPInstruction *addResumePhiRecipeForInduction( VPWidenInductionRecipe *WideIV, VPBuilder &VectorPHBuilder, VPBuilder &ScalarPHBuilder, VPTypeAnalysis &TypeInfo, VPValue *VectorTC) { auto *WideIntOrFp = dyn_cast(WideIV); // Truncated wide inductions resume from the last lane of their vector value // in the last vector iteration which is handled elsewhere. if (WideIntOrFp && WideIntOrFp->getTruncInst()) return nullptr; VPValue *Start = WideIV->getStartValue(); VPValue *Step = WideIV->getStepValue(); const InductionDescriptor &ID = WideIV->getInductionDescriptor(); VPValue *EndValue = VectorTC; if (!WideIntOrFp || !WideIntOrFp->isCanonical()) { EndValue = VectorPHBuilder.createDerivedIV( ID.getKind(), dyn_cast_or_null(ID.getInductionBinOp()), Start, VectorTC, Step); } // EndValue is derived from the vector trip count (which has the same type as // the widest induction) and thus may be wider than the induction here. Type *ScalarTypeOfWideIV = TypeInfo.inferScalarType(WideIV); if (ScalarTypeOfWideIV != TypeInfo.inferScalarType(EndValue)) { EndValue = VectorPHBuilder.createScalarCast(Instruction::Trunc, EndValue, ScalarTypeOfWideIV, WideIV->getDebugLoc()); } auto *ResumePhiRecipe = ScalarPHBuilder.createScalarPhi( {EndValue, Start}, WideIV->getDebugLoc(), "bc.resume.val"); return ResumePhiRecipe; } /// Create resume phis in the scalar preheader for first-order recurrences, /// reductions and inductions, and update the VPIRInstructions wrapping the /// original phis in the scalar header. End values for inductions are added to /// \p IVEndValues. static void addScalarResumePhis(VPRecipeBuilder &Builder, VPlan &Plan, DenseMap &IVEndValues) { VPTypeAnalysis TypeInfo(Plan); auto *ScalarPH = Plan.getScalarPreheader(); auto *MiddleVPBB = cast(ScalarPH->getPredecessors()[0]); VPRegionBlock *VectorRegion = Plan.getVectorLoopRegion(); VPBuilder VectorPHBuilder( cast(VectorRegion->getSinglePredecessor())); VPBuilder MiddleBuilder(MiddleVPBB, MiddleVPBB->getFirstNonPhi()); VPBuilder ScalarPHBuilder(ScalarPH); for (VPRecipeBase &ScalarPhiR : Plan.getScalarHeader()->phis()) { auto *ScalarPhiIRI = cast(&ScalarPhiR); // TODO: Extract final value from induction recipe initially, optimize to // pre-computed end value together in optimizeInductionExitUsers. auto *VectorPhiR = cast(Builder.getRecipe(&ScalarPhiIRI->getIRPhi())); if (auto *WideIVR = dyn_cast(VectorPhiR)) { if (VPInstruction *ResumePhi = addResumePhiRecipeForInduction( WideIVR, VectorPHBuilder, ScalarPHBuilder, TypeInfo, &Plan.getVectorTripCount())) { assert(isa(ResumePhi) && "Expected a phi"); IVEndValues[WideIVR] = ResumePhi->getOperand(0); ScalarPhiIRI->addOperand(ResumePhi); continue; } // TODO: Also handle truncated inductions here. Computing end-values // separately should be done as VPlan-to-VPlan optimization, after // legalizing all resume values to use the last lane from the loop. assert(cast(VectorPhiR)->getTruncInst() && "should only skip truncated wide inductions"); continue; } // The backedge value provides the value to resume coming out of a loop, // which for FORs is a vector whose last element needs to be extracted. The // start value provides the value if the loop is bypassed. bool IsFOR = isa(VectorPhiR); auto *ResumeFromVectorLoop = VectorPhiR->getBackedgeValue(); assert(VectorRegion->getSingleSuccessor() == Plan.getMiddleBlock() && "Cannot handle loops with uncountable early exits"); if (IsFOR) ResumeFromVectorLoop = MiddleBuilder.createNaryOp( VPInstruction::ExtractLastElement, {ResumeFromVectorLoop}, {}, "vector.recur.extract"); StringRef Name = IsFOR ? "scalar.recur.init" : "bc.merge.rdx"; auto *ResumePhiR = ScalarPHBuilder.createScalarPhi( {ResumeFromVectorLoop, VectorPhiR->getStartValue()}, {}, Name); ScalarPhiIRI->addOperand(ResumePhiR); } } /// Handle users in the exit block for first order reductions in the original /// exit block. The penultimate value of recurrences is fed to their LCSSA phi /// users in the original exit block using the VPIRInstruction wrapping to the /// LCSSA phi. static void addExitUsersForFirstOrderRecurrences(VPlan &Plan, VFRange &Range) { VPRegionBlock *VectorRegion = Plan.getVectorLoopRegion(); auto *ScalarPHVPBB = Plan.getScalarPreheader(); auto *MiddleVPBB = Plan.getMiddleBlock(); VPBuilder ScalarPHBuilder(ScalarPHVPBB); VPBuilder MiddleBuilder(MiddleVPBB, MiddleVPBB->getFirstNonPhi()); auto IsScalableOne = [](ElementCount VF) -> bool { return VF == ElementCount::getScalable(1); }; for (auto &HeaderPhi : VectorRegion->getEntryBasicBlock()->phis()) { auto *FOR = dyn_cast(&HeaderPhi); if (!FOR) continue; assert(VectorRegion->getSingleSuccessor() == Plan.getMiddleBlock() && "Cannot handle loops with uncountable early exits"); // This is the second phase of vectorizing first-order recurrences, creating // extract for users outside the loop. An overview of the transformation is // described below. Suppose we have the following loop with some use after // the loop of the last a[i-1], // // for (int i = 0; i < n; ++i) { // t = a[i - 1]; // b[i] = a[i] - t; // } // use t; // // There is a first-order recurrence on "a". For this loop, the shorthand // scalar IR looks like: // // scalar.ph: // s.init = a[-1] // br scalar.body // // scalar.body: // i = phi [0, scalar.ph], [i+1, scalar.body] // s1 = phi [s.init, scalar.ph], [s2, scalar.body] // s2 = a[i] // b[i] = s2 - s1 // br cond, scalar.body, exit.block // // exit.block: // use = lcssa.phi [s1, scalar.body] // // In this example, s1 is a recurrence because it's value depends on the // previous iteration. In the first phase of vectorization, we created a // VPFirstOrderRecurrencePHIRecipe v1 for s1. Now we create the extracts // for users in the scalar preheader and exit block. // // vector.ph: // v_init = vector(..., ..., ..., a[-1]) // br vector.body // // vector.body // i = phi [0, vector.ph], [i+4, vector.body] // v1 = phi [v_init, vector.ph], [v2, vector.body] // v2 = a[i, i+1, i+2, i+3] // b[i] = v2 - v1 // // Next, third phase will introduce v1' = splice(v1(3), v2(0, 1, 2)) // b[i, i+1, i+2, i+3] = v2 - v1 // br cond, vector.body, middle.block // // middle.block: // vector.recur.extract.for.phi = v2(2) // vector.recur.extract = v2(3) // br cond, scalar.ph, exit.block // // scalar.ph: // scalar.recur.init = phi [vector.recur.extract, middle.block], // [s.init, otherwise] // br scalar.body // // scalar.body: // i = phi [0, scalar.ph], [i+1, scalar.body] // s1 = phi [scalar.recur.init, scalar.ph], [s2, scalar.body] // s2 = a[i] // b[i] = s2 - s1 // br cond, scalar.body, exit.block // // exit.block: // lo = lcssa.phi [s1, scalar.body], // [vector.recur.extract.for.phi, middle.block] // // Now update VPIRInstructions modeling LCSSA phis in the exit block. // Extract the penultimate value of the recurrence and use it as operand for // the VPIRInstruction modeling the phi. for (VPUser *U : FOR->users()) { using namespace llvm::VPlanPatternMatch; if (!match(U, m_VPInstruction( m_Specific(FOR)))) continue; // For VF vscale x 1, if vscale = 1, we are unable to extract the // penultimate value of the recurrence. Instead we rely on the existing // extract of the last element from the result of // VPInstruction::FirstOrderRecurrenceSplice. // TODO: Consider vscale_range info and UF. if (LoopVectorizationPlanner::getDecisionAndClampRange(IsScalableOne, Range)) return; VPValue *PenultimateElement = MiddleBuilder.createNaryOp( VPInstruction::ExtractPenultimateElement, {FOR->getBackedgeValue()}, {}, "vector.recur.extract.for.phi"); cast(U)->replaceAllUsesWith(PenultimateElement); } } } VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes( VPlanPtr Plan, VFRange &Range, LoopVersioning *LVer) { using namespace llvm::VPlanPatternMatch; SmallPtrSet *, 1> InterleaveGroups; // --------------------------------------------------------------------------- // Build initial VPlan: Scan the body of the loop in a topological order to // visit each basic block after having visited its predecessor basic blocks. // --------------------------------------------------------------------------- bool RequiresScalarEpilogueCheck = LoopVectorizationPlanner::getDecisionAndClampRange( [this](ElementCount VF) { return !CM.requiresScalarEpilogue(VF.isVector()); }, Range); VPlanTransforms::handleEarlyExits(*Plan, Legal->hasUncountableEarlyExit(), Range); VPlanTransforms::addMiddleCheck(*Plan, RequiresScalarEpilogueCheck, CM.foldTailByMasking()); VPlanTransforms::createLoopRegions(*Plan); VPlanTransforms::createExtractsForLiveOuts(*Plan); // Don't use getDecisionAndClampRange here, because we don't know the UF // so this function is better to be conservative, rather than to split // it up into different VPlans. // TODO: Consider using getDecisionAndClampRange here to split up VPlans. bool IVUpdateMayOverflow = false; for (ElementCount VF : Range) IVUpdateMayOverflow |= !isIndvarOverflowCheckKnownFalse(&CM, VF); TailFoldingStyle Style = CM.getTailFoldingStyle(IVUpdateMayOverflow); // Use NUW for the induction increment if we proved that it won't overflow in // the vector loop or when not folding the tail. In the later case, we know // that the canonical induction increment will not overflow as the vector trip // count is >= increment and a multiple of the increment. bool HasNUW = !IVUpdateMayOverflow || Style == TailFoldingStyle::None; if (!HasNUW) { auto *IVInc = Plan->getVectorLoopRegion() ->getExitingBasicBlock() ->getTerminator() ->getOperand(0); assert(match(IVInc, m_VPInstruction( m_Specific(Plan->getCanonicalIV()), m_VPValue())) && "Did not find the canonical IV increment"); cast(IVInc)->dropPoisonGeneratingFlags(); } // --------------------------------------------------------------------------- // Pre-construction: record ingredients whose recipes we'll need to further // process after constructing the initial VPlan. // --------------------------------------------------------------------------- // For each interleave group which is relevant for this (possibly trimmed) // Range, add it to the set of groups to be later applied to the VPlan and add // placeholders for its members' Recipes which we'll be replacing with a // single VPInterleaveRecipe. for (InterleaveGroup *IG : IAI.getInterleaveGroups()) { auto ApplyIG = [IG, this](ElementCount VF) -> bool { bool Result = (VF.isVector() && // Query is illegal for VF == 1 CM.getWideningDecision(IG->getInsertPos(), VF) == LoopVectorizationCostModel::CM_Interleave); // For scalable vectors, the interleave factors must be <= 8 since we // require the (de)interleaveN intrinsics instead of shufflevectors. assert((!Result || !VF.isScalable() || IG->getFactor() <= 8) && "Unsupported interleave factor for scalable vectors"); return Result; }; if (!getDecisionAndClampRange(ApplyIG, Range)) continue; InterleaveGroups.insert(IG); } // --------------------------------------------------------------------------- // Predicate and linearize the top-level loop region. // --------------------------------------------------------------------------- auto BlockMaskCache = VPlanTransforms::introduceMasksAndLinearize( *Plan, CM.foldTailByMasking()); // --------------------------------------------------------------------------- // Construct wide recipes and apply predication for original scalar // VPInstructions in the loop. // --------------------------------------------------------------------------- VPRecipeBuilder RecipeBuilder(*Plan, OrigLoop, TLI, &TTI, Legal, CM, PSE, Builder, BlockMaskCache, LVer); RecipeBuilder.collectScaledReductions(Range); // Scan the body of the loop in a topological order to visit each basic block // after having visited its predecessor basic blocks. VPRegionBlock *LoopRegion = Plan->getVectorLoopRegion(); VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock(); ReversePostOrderTraversal> RPOT( HeaderVPBB); auto *MiddleVPBB = Plan->getMiddleBlock(); VPBasicBlock::iterator MBIP = MiddleVPBB->getFirstNonPhi(); // Mapping from VPValues in the initial plan to their widened VPValues. Needed // temporarily to update created block masks. DenseMap Old2New; for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly(RPOT)) { // Convert input VPInstructions to widened recipes. for (VPRecipeBase &R : make_early_inc_range(*VPBB)) { auto *SingleDef = cast(&R); auto *UnderlyingValue = SingleDef->getUnderlyingValue(); // Skip recipes that do not need transforming, including canonical IV, // wide canonical IV and VPInstructions without underlying values. The // latter are added above for masking. // FIXME: Migrate code relying on the underlying instruction from VPlan0 // to construct recipes below to not use the underlying instruction. if (isa( &R) || (isa(&R) && !UnderlyingValue)) continue; // FIXME: VPlan0, which models a copy of the original scalar loop, should // not use VPWidenPHIRecipe to model the phis. assert((isa(&R) || isa(&R)) && UnderlyingValue && "unsupported recipe"); // TODO: Gradually replace uses of underlying instruction by analyses on // VPlan. Instruction *Instr = cast(UnderlyingValue); Builder.setInsertPoint(SingleDef); // The stores with invariant address inside the loop will be deleted, and // in the exit block, a uniform store recipe will be created for the final // invariant store of the reduction. StoreInst *SI; if ((SI = dyn_cast(Instr)) && Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) { // Only create recipe for the final invariant store of the reduction. if (Legal->isInvariantStoreOfReduction(SI)) { auto *Recipe = new VPReplicateRecipe(SI, R.operands(), true /* IsUniform */, nullptr /*Mask*/, VPIRMetadata(*SI, LVer)); Recipe->insertBefore(*MiddleVPBB, MBIP); } R.eraseFromParent(); continue; } VPRecipeBase *Recipe = RecipeBuilder.tryToCreateWidenRecipe(SingleDef, Range); if (!Recipe) { SmallVector Operands(R.operands()); Recipe = RecipeBuilder.handleReplication(Instr, Operands, Range); } RecipeBuilder.setRecipe(Instr, Recipe); if (isa(Recipe) && isa(Instr)) { // Optimized a truncate to VPWidenIntOrFpInductionRecipe. It needs to be // moved to the phi section in the header. Recipe->insertBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi()); } else { Builder.insert(Recipe); } if (Recipe->getNumDefinedValues() == 1) { SingleDef->replaceAllUsesWith(Recipe->getVPSingleValue()); Old2New[SingleDef] = Recipe->getVPSingleValue(); } else { assert(Recipe->getNumDefinedValues() == 0 && "Unexpected multidef recipe"); R.eraseFromParent(); } } } // replaceAllUsesWith above may invalidate the block masks. Update them here. // TODO: Include the masks as operands in the predicated VPlan directly // to remove the need to keep a map of masks beyond the predication // transform. RecipeBuilder.updateBlockMaskCache(Old2New); for (const auto &[Old, _] : Old2New) Old->getDefiningRecipe()->eraseFromParent(); assert(isa(Plan->getVectorLoopRegion()) && !Plan->getVectorLoopRegion()->getEntryBasicBlock()->empty() && "entry block must be set to a VPRegionBlock having a non-empty entry " "VPBasicBlock"); // Update wide induction increments to use the same step as the corresponding // wide induction. This enables detecting induction increments directly in // VPlan and removes redundant splats. for (const auto &[Phi, ID] : Legal->getInductionVars()) { auto *IVInc = cast( Phi->getIncomingValueForBlock(OrigLoop->getLoopLatch())); if (IVInc->getOperand(0) != Phi || IVInc->getOpcode() != Instruction::Add) continue; VPWidenInductionRecipe *WideIV = cast(RecipeBuilder.getRecipe(Phi)); VPRecipeBase *R = RecipeBuilder.getRecipe(IVInc); R->setOperand(1, WideIV->getStepValue()); } addExitUsersForFirstOrderRecurrences(*Plan, Range); DenseMap IVEndValues; addScalarResumePhis(RecipeBuilder, *Plan, IVEndValues); // --------------------------------------------------------------------------- // Transform initial VPlan: Apply previously taken decisions, in order, to // bring the VPlan to its final state. // --------------------------------------------------------------------------- // Adjust the recipes for any inloop reductions. adjustRecipesForReductions(Plan, RecipeBuilder, Range.Start); // Apply mandatory transformation to handle FP maxnum/minnum reduction with // NaNs if possible, bail out otherwise. if (!VPlanTransforms::runPass(VPlanTransforms::handleMaxMinNumReductions, *Plan)) return nullptr; // Transform recipes to abstract recipes if it is legal and beneficial and // clamp the range for better cost estimation. // TODO: Enable following transform when the EVL-version of extended-reduction // and mulacc-reduction are implemented. if (!CM.foldTailWithEVL()) { VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind); VPlanTransforms::runPass(VPlanTransforms::convertToAbstractRecipes, *Plan, CostCtx, Range); } for (ElementCount VF : Range) Plan->addVF(VF); Plan->setName("Initial VPlan"); // Interleave memory: for each Interleave Group we marked earlier as relevant // for this VPlan, replace the Recipes widening its memory instructions with a // single VPInterleaveRecipe at its insertion point. VPlanTransforms::runPass(VPlanTransforms::createInterleaveGroups, *Plan, InterleaveGroups, RecipeBuilder, CM.isScalarEpilogueAllowed()); // Replace VPValues for known constant strides guaranteed by predicate scalar // evolution. auto CanUseVersionedStride = [&Plan](VPUser &U, unsigned) { auto *R = cast(&U); return R->getParent()->getParent() || R->getParent() == Plan->getVectorLoopRegion()->getSinglePredecessor(); }; for (auto [_, Stride] : Legal->getLAI()->getSymbolicStrides()) { auto *StrideV = cast(Stride)->getValue(); auto *ScevStride = dyn_cast(PSE.getSCEV(StrideV)); // Only handle constant strides for now. if (!ScevStride) continue; auto *CI = Plan->getOrAddLiveIn( ConstantInt::get(Stride->getType(), ScevStride->getAPInt())); if (VPValue *StrideVPV = Plan->getLiveIn(StrideV)) StrideVPV->replaceUsesWithIf(CI, CanUseVersionedStride); // The versioned value may not be used in the loop directly but through a // sext/zext. Add new live-ins in those cases. for (Value *U : StrideV->users()) { if (!isa(U)) continue; VPValue *StrideVPV = Plan->getLiveIn(U); if (!StrideVPV) continue; unsigned BW = U->getType()->getScalarSizeInBits(); APInt C = isa(U) ? ScevStride->getAPInt().sext(BW) : ScevStride->getAPInt().zext(BW); VPValue *CI = Plan->getOrAddLiveIn(ConstantInt::get(U->getType(), C)); StrideVPV->replaceUsesWithIf(CI, CanUseVersionedStride); } } auto BlockNeedsPredication = [this](BasicBlock *BB) { return Legal->blockNeedsPredication(BB); }; VPlanTransforms::runPass(VPlanTransforms::dropPoisonGeneratingRecipes, *Plan, BlockNeedsPredication); // Sink users of fixed-order recurrence past the recipe defining the previous // value and introduce FirstOrderRecurrenceSplice VPInstructions. if (!VPlanTransforms::runPass(VPlanTransforms::adjustFixedOrderRecurrences, *Plan, Builder)) return nullptr; if (useActiveLaneMask(Style)) { // TODO: Move checks to VPlanTransforms::addActiveLaneMask once // TailFoldingStyle is visible there. bool ForControlFlow = useActiveLaneMaskForControlFlow(Style); bool WithoutRuntimeCheck = Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck; VPlanTransforms::addActiveLaneMask(*Plan, ForControlFlow, WithoutRuntimeCheck); } VPlanTransforms::optimizeInductionExitUsers(*Plan, IVEndValues); assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid"); return Plan; } VPlanPtr LoopVectorizationPlanner::tryToBuildVPlan(VFRange &Range) { // Outer loop handling: They may require CFG and instruction level // transformations before even evaluating whether vectorization is profitable. // Since we cannot modify the incoming IR, we need to build VPlan upfront in // the vectorization pipeline. assert(!OrigLoop->isInnermost()); assert(EnableVPlanNativePath && "VPlan-native path is not enabled."); auto Plan = VPlanTransforms::buildVPlan0( OrigLoop, *LI, Legal->getWidestInductionType(), getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()), PSE); VPlanTransforms::handleEarlyExits(*Plan, /*HasUncountableExit*/ false, Range); VPlanTransforms::addMiddleCheck(*Plan, /*RequiresScalarEpilogue*/ true, /*TailFolded*/ false); VPlanTransforms::createLoopRegions(*Plan); for (ElementCount VF : Range) Plan->addVF(VF); if (!VPlanTransforms::tryToConvertVPInstructionsToVPRecipes( Plan, [this](PHINode *P) { return Legal->getIntOrFpInductionDescriptor(P); }, *PSE.getSE(), *TLI)) return nullptr; // Collect mapping of IR header phis to header phi recipes, to be used in // addScalarResumePhis. DenseMap BlockMaskCache; VPRecipeBuilder RecipeBuilder(*Plan, OrigLoop, TLI, &TTI, Legal, CM, PSE, Builder, BlockMaskCache, nullptr /*LVer*/); for (auto &R : Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { if (isa(&R)) continue; auto *HeaderR = cast(&R); RecipeBuilder.setRecipe(HeaderR->getUnderlyingInstr(), HeaderR); } DenseMap IVEndValues; // TODO: IVEndValues are not used yet in the native path, to optimize exit // values. addScalarResumePhis(RecipeBuilder, *Plan, IVEndValues); assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid"); return Plan; } // Adjust the recipes for reductions. For in-loop reductions the chain of // instructions leading from the loop exit instr to the phi need to be converted // to reductions, with one operand being vector and the other being the scalar // reduction chain. For other reductions, a select is introduced between the phi // and users outside the vector region when folding the tail. // // A ComputeReductionResult recipe is added to the middle block, also for // in-loop reductions which compute their result in-loop, because generating // the subsequent bc.merge.rdx phi is driven by ComputeReductionResult recipes. // // Adjust AnyOf reductions; replace the reduction phi for the selected value // with a boolean reduction phi node to check if the condition is true in any // iteration. The final value is selected by the final ComputeReductionResult. void LoopVectorizationPlanner::adjustRecipesForReductions( VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder, ElementCount MinVF) { using namespace VPlanPatternMatch; VPRegionBlock *VectorLoopRegion = Plan->getVectorLoopRegion(); VPBasicBlock *Header = VectorLoopRegion->getEntryBasicBlock(); VPBasicBlock *MiddleVPBB = Plan->getMiddleBlock(); SmallVector ToDelete; for (VPRecipeBase &R : Header->phis()) { auto *PhiR = dyn_cast(&R); if (!PhiR || !PhiR->isInLoop() || (MinVF.isScalar() && !PhiR->isOrdered())) continue; RecurKind Kind = PhiR->getRecurrenceKind(); assert( !RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) && !RecurrenceDescriptor::isFindIVRecurrenceKind(Kind) && "AnyOf and FindIV reductions are not allowed for in-loop reductions"); // Collect the chain of "link" recipes for the reduction starting at PhiR. SetVector Worklist; Worklist.insert(PhiR); for (unsigned I = 0; I != Worklist.size(); ++I) { VPSingleDefRecipe *Cur = Worklist[I]; for (VPUser *U : Cur->users()) { auto *UserRecipe = cast(U); if (!UserRecipe->getParent()->getEnclosingLoopRegion()) { assert((UserRecipe->getParent() == MiddleVPBB || UserRecipe->getParent() == Plan->getScalarPreheader()) && "U must be either in the loop region, the middle block or the " "scalar preheader."); continue; } Worklist.insert(UserRecipe); } } // Visit operation "Links" along the reduction chain top-down starting from // the phi until LoopExitValue. We keep track of the previous item // (PreviousLink) to tell which of the two operands of a Link will remain // scalar and which will be reduced. For minmax by select(cmp), Link will be // the select instructions. Blend recipes of in-loop reduction phi's will // get folded to their non-phi operand, as the reduction recipe handles the // condition directly. VPSingleDefRecipe *PreviousLink = PhiR; // Aka Worklist[0]. for (VPSingleDefRecipe *CurrentLink : drop_begin(Worklist)) { if (auto *Blend = dyn_cast(CurrentLink)) { assert(Blend->getNumIncomingValues() == 2 && "Blend must have 2 incoming values"); if (Blend->getIncomingValue(0) == PhiR) { Blend->replaceAllUsesWith(Blend->getIncomingValue(1)); } else { assert(Blend->getIncomingValue(1) == PhiR && "PhiR must be an operand of the blend"); Blend->replaceAllUsesWith(Blend->getIncomingValue(0)); } continue; } Instruction *CurrentLinkI = CurrentLink->getUnderlyingInstr(); // Index of the first operand which holds a non-mask vector operand. unsigned IndexOfFirstOperand; // Recognize a call to the llvm.fmuladd intrinsic. bool IsFMulAdd = (Kind == RecurKind::FMulAdd); VPValue *VecOp; VPBasicBlock *LinkVPBB = CurrentLink->getParent(); if (IsFMulAdd) { assert( RecurrenceDescriptor::isFMulAddIntrinsic(CurrentLinkI) && "Expected instruction to be a call to the llvm.fmuladd intrinsic"); assert(((MinVF.isScalar() && isa(CurrentLink)) || isa(CurrentLink)) && CurrentLink->getOperand(2) == PreviousLink && "expected a call where the previous link is the added operand"); // If the instruction is a call to the llvm.fmuladd intrinsic then we // need to create an fmul recipe (multiplying the first two operands of // the fmuladd together) to use as the vector operand for the fadd // reduction. VPInstruction *FMulRecipe = new VPInstruction( Instruction::FMul, {CurrentLink->getOperand(0), CurrentLink->getOperand(1)}, CurrentLinkI->getFastMathFlags()); LinkVPBB->insert(FMulRecipe, CurrentLink->getIterator()); VecOp = FMulRecipe; } else if (PhiR->isInLoop() && Kind == RecurKind::AddChainWithSubs && CurrentLinkI->getOpcode() == Instruction::Sub) { Type *PhiTy = PhiR->getUnderlyingValue()->getType(); auto *Zero = Plan->getOrAddLiveIn(ConstantInt::get(PhiTy, 0)); VPWidenRecipe *Sub = new VPWidenRecipe( Instruction::Sub, {Zero, CurrentLink->getOperand(1)}, {}, VPIRMetadata(), CurrentLinkI->getDebugLoc()); Sub->setUnderlyingValue(CurrentLinkI); LinkVPBB->insert(Sub, CurrentLink->getIterator()); VecOp = Sub; } else { if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) { if (isa(CurrentLink)) { assert(isa(CurrentLinkI) && "need to have the compare of the select"); continue; } assert(isa(CurrentLink) && "must be a select recipe"); IndexOfFirstOperand = 1; } else { assert((MinVF.isScalar() || isa(CurrentLink)) && "Expected to replace a VPWidenSC"); IndexOfFirstOperand = 0; } // Note that for non-commutable operands (cmp-selects), the semantics of // the cmp-select are captured in the recurrence kind. unsigned VecOpId = CurrentLink->getOperand(IndexOfFirstOperand) == PreviousLink ? IndexOfFirstOperand + 1 : IndexOfFirstOperand; VecOp = CurrentLink->getOperand(VecOpId); assert(VecOp != PreviousLink && CurrentLink->getOperand(CurrentLink->getNumOperands() - 1 - (VecOpId - IndexOfFirstOperand)) == PreviousLink && "PreviousLink must be the operand other than VecOp"); } VPValue *CondOp = nullptr; if (CM.blockNeedsPredicationForAnyReason(CurrentLinkI->getParent())) CondOp = RecipeBuilder.getBlockInMask(CurrentLink->getParent()); // TODO: Retrieve FMFs from recipes directly. RecurrenceDescriptor RdxDesc = Legal->getRecurrenceDescriptor( cast(PhiR->getUnderlyingInstr())); // Non-FP RdxDescs will have all fast math flags set, so clear them. FastMathFlags FMFs = isa(CurrentLinkI) ? RdxDesc.getFastMathFlags() : FastMathFlags(); auto *RedRecipe = new VPReductionRecipe( Kind, FMFs, CurrentLinkI, PreviousLink, VecOp, CondOp, PhiR->isOrdered(), CurrentLinkI->getDebugLoc()); // Append the recipe to the end of the VPBasicBlock because we need to // ensure that it comes after all of it's inputs, including CondOp. // Delete CurrentLink as it will be invalid if its operand is replaced // with a reduction defined at the bottom of the block in the next link. if (LinkVPBB->getNumSuccessors() == 0) RedRecipe->insertBefore(&*std::prev(std::prev(LinkVPBB->end()))); else LinkVPBB->appendRecipe(RedRecipe); CurrentLink->replaceAllUsesWith(RedRecipe); ToDelete.push_back(CurrentLink); PreviousLink = RedRecipe; } } VPBasicBlock *LatchVPBB = VectorLoopRegion->getExitingBasicBlock(); Builder.setInsertPoint(&*std::prev(std::prev(LatchVPBB->end()))); VPBasicBlock::iterator IP = MiddleVPBB->getFirstNonPhi(); for (VPRecipeBase &R : Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) { VPReductionPHIRecipe *PhiR = dyn_cast(&R); if (!PhiR) continue; const RecurrenceDescriptor &RdxDesc = Legal->getRecurrenceDescriptor( cast(PhiR->getUnderlyingInstr())); Type *PhiTy = PhiR->getUnderlyingValue()->getType(); // If tail is folded by masking, introduce selects between the phi // and the users outside the vector region of each reduction, at the // beginning of the dedicated latch block. auto *OrigExitingVPV = PhiR->getBackedgeValue(); auto *NewExitingVPV = PhiR->getBackedgeValue(); // Don't output selects for partial reductions because they have an output // with fewer lanes than the VF. So the operands of the select would have // different numbers of lanes. Partial reductions mask the input instead. if (!PhiR->isInLoop() && CM.foldTailByMasking() && !isa(OrigExitingVPV->getDefiningRecipe())) { VPValue *Cond = RecipeBuilder.getBlockInMask(PhiR->getParent()); std::optional FMFs = PhiTy->isFloatingPointTy() ? std::make_optional(RdxDesc.getFastMathFlags()) : std::nullopt; NewExitingVPV = Builder.createSelect(Cond, OrigExitingVPV, PhiR, {}, "", FMFs); OrigExitingVPV->replaceUsesWithIf(NewExitingVPV, [](VPUser &U, unsigned) { return isa(&U) && (cast(&U)->getOpcode() == VPInstruction::ComputeAnyOfResult || cast(&U)->getOpcode() == VPInstruction::ComputeReductionResult || cast(&U)->getOpcode() == VPInstruction::ComputeFindIVResult); }); if (CM.usePredicatedReductionSelect()) PhiR->setOperand(1, NewExitingVPV); } // We want code in the middle block to appear to execute on the location of // the scalar loop's latch terminator because: (a) it is all compiler // generated, (b) these instructions are always executed after evaluating // the latch conditional branch, and (c) other passes may add new // predecessors which terminate on this line. This is the easiest way to // ensure we don't accidentally cause an extra step back into the loop while // debugging. DebugLoc ExitDL = OrigLoop->getLoopLatch()->getTerminator()->getDebugLoc(); // TODO: At the moment ComputeReductionResult also drives creation of the // bc.merge.rdx phi nodes, hence it needs to be created unconditionally here // even for in-loop reductions, until the reduction resume value handling is // also modeled in VPlan. VPInstruction *FinalReductionResult; VPBuilder::InsertPointGuard Guard(Builder); Builder.setInsertPoint(MiddleVPBB, IP); RecurKind RecurrenceKind = PhiR->getRecurrenceKind(); if (RecurrenceDescriptor::isFindIVRecurrenceKind(RecurrenceKind)) { VPValue *Start = PhiR->getStartValue(); VPValue *Sentinel = Plan->getOrAddLiveIn(RdxDesc.getSentinelValue()); FinalReductionResult = Builder.createNaryOp(VPInstruction::ComputeFindIVResult, {PhiR, Start, Sentinel, NewExitingVPV}, ExitDL); } else if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RecurrenceKind)) { VPValue *Start = PhiR->getStartValue(); FinalReductionResult = Builder.createNaryOp(VPInstruction::ComputeAnyOfResult, {PhiR, Start, NewExitingVPV}, ExitDL); } else { VPIRFlags Flags = RecurrenceDescriptor::isFloatingPointRecurrenceKind(RecurrenceKind) ? VPIRFlags(RdxDesc.getFastMathFlags()) : VPIRFlags(); FinalReductionResult = Builder.createNaryOp(VPInstruction::ComputeReductionResult, {PhiR, NewExitingVPV}, Flags, ExitDL); } // If the vector reduction can be performed in a smaller type, we truncate // then extend the loop exit value to enable InstCombine to evaluate the // entire expression in the smaller type. if (MinVF.isVector() && PhiTy != RdxDesc.getRecurrenceType() && !RecurrenceDescriptor::isAnyOfRecurrenceKind(RecurrenceKind)) { assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!"); assert(!RecurrenceDescriptor::isMinMaxRecurrenceKind(RecurrenceKind) && "Unexpected truncated min-max recurrence!"); Type *RdxTy = RdxDesc.getRecurrenceType(); auto *Trunc = new VPWidenCastRecipe(Instruction::Trunc, NewExitingVPV, RdxTy); Instruction::CastOps ExtendOpc = RdxDesc.isSigned() ? Instruction::SExt : Instruction::ZExt; auto *Extnd = new VPWidenCastRecipe(ExtendOpc, Trunc, PhiTy); Trunc->insertAfter(NewExitingVPV->getDefiningRecipe()); Extnd->insertAfter(Trunc); if (PhiR->getOperand(1) == NewExitingVPV) PhiR->setOperand(1, Extnd->getVPSingleValue()); // Update ComputeReductionResult with the truncated exiting value and // extend its result. FinalReductionResult->setOperand(1, Trunc); FinalReductionResult = Builder.createScalarCast(ExtendOpc, FinalReductionResult, PhiTy, {}); } // Update all users outside the vector region. Also replace redundant // ExtractLastElement. for (auto *U : to_vector(OrigExitingVPV->users())) { auto *Parent = cast(U)->getParent(); if (FinalReductionResult == U || Parent->getParent()) continue; U->replaceUsesOfWith(OrigExitingVPV, FinalReductionResult); if (match(U, m_VPInstruction( m_VPValue()))) cast(U)->replaceAllUsesWith(FinalReductionResult); } // Adjust AnyOf reductions; replace the reduction phi for the selected value // with a boolean reduction phi node to check if the condition is true in // any iteration. The final value is selected by the final // ComputeReductionResult. if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RecurrenceKind)) { auto *Select = cast(*find_if(PhiR->users(), [](VPUser *U) { return isa(U) || (isa(U) && cast(U)->getUnderlyingInstr()->getOpcode() == Instruction::Select); })); VPValue *Cmp = Select->getOperand(0); // If the compare is checking the reduction PHI node, adjust it to check // the start value. if (VPRecipeBase *CmpR = Cmp->getDefiningRecipe()) CmpR->replaceUsesOfWith(PhiR, PhiR->getStartValue()); Builder.setInsertPoint(Select); // If the true value of the select is the reduction phi, the new value is // selected if the negated condition is true in any iteration. if (Select->getOperand(1) == PhiR) Cmp = Builder.createNot(Cmp); VPValue *Or = Builder.createOr(PhiR, Cmp); Select->getVPSingleValue()->replaceAllUsesWith(Or); // Delete Select now that it has invalid types. ToDelete.push_back(Select); // Convert the reduction phi to operate on bools. PhiR->setOperand(0, Plan->getOrAddLiveIn(ConstantInt::getFalse( OrigLoop->getHeader()->getContext()))); continue; } if (RecurrenceDescriptor::isFindIVRecurrenceKind( RdxDesc.getRecurrenceKind())) { // Adjust the start value for FindFirstIV/FindLastIV recurrences to use // the sentinel value after generating the ResumePhi recipe, which uses // the original start value. PhiR->setOperand(0, Plan->getOrAddLiveIn(RdxDesc.getSentinelValue())); } RecurKind RK = RdxDesc.getRecurrenceKind(); if ((!RecurrenceDescriptor::isAnyOfRecurrenceKind(RK) && !RecurrenceDescriptor::isFindIVRecurrenceKind(RK) && !RecurrenceDescriptor::isMinMaxRecurrenceKind(RK))) { VPBuilder PHBuilder(Plan->getVectorPreheader()); VPValue *Iden = Plan->getOrAddLiveIn( getRecurrenceIdentity(RK, PhiTy, RdxDesc.getFastMathFlags())); // If the PHI is used by a partial reduction, set the scale factor. unsigned ScaleFactor = RecipeBuilder.getScalingForReduction(RdxDesc.getLoopExitInstr()) .value_or(1); Type *I32Ty = IntegerType::getInt32Ty(PhiTy->getContext()); auto *ScaleFactorVPV = Plan->getOrAddLiveIn(ConstantInt::get(I32Ty, ScaleFactor)); VPValue *StartV = PHBuilder.createNaryOp( VPInstruction::ReductionStartVector, {PhiR->getStartValue(), Iden, ScaleFactorVPV}, PhiTy->isFloatingPointTy() ? RdxDesc.getFastMathFlags() : FastMathFlags()); PhiR->setOperand(0, StartV); } } for (VPRecipeBase *R : ToDelete) R->eraseFromParent(); VPlanTransforms::runPass(VPlanTransforms::clearReductionWrapFlags, *Plan); } void LoopVectorizationPlanner::attachRuntimeChecks( VPlan &Plan, GeneratedRTChecks &RTChecks, bool HasBranchWeights) const { const auto &[SCEVCheckCond, SCEVCheckBlock] = RTChecks.getSCEVChecks(); if (SCEVCheckBlock) { assert((!CM.OptForSize || CM.Hints->getForce() == LoopVectorizeHints::FK_Enabled) && "Cannot SCEV check stride or overflow when optimizing for size"); VPlanTransforms::attachCheckBlock(Plan, SCEVCheckCond, SCEVCheckBlock, HasBranchWeights); } const auto &[MemCheckCond, MemCheckBlock] = RTChecks.getMemRuntimeChecks(); if (MemCheckBlock) { // VPlan-native path does not do any analysis for runtime checks // currently. assert((!EnableVPlanNativePath || OrigLoop->isInnermost()) && "Runtime checks are not supported for outer loops yet"); if (CM.OptForSize) { assert( CM.Hints->getForce() == LoopVectorizeHints::FK_Enabled && "Cannot emit memory checks when optimizing for size, unless forced " "to vectorize."); ORE->emit([&]() { return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationCodeSize", OrigLoop->getStartLoc(), OrigLoop->getHeader()) << "Code-size may be reduced by not forcing " "vectorization, or by source-code modifications " "eliminating the need for runtime checks " "(e.g., adding 'restrict')."; }); } VPlanTransforms::attachCheckBlock(Plan, MemCheckCond, MemCheckBlock, HasBranchWeights); } } void VPDerivedIVRecipe::execute(VPTransformState &State) { assert(!State.Lane && "VPDerivedIVRecipe being replicated."); // Fast-math-flags propagate from the original induction instruction. IRBuilder<>::FastMathFlagGuard FMFG(State.Builder); if (FPBinOp) State.Builder.setFastMathFlags(FPBinOp->getFastMathFlags()); Value *Step = State.get(getStepValue(), VPLane(0)); Value *Index = State.get(getOperand(1), VPLane(0)); Value *DerivedIV = emitTransformedIndex( State.Builder, Index, getStartValue()->getLiveInIRValue(), Step, Kind, cast_if_present(FPBinOp)); DerivedIV->setName(Name); State.set(this, DerivedIV, VPLane(0)); } // Determine how to lower the scalar epilogue, which depends on 1) optimising // for minimum code-size, 2) predicate compiler options, 3) loop hints forcing // predication, and 4) a TTI hook that analyses whether the loop is suitable // for predication. static ScalarEpilogueLowering getScalarEpilogueLowering( Function *F, Loop *L, LoopVectorizeHints &Hints, ProfileSummaryInfo *PSI, BlockFrequencyInfo *BFI, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, LoopVectorizationLegality &LVL, InterleavedAccessInfo *IAI) { // 1) OptSize takes precedence over all other options, i.e. if this is set, // don't look at hints or options, and don't request a scalar epilogue. // (For PGSO, as shouldOptimizeForSize isn't currently accessible from // LoopAccessInfo (due to code dependency and not being able to reliably get // PSI/BFI from a loop analysis under NPM), we cannot suppress the collection // of strides in LoopAccessInfo::analyzeLoop() and vectorize without // versioning when the vectorization is forced, unlike hasOptSize. So revert // back to the old way and vectorize with versioning when forced. See D81345.) if (F->hasOptSize() || (llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI, PGSOQueryType::IRPass) && Hints.getForce() != LoopVectorizeHints::FK_Enabled)) return CM_ScalarEpilogueNotAllowedOptSize; // 2) If set, obey the directives if (PreferPredicateOverEpilogue.getNumOccurrences()) { switch (PreferPredicateOverEpilogue) { case PreferPredicateTy::ScalarEpilogue: return CM_ScalarEpilogueAllowed; case PreferPredicateTy::PredicateElseScalarEpilogue: return CM_ScalarEpilogueNotNeededUsePredicate; case PreferPredicateTy::PredicateOrDontVectorize: return CM_ScalarEpilogueNotAllowedUsePredicate; }; } // 3) If set, obey the hints switch (Hints.getPredicate()) { case LoopVectorizeHints::FK_Enabled: return CM_ScalarEpilogueNotNeededUsePredicate; case LoopVectorizeHints::FK_Disabled: return CM_ScalarEpilogueAllowed; }; // 4) if the TTI hook indicates this is profitable, request predication. TailFoldingInfo TFI(TLI, &LVL, IAI); if (TTI->preferPredicateOverEpilogue(&TFI)) return CM_ScalarEpilogueNotNeededUsePredicate; return CM_ScalarEpilogueAllowed; } // Process the loop in the VPlan-native vectorization path. This path builds // VPlan upfront in the vectorization pipeline, which allows to apply // VPlan-to-VPlan transformations from the very beginning without modifying the // input LLVM IR. static bool processLoopInVPlanNativePath( Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT, LoopVectorizationLegality *LVL, TargetTransformInfo *TTI, TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC, OptimizationRemarkEmitter *ORE, BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI, LoopVectorizeHints &Hints, LoopVectorizationRequirements &Requirements) { if (isa(PSE.getBackedgeTakenCount())) { LLVM_DEBUG(dbgs() << "LV: cannot compute the outer-loop trip count\n"); return false; } assert(EnableVPlanNativePath && "VPlan-native path is disabled."); Function *F = L->getHeader()->getParent(); InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI()); ScalarEpilogueLowering SEL = getScalarEpilogueLowering(F, L, Hints, PSI, BFI, TTI, TLI, *LVL, &IAI); LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F, &Hints, IAI, PSI, BFI); // Use the planner for outer loop vectorization. // TODO: CM is not used at this point inside the planner. Turn CM into an // optional argument if we don't need it in the future. LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, LVL, CM, IAI, PSE, Hints, ORE); // Get user vectorization factor. ElementCount UserVF = Hints.getWidth(); CM.collectElementTypesForWidening(); // Plan how to best vectorize, return the best VF and its cost. const VectorizationFactor VF = LVP.planInVPlanNativePath(UserVF); // If we are stress testing VPlan builds, do not attempt to generate vector // code. Masked vector code generation support will follow soon. // Also, do not attempt to vectorize if no vector code will be produced. if (VPlanBuildStressTest || VectorizationFactor::Disabled() == VF) return false; VPlan &BestPlan = LVP.getPlanFor(VF.Width); { GeneratedRTChecks Checks(PSE, DT, LI, TTI, F->getDataLayout(), CM.CostKind); InnerLoopVectorizer LB(L, PSE, LI, DT, TTI, AC, VF.Width, VF.Width, 1, &CM, BFI, PSI, Checks, BestPlan); LLVM_DEBUG(dbgs() << "Vectorizing outer loop in \"" << L->getHeader()->getParent()->getName() << "\"\n"); LVP.executePlan(VF.Width, 1, BestPlan, LB, DT, false); } reportVectorization(ORE, L, VF, 1); // Mark the loop as already vectorized to avoid vectorizing again. Hints.setAlreadyVectorized(); assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs())); return true; } // Emit a remark if there are stores to floats that required a floating point // extension. If the vectorized loop was generated with floating point there // will be a performance penalty from the conversion overhead and the change in // the vector width. static void checkMixedPrecision(Loop *L, OptimizationRemarkEmitter *ORE) { SmallVector Worklist; for (BasicBlock *BB : L->getBlocks()) { for (Instruction &Inst : *BB) { if (auto *S = dyn_cast(&Inst)) { if (S->getValueOperand()->getType()->isFloatTy()) Worklist.push_back(S); } } } // Traverse the floating point stores upwards searching, for floating point // conversions. SmallPtrSet Visited; SmallPtrSet EmittedRemark; while (!Worklist.empty()) { auto *I = Worklist.pop_back_val(); if (!L->contains(I)) continue; if (!Visited.insert(I).second) continue; // Emit a remark if the floating point store required a floating // point conversion. // TODO: More work could be done to identify the root cause such as a // constant or a function return type and point the user to it. if (isa(I) && EmittedRemark.insert(I).second) ORE->emit([&]() { return OptimizationRemarkAnalysis(LV_NAME, "VectorMixedPrecision", I->getDebugLoc(), L->getHeader()) << "floating point conversion changes vector width. " << "Mixed floating point precision requires an up/down " << "cast that will negatively impact performance."; }); for (Use &Op : I->operands()) if (auto *OpI = dyn_cast(Op)) Worklist.push_back(OpI); } } /// For loops with uncountable early exits, find the cost of doing work when /// exiting the loop early, such as calculating the final exit values of /// variables used outside the loop. /// TODO: This is currently overly pessimistic because the loop may not take /// the early exit, but better to keep this conservative for now. In future, /// it might be possible to relax this by using branch probabilities. static InstructionCost calculateEarlyExitCost(VPCostContext &CostCtx, VPlan &Plan, ElementCount VF) { InstructionCost Cost = 0; for (auto *ExitVPBB : Plan.getExitBlocks()) { for (auto *PredVPBB : ExitVPBB->getPredecessors()) { // If the predecessor is not the middle.block, then it must be the // vector.early.exit block, which may contain work to calculate the exit // values of variables used outside the loop. if (PredVPBB != Plan.getMiddleBlock()) { LLVM_DEBUG(dbgs() << "Calculating cost of work in exit block " << PredVPBB->getName() << ":\n"); Cost += PredVPBB->cost(VF, CostCtx); } } } return Cost; } /// This function determines whether or not it's still profitable to vectorize /// the loop given the extra work we have to do outside of the loop: /// 1. Perform the runtime checks before entering the loop to ensure it's safe /// to vectorize. /// 2. In the case of loops with uncountable early exits, we may have to do /// extra work when exiting the loop early, such as calculating the final /// exit values of variables used outside the loop. static bool isOutsideLoopWorkProfitable(GeneratedRTChecks &Checks, VectorizationFactor &VF, Loop *L, PredicatedScalarEvolution &PSE, VPCostContext &CostCtx, VPlan &Plan, ScalarEpilogueLowering SEL, std::optional VScale) { InstructionCost TotalCost = Checks.getCost(); if (!TotalCost.isValid()) return false; // Add on the cost of any work required in the vector early exit block, if // one exists. TotalCost += calculateEarlyExitCost(CostCtx, Plan, VF.Width); // When interleaving only scalar and vector cost will be equal, which in turn // would lead to a divide by 0. Fall back to hard threshold. if (VF.Width.isScalar()) { // TODO: Should we rename VectorizeMemoryCheckThreshold? if (TotalCost > VectorizeMemoryCheckThreshold) { LLVM_DEBUG( dbgs() << "LV: Interleaving only is not profitable due to runtime checks\n"); return false; } return true; } // The scalar cost should only be 0 when vectorizing with a user specified // VF/IC. In those cases, runtime checks should always be generated. uint64_t ScalarC = VF.ScalarCost.getValue(); if (ScalarC == 0) return true; // First, compute the minimum iteration count required so that the vector // loop outperforms the scalar loop. // The total cost of the scalar loop is // ScalarC * TC // where // * TC is the actual trip count of the loop. // * ScalarC is the cost of a single scalar iteration. // // The total cost of the vector loop is // RtC + VecC * (TC / VF) + EpiC // where // * RtC is the cost of the generated runtime checks plus the cost of // performing any additional work in the vector.early.exit block for loops // with uncountable early exits. // * VecC is the cost of a single vector iteration. // * TC is the actual trip count of the loop // * VF is the vectorization factor // * EpiCost is the cost of the generated epilogue, including the cost // of the remaining scalar operations. // // Vectorization is profitable once the total vector cost is less than the // total scalar cost: // RtC + VecC * (TC / VF) + EpiC < ScalarC * TC // // Now we can compute the minimum required trip count TC as // VF * (RtC + EpiC) / (ScalarC * VF - VecC) < TC // // For now we assume the epilogue cost EpiC = 0 for simplicity. Note that // the computations are performed on doubles, not integers and the result // is rounded up, hence we get an upper estimate of the TC. unsigned IntVF = estimateElementCount(VF.Width, VScale); uint64_t RtC = TotalCost.getValue(); uint64_t Div = ScalarC * IntVF - VF.Cost.getValue(); uint64_t MinTC1 = Div == 0 ? 0 : divideCeil(RtC * IntVF, Div); // Second, compute a minimum iteration count so that the cost of the // runtime checks is only a fraction of the total scalar loop cost. This // adds a loop-dependent bound on the overhead incurred if the runtime // checks fail. In case the runtime checks fail, the cost is RtC + ScalarC // * TC. To bound the runtime check to be a fraction 1/X of the scalar // cost, compute // RtC < ScalarC * TC * (1 / X) ==> RtC * X / ScalarC < TC uint64_t MinTC2 = divideCeil(RtC * 10, ScalarC); // Now pick the larger minimum. If it is not a multiple of VF and a scalar // epilogue is allowed, choose the next closest multiple of VF. This should // partly compensate for ignoring the epilogue cost. uint64_t MinTC = std::max(MinTC1, MinTC2); if (SEL == CM_ScalarEpilogueAllowed) MinTC = alignTo(MinTC, IntVF); VF.MinProfitableTripCount = ElementCount::getFixed(MinTC); LLVM_DEBUG( dbgs() << "LV: Minimum required TC for runtime checks to be profitable:" << VF.MinProfitableTripCount << "\n"); // Skip vectorization if the expected trip count is less than the minimum // required trip count. if (auto ExpectedTC = getSmallBestKnownTC(PSE, L)) { if (ElementCount::isKnownLT(*ExpectedTC, VF.MinProfitableTripCount)) { LLVM_DEBUG(dbgs() << "LV: Vectorization is not beneficial: expected " "trip count < minimum profitable VF (" << *ExpectedTC << " < " << VF.MinProfitableTripCount << ")\n"); return false; } } return true; } LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts) : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced || !EnableLoopInterleaving), VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced || !EnableLoopVectorization) {} /// Prepare \p MainPlan for vectorizing the main vector loop during epilogue /// vectorization. Remove ResumePhis from \p MainPlan for inductions that /// don't have a corresponding wide induction in \p EpiPlan. static void preparePlanForMainVectorLoop(VPlan &MainPlan, VPlan &EpiPlan) { // Collect PHI nodes of widened phis in the VPlan for the epilogue. Those // will need their resume-values computed in the main vector loop. Others // can be removed from the main VPlan. SmallPtrSet EpiWidenedPhis; for (VPRecipeBase &R : EpiPlan.getVectorLoopRegion()->getEntryBasicBlock()->phis()) { if (isa(&R)) continue; EpiWidenedPhis.insert( cast(R.getVPSingleValue()->getUnderlyingValue())); } for (VPRecipeBase &R : make_early_inc_range(MainPlan.getScalarHeader()->phis())) { auto *VPIRInst = cast(&R); if (EpiWidenedPhis.contains(&VPIRInst->getIRPhi())) continue; // There is no corresponding wide induction in the epilogue plan that would // need a resume value. Remove the VPIRInst wrapping the scalar header phi // together with the corresponding ResumePhi. The resume values for the // scalar loop will be created during execution of EpiPlan. VPRecipeBase *ResumePhi = VPIRInst->getOperand(0)->getDefiningRecipe(); VPIRInst->eraseFromParent(); ResumePhi->eraseFromParent(); } VPlanTransforms::runPass(VPlanTransforms::removeDeadRecipes, MainPlan); using namespace VPlanPatternMatch; // When vectorizing the epilogue, FindFirstIV & FindLastIV reductions can // introduce multiple uses of undef/poison. If the reduction start value may // be undef or poison it needs to be frozen and the frozen start has to be // used when computing the reduction result. We also need to use the frozen // value in the resume phi generated by the main vector loop, as this is also // used to compute the reduction result after the epilogue vector loop. auto AddFreezeForFindLastIVReductions = [](VPlan &Plan, bool UpdateResumePhis) { VPBuilder Builder(Plan.getEntry()); for (VPRecipeBase &R : *Plan.getMiddleBlock()) { auto *VPI = dyn_cast(&R); if (!VPI || VPI->getOpcode() != VPInstruction::ComputeFindIVResult) continue; VPValue *OrigStart = VPI->getOperand(1); if (isGuaranteedNotToBeUndefOrPoison(OrigStart->getLiveInIRValue())) continue; VPInstruction *Freeze = Builder.createNaryOp(Instruction::Freeze, {OrigStart}, {}, "fr"); VPI->setOperand(1, Freeze); if (UpdateResumePhis) OrigStart->replaceUsesWithIf(Freeze, [Freeze](VPUser &U, unsigned) { return Freeze != &U && isa(&U); }); } }; AddFreezeForFindLastIVReductions(MainPlan, true); AddFreezeForFindLastIVReductions(EpiPlan, false); VPBasicBlock *MainScalarPH = MainPlan.getScalarPreheader(); VPValue *VectorTC = &MainPlan.getVectorTripCount(); // If there is a suitable resume value for the canonical induction in the // scalar (which will become vector) epilogue loop, use it and move it to the // beginning of the scalar preheader. Otherwise create it below. auto ResumePhiIter = find_if(MainScalarPH->phis(), [VectorTC](VPRecipeBase &R) { return match(&R, m_VPInstruction(m_Specific(VectorTC), m_SpecificInt(0))); }); VPPhi *ResumePhi = nullptr; if (ResumePhiIter == MainScalarPH->phis().end()) { VPBuilder ScalarPHBuilder(MainScalarPH, MainScalarPH->begin()); ResumePhi = ScalarPHBuilder.createScalarPhi( {VectorTC, MainPlan.getCanonicalIV()->getStartValue()}, {}, "vec.epilog.resume.val"); } else { ResumePhi = cast(&*ResumePhiIter); if (MainScalarPH->begin() == MainScalarPH->end()) ResumePhi->moveBefore(*MainScalarPH, MainScalarPH->end()); else if (&*MainScalarPH->begin() != ResumePhi) ResumePhi->moveBefore(*MainScalarPH, MainScalarPH->begin()); } // Add a user to to make sure the resume phi won't get removed. VPBuilder(MainScalarPH) .createNaryOp(VPInstruction::ResumeForEpilogue, ResumePhi); } /// Prepare \p Plan for vectorizing the epilogue loop. That is, re-use expanded /// SCEVs from \p ExpandedSCEVs and set resume values for header recipes. static void preparePlanForEpilogueVectorLoop(VPlan &Plan, Loop *L, const SCEV2ValueTy &ExpandedSCEVs, EpilogueLoopVectorizationInfo &EPI) { VPRegionBlock *VectorLoop = Plan.getVectorLoopRegion(); VPBasicBlock *Header = VectorLoop->getEntryBasicBlock(); Header->setName("vec.epilog.vector.body"); DenseMap ToFrozen; // Ensure that the start values for all header phi recipes are updated before // vectorizing the epilogue loop. for (VPRecipeBase &R : Header->phis()) { if (auto *IV = dyn_cast(&R)) { // When vectorizing the epilogue loop, the canonical induction start // value needs to be changed from zero to the value after the main // vector loop. Find the resume value created during execution of the main // VPlan. It must be the first phi in the loop preheader. // FIXME: Improve modeling for canonical IV start values in the epilogue // loop. using namespace llvm::PatternMatch; PHINode *EPResumeVal = &*L->getLoopPreheader()->phis().begin(); for (Value *Inc : EPResumeVal->incoming_values()) { if (match(Inc, m_SpecificInt(0))) continue; assert(!EPI.VectorTripCount && "Must only have a single non-zero incoming value"); EPI.VectorTripCount = Inc; } VPValue *VPV = Plan.getOrAddLiveIn(EPResumeVal); assert(all_of(IV->users(), [](const VPUser *U) { return isa(U) || isa(U) || cast(U)->isScalarCast() || cast(U)->getOpcode() == Instruction::Add; }) && "the canonical IV should only be used by its increment or " "ScalarIVSteps when resetting the start value"); IV->setOperand(0, VPV); continue; } Value *ResumeV = nullptr; // TODO: Move setting of resume values to prepareToExecute. if (auto *ReductionPhi = dyn_cast(&R)) { auto *RdxResult = cast(*find_if(ReductionPhi->users(), [](VPUser *U) { auto *VPI = dyn_cast(U); return VPI && (VPI->getOpcode() == VPInstruction::ComputeAnyOfResult || VPI->getOpcode() == VPInstruction::ComputeReductionResult || VPI->getOpcode() == VPInstruction::ComputeFindIVResult); })); ResumeV = cast(ReductionPhi->getUnderlyingInstr()) ->getIncomingValueForBlock(L->getLoopPreheader()); RecurKind RK = ReductionPhi->getRecurrenceKind(); if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RK)) { Value *StartV = RdxResult->getOperand(1)->getLiveInIRValue(); // VPReductionPHIRecipes for AnyOf reductions expect a boolean as // start value; compare the final value from the main vector loop // to the start value. BasicBlock *PBB = cast(ResumeV)->getParent(); IRBuilder<> Builder(PBB, PBB->getFirstNonPHIIt()); ResumeV = Builder.CreateICmpNE(ResumeV, StartV); } else if (RecurrenceDescriptor::isFindIVRecurrenceKind(RK)) { Value *StartV = getStartValueFromReductionResult(RdxResult); ToFrozen[StartV] = cast(ResumeV)->getIncomingValueForBlock( EPI.MainLoopIterationCountCheck); // VPReductionPHIRecipe for FindFirstIV/FindLastIV reductions requires // an adjustment to the resume value. The resume value is adjusted to // the sentinel value when the final value from the main vector loop // equals the start value. This ensures correctness when the start value // might not be less than the minimum value of a monotonically // increasing induction variable. BasicBlock *ResumeBB = cast(ResumeV)->getParent(); IRBuilder<> Builder(ResumeBB, ResumeBB->getFirstNonPHIIt()); Value *Cmp = Builder.CreateICmpEQ(ResumeV, ToFrozen[StartV]); Value *Sentinel = RdxResult->getOperand(2)->getLiveInIRValue(); ResumeV = Builder.CreateSelect(Cmp, Sentinel, ResumeV); } else { VPValue *StartVal = Plan.getOrAddLiveIn(ResumeV); auto *PhiR = dyn_cast(&R); if (auto *VPI = dyn_cast(PhiR->getStartValue())) { assert(VPI->getOpcode() == VPInstruction::ReductionStartVector && "unexpected start value"); VPI->setOperand(0, StartVal); continue; } } } else { // Retrieve the induction resume values for wide inductions from // their original phi nodes in the scalar loop. PHINode *IndPhi = cast(&R)->getPHINode(); // Hook up to the PHINode generated by a ResumePhi recipe of main // loop VPlan, which feeds the scalar loop. ResumeV = IndPhi->getIncomingValueForBlock(L->getLoopPreheader()); } assert(ResumeV && "Must have a resume value"); VPValue *StartVal = Plan.getOrAddLiveIn(ResumeV); cast(&R)->setStartValue(StartVal); } // For some VPValues in the epilogue plan we must re-use the generated IR // values from the main plan. Replace them with live-in VPValues. // TODO: This is a workaround needed for epilogue vectorization and it // should be removed once induction resume value creation is done // directly in VPlan. for (auto &R : make_early_inc_range(*Plan.getEntry())) { // Re-use frozen values from the main plan for Freeze VPInstructions in the // epilogue plan. This ensures all users use the same frozen value. auto *VPI = dyn_cast(&R); if (VPI && VPI->getOpcode() == Instruction::Freeze) { VPI->replaceAllUsesWith(Plan.getOrAddLiveIn( ToFrozen.lookup(VPI->getOperand(0)->getLiveInIRValue()))); continue; } // Re-use the trip count and steps expanded for the main loop, as // skeleton creation needs it as a value that dominates both the scalar // and vector epilogue loops auto *ExpandR = dyn_cast(&R); if (!ExpandR) continue; VPValue *ExpandedVal = Plan.getOrAddLiveIn(ExpandedSCEVs.lookup(ExpandR->getSCEV())); ExpandR->replaceAllUsesWith(ExpandedVal); if (Plan.getTripCount() == ExpandR) Plan.resetTripCount(ExpandedVal); ExpandR->eraseFromParent(); } } // Generate bypass values from the additional bypass block. Note that when the // vectorized epilogue is skipped due to iteration count check, then the // resume value for the induction variable comes from the trip count of the // main vector loop, passed as the second argument. static Value *createInductionAdditionalBypassValues( PHINode *OrigPhi, const InductionDescriptor &II, IRBuilder<> &BypassBuilder, const SCEV2ValueTy &ExpandedSCEVs, Value *MainVectorTripCount, Instruction *OldInduction) { Value *Step = getExpandedStep(II, ExpandedSCEVs); // For the primary induction the additional bypass end value is known. // Otherwise it is computed. Value *EndValueFromAdditionalBypass = MainVectorTripCount; if (OrigPhi != OldInduction) { auto *BinOp = II.getInductionBinOp(); // Fast-math-flags propagate from the original induction instruction. if (isa_and_nonnull(BinOp)) BypassBuilder.setFastMathFlags(BinOp->getFastMathFlags()); // Compute the end value for the additional bypass. EndValueFromAdditionalBypass = emitTransformedIndex(BypassBuilder, MainVectorTripCount, II.getStartValue(), Step, II.getKind(), BinOp); EndValueFromAdditionalBypass->setName("ind.end"); } return EndValueFromAdditionalBypass; } bool LoopVectorizePass::processLoop(Loop *L) { assert((EnableVPlanNativePath || L->isInnermost()) && "VPlan-native path is not enabled. Only process inner loops."); LLVM_DEBUG(dbgs() << "\nLV: Checking a loop in '" << L->getHeader()->getParent()->getName() << "' from " << L->getLocStr() << "\n"); LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE, TTI); LLVM_DEBUG( dbgs() << "LV: Loop hints:" << " force=" << (Hints.getForce() == LoopVectorizeHints::FK_Disabled ? "disabled" : (Hints.getForce() == LoopVectorizeHints::FK_Enabled ? "enabled" : "?")) << " width=" << Hints.getWidth() << " interleave=" << Hints.getInterleave() << "\n"); // Function containing loop Function *F = L->getHeader()->getParent(); // Looking at the diagnostic output is the only way to determine if a loop // was vectorized (other than looking at the IR or machine code), so it // is important to generate an optimization remark for each loop. Most of // these messages are generated as OptimizationRemarkAnalysis. Remarks // generated as OptimizationRemark and OptimizationRemarkMissed are // less verbose reporting vectorized loops and unvectorized loops that may // benefit from vectorization, respectively. if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) { LLVM_DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n"); return false; } PredicatedScalarEvolution PSE(*SE, *L); // Check if it is legal to vectorize the loop. LoopVectorizationRequirements Requirements; LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, F, *LAIs, LI, ORE, &Requirements, &Hints, DB, AC, BFI, PSI); if (!LVL.canVectorize(EnableVPlanNativePath)) { LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n"); Hints.emitRemarkWithHints(); return false; } if (LVL.hasUncountableEarlyExit() && !EnableEarlyExitVectorization) { reportVectorizationFailure("Auto-vectorization of loops with uncountable " "early exit is not enabled", "UncountableEarlyExitLoopsDisabled", ORE, L); return false; } // Entrance to the VPlan-native vectorization path. Outer loops are processed // here. They may require CFG and instruction level transformations before // even evaluating whether vectorization is profitable. Since we cannot modify // the incoming IR, we need to build VPlan upfront in the vectorization // pipeline. if (!L->isInnermost()) return processLoopInVPlanNativePath(L, PSE, LI, DT, &LVL, TTI, TLI, DB, AC, ORE, BFI, PSI, Hints, Requirements); assert(L->isInnermost() && "Inner loop expected."); InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI()); bool UseInterleaved = TTI->enableInterleavedAccessVectorization(); // If an override option has been passed in for interleaved accesses, use it. if (EnableInterleavedMemAccesses.getNumOccurrences() > 0) UseInterleaved = EnableInterleavedMemAccesses; // Analyze interleaved memory accesses. if (UseInterleaved) IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI)); if (LVL.hasUncountableEarlyExit()) { BasicBlock *LoopLatch = L->getLoopLatch(); if (IAI.requiresScalarEpilogue() || any_of(LVL.getCountableExitingBlocks(), [LoopLatch](BasicBlock *BB) { return BB != LoopLatch; })) { reportVectorizationFailure("Auto-vectorization of early exit loops " "requiring a scalar epilogue is unsupported", "UncountableEarlyExitUnsupported", ORE, L); return false; } } // Check the function attributes and profiles to find out if this function // should be optimized for size. ScalarEpilogueLowering SEL = getScalarEpilogueLowering(F, L, Hints, PSI, BFI, TTI, TLI, LVL, &IAI); // Check the loop for a trip count threshold: vectorize loops with a tiny trip // count by optimizing for size, to minimize overheads. auto ExpectedTC = getSmallBestKnownTC(PSE, L); if (ExpectedTC && ExpectedTC->isFixed() && ExpectedTC->getFixedValue() < TinyTripCountVectorThreshold) { LLVM_DEBUG(dbgs() << "LV: Found a loop with a very small trip count. " << "This loop is worth vectorizing only if no scalar " << "iteration overheads are incurred."); if (Hints.getForce() == LoopVectorizeHints::FK_Enabled) LLVM_DEBUG(dbgs() << " But vectorizing was explicitly forced.\n"); else { LLVM_DEBUG(dbgs() << "\n"); // Predicate tail-folded loops are efficient even when the loop // iteration count is low. However, setting the epilogue policy to // `CM_ScalarEpilogueNotAllowedLowTripLoop` prevents vectorizing loops // with runtime checks. It's more effective to let // `isOutsideLoopWorkProfitable` determine if vectorization is // beneficial for the loop. if (SEL != CM_ScalarEpilogueNotNeededUsePredicate) SEL = CM_ScalarEpilogueNotAllowedLowTripLoop; } } // Check the function attributes to see if implicit floats or vectors are // allowed. if (F->hasFnAttribute(Attribute::NoImplicitFloat)) { reportVectorizationFailure( "Can't vectorize when the NoImplicitFloat attribute is used", "loop not vectorized due to NoImplicitFloat attribute", "NoImplicitFloat", ORE, L); Hints.emitRemarkWithHints(); return false; } // Check if the target supports potentially unsafe FP vectorization. // FIXME: Add a check for the type of safety issue (denormal, signaling) // for the target we're vectorizing for, to make sure none of the // additional fp-math flags can help. if (Hints.isPotentiallyUnsafe() && TTI->isFPVectorizationPotentiallyUnsafe()) { reportVectorizationFailure( "Potentially unsafe FP op prevents vectorization", "loop not vectorized due to unsafe FP support.", "UnsafeFP", ORE, L); Hints.emitRemarkWithHints(); return false; } bool AllowOrderedReductions; // If the flag is set, use that instead and override the TTI behaviour. if (ForceOrderedReductions.getNumOccurrences() > 0) AllowOrderedReductions = ForceOrderedReductions; else AllowOrderedReductions = TTI->enableOrderedReductions(); if (!LVL.canVectorizeFPMath(AllowOrderedReductions)) { ORE->emit([&]() { auto *ExactFPMathInst = Requirements.getExactFPInst(); return OptimizationRemarkAnalysisFPCommute(DEBUG_TYPE, "CantReorderFPOps", ExactFPMathInst->getDebugLoc(), ExactFPMathInst->getParent()) << "loop not vectorized: cannot prove it is safe to reorder " "floating-point operations"; }); LLVM_DEBUG(dbgs() << "LV: loop not vectorized: cannot prove it is safe to " "reorder floating-point operations\n"); Hints.emitRemarkWithHints(); return false; } // Use the cost model. LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE, F, &Hints, IAI, PSI, BFI); // Use the planner for vectorization. LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, &LVL, CM, IAI, PSE, Hints, ORE); // Get user vectorization factor and interleave count. ElementCount UserVF = Hints.getWidth(); unsigned UserIC = Hints.getInterleave(); // Plan how to best vectorize. LVP.plan(UserVF, UserIC); VectorizationFactor VF = LVP.computeBestVF(); unsigned IC = 1; if (ORE->allowExtraAnalysis(LV_NAME)) LVP.emitInvalidCostRemarks(ORE); GeneratedRTChecks Checks(PSE, DT, LI, TTI, F->getDataLayout(), CM.CostKind); if (LVP.hasPlanWithVF(VF.Width)) { // Select the interleave count. IC = LVP.selectInterleaveCount(LVP.getPlanFor(VF.Width), VF.Width, VF.Cost); unsigned SelectedIC = std::max(IC, UserIC); // Optimistically generate runtime checks if they are needed. Drop them if // they turn out to not be profitable. if (VF.Width.isVector() || SelectedIC > 1) { Checks.create(L, *LVL.getLAI(), PSE.getPredicate(), VF.Width, SelectedIC); // Bail out early if either the SCEV or memory runtime checks are known to // fail. In that case, the vector loop would never execute. using namespace llvm::PatternMatch; if (Checks.getSCEVChecks().first && match(Checks.getSCEVChecks().first, m_One())) return false; if (Checks.getMemRuntimeChecks().first && match(Checks.getMemRuntimeChecks().first, m_One())) return false; } // Check if it is profitable to vectorize with runtime checks. bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled; VPCostContext CostCtx(CM.TTI, *CM.TLI, LVP.getPlanFor(VF.Width), CM, CM.CostKind); if (!ForceVectorization && !isOutsideLoopWorkProfitable(Checks, VF, L, PSE, CostCtx, LVP.getPlanFor(VF.Width), SEL, CM.getVScaleForTuning())) { ORE->emit([&]() { return OptimizationRemarkAnalysisAliasing( DEBUG_TYPE, "CantReorderMemOps", L->getStartLoc(), L->getHeader()) << "loop not vectorized: cannot prove it is safe to reorder " "memory operations"; }); LLVM_DEBUG(dbgs() << "LV: Too many memory checks needed.\n"); Hints.emitRemarkWithHints(); return false; } } // Identify the diagnostic messages that should be produced. std::pair VecDiagMsg, IntDiagMsg; bool VectorizeLoop = true, InterleaveLoop = true; if (VF.Width.isScalar()) { LLVM_DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n"); VecDiagMsg = { "VectorizationNotBeneficial", "the cost-model indicates that vectorization is not beneficial"}; VectorizeLoop = false; } if (!LVP.hasPlanWithVF(VF.Width) && UserIC > 1) { // Tell the user interleaving was avoided up-front, despite being explicitly // requested. LLVM_DEBUG(dbgs() << "LV: Ignoring UserIC, because vectorization and " "interleaving should be avoided up front\n"); IntDiagMsg = {"InterleavingAvoided", "Ignoring UserIC, because interleaving was avoided up front"}; InterleaveLoop = false; } else if (IC == 1 && UserIC <= 1) { // Tell the user interleaving is not beneficial. LLVM_DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n"); IntDiagMsg = { "InterleavingNotBeneficial", "the cost-model indicates that interleaving is not beneficial"}; InterleaveLoop = false; if (UserIC == 1) { IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled"; IntDiagMsg.second += " and is explicitly disabled or interleave count is set to 1"; } } else if (IC > 1 && UserIC == 1) { // Tell the user interleaving is beneficial, but it explicitly disabled. LLVM_DEBUG(dbgs() << "LV: Interleaving is beneficial but is explicitly " "disabled.\n"); IntDiagMsg = {"InterleavingBeneficialButDisabled", "the cost-model indicates that interleaving is beneficial " "but is explicitly disabled or interleave count is set to 1"}; InterleaveLoop = false; } // If there is a histogram in the loop, do not just interleave without // vectorizing. The order of operations will be incorrect without the // histogram intrinsics, which are only used for recipes with VF > 1. if (!VectorizeLoop && InterleaveLoop && LVL.hasHistograms()) { LLVM_DEBUG(dbgs() << "LV: Not interleaving without vectorization due " << "to histogram operations.\n"); IntDiagMsg = { "HistogramPreventsScalarInterleaving", "Unable to interleave without vectorization due to constraints on " "the order of histogram operations"}; InterleaveLoop = false; } // Override IC if user provided an interleave count. IC = UserIC > 0 ? UserIC : IC; // Emit diagnostic messages, if any. const char *VAPassName = Hints.vectorizeAnalysisPassName(); if (!VectorizeLoop && !InterleaveLoop) { // Do not vectorize or interleaving the loop. ORE->emit([&]() { return OptimizationRemarkMissed(VAPassName, VecDiagMsg.first, L->getStartLoc(), L->getHeader()) << VecDiagMsg.second; }); ORE->emit([&]() { return OptimizationRemarkMissed(LV_NAME, IntDiagMsg.first, L->getStartLoc(), L->getHeader()) << IntDiagMsg.second; }); return false; } if (!VectorizeLoop && InterleaveLoop) { LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n'); ORE->emit([&]() { return OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first, L->getStartLoc(), L->getHeader()) << VecDiagMsg.second; }); } else if (VectorizeLoop && !InterleaveLoop) { LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << L->getLocStr() << '\n'); ORE->emit([&]() { return OptimizationRemarkAnalysis(LV_NAME, IntDiagMsg.first, L->getStartLoc(), L->getHeader()) << IntDiagMsg.second; }); } else if (VectorizeLoop && InterleaveLoop) { LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width << ") in " << L->getLocStr() << '\n'); LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n'); } bool DisableRuntimeUnroll = false; MDNode *OrigLoopID = L->getLoopID(); { using namespace ore; if (!VectorizeLoop) { assert(IC > 1 && "interleave count should not be 1 or 0"); // If we decided that it is not legal to vectorize the loop, then // interleave it. VPlan &BestPlan = LVP.getPlanFor(VF.Width); InnerLoopVectorizer Unroller( L, PSE, LI, DT, TTI, AC, ElementCount::getFixed(1), ElementCount::getFixed(1), IC, &CM, BFI, PSI, Checks, BestPlan); // TODO: Move to general VPlan pipeline once epilogue loops are also // supported. VPlanTransforms::runPass( VPlanTransforms::materializeConstantVectorTripCount, BestPlan, VF.Width, IC, PSE); LVP.executePlan(VF.Width, IC, BestPlan, Unroller, DT, false); ORE->emit([&]() { return OptimizationRemark(LV_NAME, "Interleaved", L->getStartLoc(), L->getHeader()) << "interleaved loop (interleaved count: " << NV("InterleaveCount", IC) << ")"; }); } else { // If we decided that it is *legal* to vectorize the loop, then do it. VPlan &BestPlan = LVP.getPlanFor(VF.Width); // Consider vectorizing the epilogue too if it's profitable. VectorizationFactor EpilogueVF = LVP.selectEpilogueVectorizationFactor(VF.Width, IC); if (EpilogueVF.Width.isVector()) { std::unique_ptr BestMainPlan(BestPlan.duplicate()); // The first pass vectorizes the main loop and creates a scalar epilogue // to be vectorized by executing the plan (potentially with a different // factor) again shortly afterwards. VPlan &BestEpiPlan = LVP.getPlanFor(EpilogueVF.Width); BestEpiPlan.getMiddleBlock()->setName("vec.epilog.middle.block"); preparePlanForMainVectorLoop(*BestMainPlan, BestEpiPlan); EpilogueLoopVectorizationInfo EPI(VF.Width, IC, EpilogueVF.Width, 1, BestEpiPlan); EpilogueVectorizerMainLoop MainILV(L, PSE, LI, DT, TTI, AC, EPI, &CM, BFI, PSI, Checks, *BestMainPlan); auto ExpandedSCEVs = LVP.executePlan(EPI.MainLoopVF, EPI.MainLoopUF, *BestMainPlan, MainILV, DT, false); ++LoopsVectorized; // Second pass vectorizes the epilogue and adjusts the control flow // edges from the first pass. EpilogueVectorizerEpilogueLoop EpilogILV( L, PSE, LI, DT, TTI, AC, EPI, &CM, BFI, PSI, Checks, BestEpiPlan); EpilogILV.setTripCount(MainILV.getTripCount()); preparePlanForEpilogueVectorLoop(BestEpiPlan, L, ExpandedSCEVs, EPI); LVP.executePlan(EPI.EpilogueVF, EPI.EpilogueUF, BestEpiPlan, EpilogILV, DT, true); // Fix induction resume values from the additional bypass block. BasicBlock *BypassBlock = EpilogILV.getAdditionalBypassBlock(); IRBuilder<> BypassBuilder(BypassBlock, BypassBlock->getFirstInsertionPt()); BasicBlock *PH = L->getLoopPreheader(); for (const auto &[IVPhi, II] : LVL.getInductionVars()) { auto *Inc = cast(IVPhi->getIncomingValueForBlock(PH)); Value *V = createInductionAdditionalBypassValues( IVPhi, II, BypassBuilder, ExpandedSCEVs, EPI.VectorTripCount, LVL.getPrimaryInduction()); // TODO: Directly add as extra operand to the VPResumePHI recipe. Inc->setIncomingValueForBlock(BypassBlock, V); } ++LoopsEpilogueVectorized; if (!Checks.hasChecks()) DisableRuntimeUnroll = true; } else { InnerLoopVectorizer LB(L, PSE, LI, DT, TTI, AC, VF.Width, VF.MinProfitableTripCount, IC, &CM, BFI, PSI, Checks, BestPlan); // TODO: Move to general VPlan pipeline once epilogue loops are also // supported. VPlanTransforms::runPass( VPlanTransforms::materializeConstantVectorTripCount, BestPlan, VF.Width, IC, PSE); LVP.executePlan(VF.Width, IC, BestPlan, LB, DT, false); ++LoopsVectorized; // Add metadata to disable runtime unrolling a scalar loop when there // are no runtime checks about strides and memory. A scalar loop that is // rarely used is not worth unrolling. if (!Checks.hasChecks()) DisableRuntimeUnroll = true; } // Report the vectorization decision. reportVectorization(ORE, L, VF, IC); } if (ORE->allowExtraAnalysis(LV_NAME)) checkMixedPrecision(L, ORE); } assert(DT->verify(DominatorTree::VerificationLevel::Fast) && "DT not preserved correctly"); std::optional RemainderLoopID = makeFollowupLoopID(OrigLoopID, {LLVMLoopVectorizeFollowupAll, LLVMLoopVectorizeFollowupEpilogue}); if (RemainderLoopID) { L->setLoopID(*RemainderLoopID); } else { if (DisableRuntimeUnroll) addRuntimeUnrollDisableMetaData(L); // Mark the loop as already vectorized to avoid vectorizing again. Hints.setAlreadyVectorized(); } assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs())); return true; } LoopVectorizeResult LoopVectorizePass::runImpl(Function &F) { // Don't attempt if // 1. the target claims to have no vector registers, and // 2. interleaving won't help ILP. // // The second condition is necessary because, even if the target has no // vector registers, loop vectorization may still enable scalar // interleaving. if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true)) && TTI->getMaxInterleaveFactor(ElementCount::getFixed(1)) < 2) return LoopVectorizeResult(false, false); bool Changed = false, CFGChanged = false; // The vectorizer requires loops to be in simplified form. // Since simplification may add new inner loops, it has to run before the // legality and profitability checks. This means running the loop vectorizer // will simplify all loops, regardless of whether anything end up being // vectorized. for (const auto &L : *LI) Changed |= CFGChanged |= simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */); // Build up a worklist of inner-loops to vectorize. This is necessary as // the act of vectorizing or partially unrolling a loop creates new loops // and can invalidate iterators across the loops. SmallVector Worklist; for (Loop *L : *LI) collectSupportedLoops(*L, LI, ORE, Worklist); LoopsAnalyzed += Worklist.size(); // Now walk the identified inner loops. while (!Worklist.empty()) { Loop *L = Worklist.pop_back_val(); // For the inner loops we actually process, form LCSSA to simplify the // transform. Changed |= formLCSSARecursively(*L, *DT, LI, SE); Changed |= CFGChanged |= processLoop(L); if (Changed) { LAIs->clear(); #ifndef NDEBUG if (VerifySCEV) SE->verify(); #endif } } // Process each loop nest in the function. return LoopVectorizeResult(Changed, CFGChanged); } PreservedAnalyses LoopVectorizePass::run(Function &F, FunctionAnalysisManager &AM) { LI = &AM.getResult(F); // There are no loops in the function. Return before computing other // expensive analyses. if (LI->empty()) return PreservedAnalyses::all(); SE = &AM.getResult(F); TTI = &AM.getResult(F); DT = &AM.getResult(F); TLI = &AM.getResult(F); AC = &AM.getResult(F); DB = &AM.getResult(F); ORE = &AM.getResult(F); LAIs = &AM.getResult(F); auto &MAMProxy = AM.getResult(F); PSI = MAMProxy.getCachedResult(*F.getParent()); BFI = nullptr; if (PSI && PSI->hasProfileSummary()) BFI = &AM.getResult(F); LoopVectorizeResult Result = runImpl(F); if (!Result.MadeAnyChange) return PreservedAnalyses::all(); PreservedAnalyses PA; if (isAssignmentTrackingEnabled(*F.getParent())) { for (auto &BB : F) RemoveRedundantDbgInstrs(&BB); } PA.preserve(); PA.preserve(); PA.preserve(); PA.preserve(); if (Result.MadeCFGChange) { // Making CFG changes likely means a loop got vectorized. Indicate that // extra simplification passes should be run. // TODO: MadeCFGChanges is not a prefect proxy. Extra passes should only // be run if runtime checks have been added. AM.getResult(F); PA.preserve(); } else { PA.preserveSet(); } return PA; } void LoopVectorizePass::printPipeline( raw_ostream &OS, function_ref MapClassName2PassName) { static_cast *>(this)->printPipeline( OS, MapClassName2PassName); OS << '<'; OS << (InterleaveOnlyWhenForced ? "" : "no-") << "interleave-forced-only;"; OS << (VectorizeOnlyWhenForced ? "" : "no-") << "vectorize-forced-only;"; OS << '>'; }