brintos / llvm-project-archived public Read only

1//===- LoopVectorize.cpp - A Loop Vectorizer ------------------------------===//2//3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.4// See https://llvm.org/LICENSE.txt for license information.5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception6//7//===----------------------------------------------------------------------===//8//9// This is the LLVM loop vectorizer. This pass modifies 'vectorizable' loops10// and generates target-independent LLVM-IR.11// The vectorizer uses the TargetTransformInfo analysis to estimate the costs12// of instructions in order to estimate the profitability of vectorization.13//14// The loop vectorizer combines consecutive loop iterations into a single15// 'wide' iteration. After this transformation the index is incremented16// by the SIMD vector width, and not by one.17//18// This pass has three parts:19// 1. The main loop pass that drives the different parts.20// 2. LoopVectorizationLegality - A unit that checks for the legality21//    of the vectorization.22// 3. InnerLoopVectorizer - A unit that performs the actual23//    widening of instructions.24// 4. LoopVectorizationCostModel - A unit that checks for the profitability25//    of vectorization. It decides on the optimal vector width, which26//    can be one, if vectorization is not profitable.27//28// There is a development effort going on to migrate loop vectorizer to the29// VPlan infrastructure and to introduce outer loop vectorization support (see30// docs/VectorizationPlan.rst and31// http://lists.llvm.org/pipermail/llvm-dev/2017-December/119523.html). For this32// purpose, we temporarily introduced the VPlan-native vectorization path: an33// alternative vectorization path that is natively implemented on top of the34// VPlan infrastructure. See EnableVPlanNativePath for enabling.35//36//===----------------------------------------------------------------------===//37//38// The reduction-variable vectorization is based on the paper:39//  D. Nuzman and R. Henderson. Multi-platform Auto-vectorization.40//41// Variable uniformity checks are inspired by:42//  Karrenberg, R. and Hack, S. Whole Function Vectorization.43//44// The interleaved access vectorization is based on the paper:45//  Dorit Nuzman, Ira Rosen and Ayal Zaks.  Auto-Vectorization of Interleaved46//  Data for SIMD47//48// Other ideas/concepts are from:49//  A. Zaks and D. Nuzman. Autovectorization in GCC-two years later.50//51//  S. Maleki, Y. Gao, M. Garzaran, T. Wong and D. Padua.  An Evaluation of52//  Vectorizing Compilers.53//54//===----------------------------------------------------------------------===//55 56#include "llvm/Transforms/Vectorize/LoopVectorize.h"57#include "LoopVectorizationPlanner.h"58#include "VPRecipeBuilder.h"59#include "VPlan.h"60#include "VPlanAnalysis.h"61#include "VPlanCFG.h"62#include "VPlanHelpers.h"63#include "VPlanPatternMatch.h"64#include "VPlanTransforms.h"65#include "VPlanUtils.h"66#include "VPlanVerifier.h"67#include "llvm/ADT/APInt.h"68#include "llvm/ADT/ArrayRef.h"69#include "llvm/ADT/DenseMap.h"70#include "llvm/ADT/DenseMapInfo.h"71#include "llvm/ADT/Hashing.h"72#include "llvm/ADT/MapVector.h"73#include "llvm/ADT/STLExtras.h"74#include "llvm/ADT/SmallPtrSet.h"75#include "llvm/ADT/SmallVector.h"76#include "llvm/ADT/Statistic.h"77#include "llvm/ADT/StringRef.h"78#include "llvm/ADT/Twine.h"79#include "llvm/ADT/TypeSwitch.h"80#include "llvm/ADT/iterator_range.h"81#include "llvm/Analysis/AssumptionCache.h"82#include "llvm/Analysis/BasicAliasAnalysis.h"83#include "llvm/Analysis/BlockFrequencyInfo.h"84#include "llvm/Analysis/CFG.h"85#include "llvm/Analysis/CodeMetrics.h"86#include "llvm/Analysis/DemandedBits.h"87#include "llvm/Analysis/GlobalsModRef.h"88#include "llvm/Analysis/LoopAccessAnalysis.h"89#include "llvm/Analysis/LoopAnalysisManager.h"90#include "llvm/Analysis/LoopInfo.h"91#include "llvm/Analysis/LoopIterator.h"92#include "llvm/Analysis/OptimizationRemarkEmitter.h"93#include "llvm/Analysis/ProfileSummaryInfo.h"94#include "llvm/Analysis/ScalarEvolution.h"95#include "llvm/Analysis/ScalarEvolutionExpressions.h"96#include "llvm/Analysis/ScalarEvolutionPatternMatch.h"97#include "llvm/Analysis/TargetLibraryInfo.h"98#include "llvm/Analysis/TargetTransformInfo.h"99#include "llvm/Analysis/ValueTracking.h"100#include "llvm/Analysis/VectorUtils.h"101#include "llvm/IR/Attributes.h"102#include "llvm/IR/BasicBlock.h"103#include "llvm/IR/CFG.h"104#include "llvm/IR/Constant.h"105#include "llvm/IR/Constants.h"106#include "llvm/IR/DataLayout.h"107#include "llvm/IR/DebugInfo.h"108#include "llvm/IR/DebugLoc.h"109#include "llvm/IR/DerivedTypes.h"110#include "llvm/IR/DiagnosticInfo.h"111#include "llvm/IR/Dominators.h"112#include "llvm/IR/Function.h"113#include "llvm/IR/IRBuilder.h"114#include "llvm/IR/InstrTypes.h"115#include "llvm/IR/Instruction.h"116#include "llvm/IR/Instructions.h"117#include "llvm/IR/IntrinsicInst.h"118#include "llvm/IR/Intrinsics.h"119#include "llvm/IR/MDBuilder.h"120#include "llvm/IR/Metadata.h"121#include "llvm/IR/Module.h"122#include "llvm/IR/Operator.h"123#include "llvm/IR/PatternMatch.h"124#include "llvm/IR/ProfDataUtils.h"125#include "llvm/IR/Type.h"126#include "llvm/IR/Use.h"127#include "llvm/IR/User.h"128#include "llvm/IR/Value.h"129#include "llvm/IR/Verifier.h"130#include "llvm/Support/Casting.h"131#include "llvm/Support/CommandLine.h"132#include "llvm/Support/Debug.h"133#include "llvm/Support/ErrorHandling.h"134#include "llvm/Support/InstructionCost.h"135#include "llvm/Support/MathExtras.h"136#include "llvm/Support/NativeFormatting.h"137#include "llvm/Support/raw_ostream.h"138#include "llvm/Transforms/Utils/BasicBlockUtils.h"139#include "llvm/Transforms/Utils/InjectTLIMappings.h"140#include "llvm/Transforms/Utils/Local.h"141#include "llvm/Transforms/Utils/LoopSimplify.h"142#include "llvm/Transforms/Utils/LoopUtils.h"143#include "llvm/Transforms/Utils/LoopVersioning.h"144#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"145#include "llvm/Transforms/Utils/SizeOpts.h"146#include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h"147#include <algorithm>148#include <cassert>149#include <cstdint>150#include <functional>151#include <iterator>152#include <limits>153#include <memory>154#include <string>155#include <tuple>156#include <utility>157 158using namespace llvm;159using namespace SCEVPatternMatch;160 161#define LV_NAME "loop-vectorize"162#define DEBUG_TYPE LV_NAME163 164#ifndef NDEBUG165const char VerboseDebug[] = DEBUG_TYPE "-verbose";166#endif167 168STATISTIC(LoopsVectorized, "Number of loops vectorized");169STATISTIC(LoopsAnalyzed, "Number of loops analyzed for vectorization");170STATISTIC(LoopsEpilogueVectorized, "Number of epilogues vectorized");171STATISTIC(LoopsEarlyExitVectorized, "Number of early exit loops vectorized");172 173static cl::opt<bool> EnableEpilogueVectorization(174    "enable-epilogue-vectorization", cl::init(true), cl::Hidden,175    cl::desc("Enable vectorization of epilogue loops."));176 177static cl::opt<unsigned> EpilogueVectorizationForceVF(178    "epilogue-vectorization-force-VF", cl::init(1), cl::Hidden,179    cl::desc("When epilogue vectorization is enabled, and a value greater than "180             "1 is specified, forces the given VF for all applicable epilogue "181             "loops."));182 183static cl::opt<unsigned> EpilogueVectorizationMinVF(184    "epilogue-vectorization-minimum-VF", cl::Hidden,185    cl::desc("Only loops with vectorization factor equal to or larger than "186             "the specified value are considered for epilogue vectorization."));187 188/// Loops with a known constant trip count below this number are vectorized only189/// if no scalar iteration overheads are incurred.190static cl::opt<unsigned> TinyTripCountVectorThreshold(191    "vectorizer-min-trip-count", cl::init(16), cl::Hidden,192    cl::desc("Loops with a constant trip count that is smaller than this "193             "value are vectorized only if no scalar iteration overheads "194             "are incurred."));195 196static cl::opt<unsigned> VectorizeMemoryCheckThreshold(197    "vectorize-memory-check-threshold", cl::init(128), cl::Hidden,198    cl::desc("The maximum allowed number of runtime memory checks"));199 200// Option prefer-predicate-over-epilogue indicates that an epilogue is undesired,201// that predication is preferred, and this lists all options. I.e., the202// vectorizer will try to fold the tail-loop (epilogue) into the vector body203// and predicate the instructions accordingly. If tail-folding fails, there are204// different fallback strategies depending on these values:205namespace PreferPredicateTy {206  enum Option {207    ScalarEpilogue = 0,208    PredicateElseScalarEpilogue,209    PredicateOrDontVectorize210  };211} // namespace PreferPredicateTy212 213static cl::opt<PreferPredicateTy::Option> PreferPredicateOverEpilogue(214    "prefer-predicate-over-epilogue",215    cl::init(PreferPredicateTy::ScalarEpilogue),216    cl::Hidden,217    cl::desc("Tail-folding and predication preferences over creating a scalar "218             "epilogue loop."),219    cl::values(clEnumValN(PreferPredicateTy::ScalarEpilogue,220                         "scalar-epilogue",221                         "Don't tail-predicate loops, create scalar epilogue"),222              clEnumValN(PreferPredicateTy::PredicateElseScalarEpilogue,223                         "predicate-else-scalar-epilogue",224                         "prefer tail-folding, create scalar epilogue if tail "225                         "folding fails."),226              clEnumValN(PreferPredicateTy::PredicateOrDontVectorize,227                         "predicate-dont-vectorize",228                         "prefers tail-folding, don't attempt vectorization if "229                         "tail-folding fails.")));230 231static cl::opt<TailFoldingStyle> ForceTailFoldingStyle(232    "force-tail-folding-style", cl::desc("Force the tail folding style"),233    cl::init(TailFoldingStyle::None),234    cl::values(235        clEnumValN(TailFoldingStyle::None, "none", "Disable tail folding"),236        clEnumValN(237            TailFoldingStyle::Data, "data",238            "Create lane mask for data only, using active.lane.mask intrinsic"),239        clEnumValN(TailFoldingStyle::DataWithoutLaneMask,240                   "data-without-lane-mask",241                   "Create lane mask with compare/stepvector"),242        clEnumValN(TailFoldingStyle::DataAndControlFlow, "data-and-control",243                   "Create lane mask using active.lane.mask intrinsic, and use "244                   "it for both data and control flow"),245        clEnumValN(TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck,246                   "data-and-control-without-rt-check",247                   "Similar to data-and-control, but remove the runtime check"),248        clEnumValN(TailFoldingStyle::DataWithEVL, "data-with-evl",249                   "Use predicated EVL instructions for tail folding. If EVL "250                   "is unsupported, fallback to data-without-lane-mask.")));251 252cl::opt<bool> llvm::EnableWideActiveLaneMask(253    "enable-wide-lane-mask", cl::init(false), cl::Hidden,254    cl::desc("Enable use of wide lane masks when used for control flow in "255             "tail-folded loops"));256 257static cl::opt<bool> MaximizeBandwidth(258    "vectorizer-maximize-bandwidth", cl::init(false), cl::Hidden,259    cl::desc("Maximize bandwidth when selecting vectorization factor which "260             "will be determined by the smallest type in loop."));261 262static cl::opt<bool> EnableInterleavedMemAccesses(263    "enable-interleaved-mem-accesses", cl::init(false), cl::Hidden,264    cl::desc("Enable vectorization on interleaved memory accesses in a loop"));265 266/// An interleave-group may need masking if it resides in a block that needs267/// predication, or in order to mask away gaps.268static cl::opt<bool> EnableMaskedInterleavedMemAccesses(269    "enable-masked-interleaved-mem-accesses", cl::init(false), cl::Hidden,270    cl::desc("Enable vectorization on masked interleaved memory accesses in a loop"));271 272static cl::opt<unsigned> ForceTargetNumScalarRegs(273    "force-target-num-scalar-regs", cl::init(0), cl::Hidden,274    cl::desc("A flag that overrides the target's number of scalar registers."));275 276static cl::opt<unsigned> ForceTargetNumVectorRegs(277    "force-target-num-vector-regs", cl::init(0), cl::Hidden,278    cl::desc("A flag that overrides the target's number of vector registers."));279 280static cl::opt<unsigned> ForceTargetMaxScalarInterleaveFactor(281    "force-target-max-scalar-interleave", cl::init(0), cl::Hidden,282    cl::desc("A flag that overrides the target's max interleave factor for "283             "scalar loops."));284 285static cl::opt<unsigned> ForceTargetMaxVectorInterleaveFactor(286    "force-target-max-vector-interleave", cl::init(0), cl::Hidden,287    cl::desc("A flag that overrides the target's max interleave factor for "288             "vectorized loops."));289 290cl::opt<unsigned> llvm::ForceTargetInstructionCost(291    "force-target-instruction-cost", cl::init(0), cl::Hidden,292    cl::desc("A flag that overrides the target's expected cost for "293             "an instruction to a single constant value. Mostly "294             "useful for getting consistent testing."));295 296static cl::opt<bool> ForceTargetSupportsScalableVectors(297    "force-target-supports-scalable-vectors", cl::init(false), cl::Hidden,298    cl::desc(299        "Pretend that scalable vectors are supported, even if the target does "300        "not support them. This flag should only be used for testing."));301 302static cl::opt<unsigned> SmallLoopCost(303    "small-loop-cost", cl::init(20), cl::Hidden,304    cl::desc(305        "The cost of a loop that is considered 'small' by the interleaver."));306 307static cl::opt<bool> LoopVectorizeWithBlockFrequency(308    "loop-vectorize-with-block-frequency", cl::init(true), cl::Hidden,309    cl::desc("Enable the use of the block frequency analysis to access PGO "310             "heuristics minimizing code growth in cold regions and being more "311             "aggressive in hot regions."));312 313// Runtime interleave loops for load/store throughput.314static cl::opt<bool> EnableLoadStoreRuntimeInterleave(315    "enable-loadstore-runtime-interleave", cl::init(true), cl::Hidden,316    cl::desc(317        "Enable runtime interleaving until load/store ports are saturated"));318 319/// The number of stores in a loop that are allowed to need predication.320static cl::opt<unsigned> NumberOfStoresToPredicate(321    "vectorize-num-stores-pred", cl::init(1), cl::Hidden,322    cl::desc("Max number of stores to be predicated behind an if."));323 324static cl::opt<bool> EnableIndVarRegisterHeur(325    "enable-ind-var-reg-heur", cl::init(true), cl::Hidden,326    cl::desc("Count the induction variable only once when interleaving"));327 328static cl::opt<bool> EnableCondStoresVectorization(329    "enable-cond-stores-vec", cl::init(true), cl::Hidden,330    cl::desc("Enable if predication of stores during vectorization."));331 332static cl::opt<unsigned> MaxNestedScalarReductionIC(333    "max-nested-scalar-reduction-interleave", cl::init(2), cl::Hidden,334    cl::desc("The maximum interleave count to use when interleaving a scalar "335             "reduction in a nested loop."));336 337static cl::opt<bool>338    PreferInLoopReductions("prefer-inloop-reductions", cl::init(false),339                           cl::Hidden,340                           cl::desc("Prefer in-loop vector reductions, "341                                    "overriding the targets preference."));342 343static cl::opt<bool> ForceOrderedReductions(344    "force-ordered-reductions", cl::init(false), cl::Hidden,345    cl::desc("Enable the vectorisation of loops with in-order (strict) "346             "FP reductions"));347 348static cl::opt<bool> PreferPredicatedReductionSelect(349    "prefer-predicated-reduction-select", cl::init(false), cl::Hidden,350    cl::desc(351        "Prefer predicating a reduction operation over an after loop select."));352 353cl::opt<bool> llvm::EnableVPlanNativePath(354    "enable-vplan-native-path", cl::Hidden,355    cl::desc("Enable VPlan-native vectorization path with "356             "support for outer loop vectorization."));357 358cl::opt<bool>359    llvm::VerifyEachVPlan("vplan-verify-each",360#ifdef EXPENSIVE_CHECKS361                          cl::init(true),362#else363                          cl::init(false),364#endif365                          cl::Hidden,366                          cl::desc("Verfiy VPlans after VPlan transforms."));367 368// This flag enables the stress testing of the VPlan H-CFG construction in the369// VPlan-native vectorization path. It must be used in conjuction with370// -enable-vplan-native-path. -vplan-verify-hcfg can also be used to enable the371// verification of the H-CFGs built.372static cl::opt<bool> VPlanBuildStressTest(373    "vplan-build-stress-test", cl::init(false), cl::Hidden,374    cl::desc(375        "Build VPlan for every supported loop nest in the function and bail "376        "out right after the build (stress test the VPlan H-CFG construction "377        "in the VPlan-native vectorization path)."));378 379cl::opt<bool> llvm::EnableLoopInterleaving(380    "interleave-loops", cl::init(true), cl::Hidden,381    cl::desc("Enable loop interleaving in Loop vectorization passes"));382cl::opt<bool> llvm::EnableLoopVectorization(383    "vectorize-loops", cl::init(true), cl::Hidden,384    cl::desc("Run the Loop vectorization passes"));385 386static cl::opt<cl::boolOrDefault> ForceSafeDivisor(387    "force-widen-divrem-via-safe-divisor", cl::Hidden,388    cl::desc(389        "Override cost based safe divisor widening for div/rem instructions"));390 391static cl::opt<bool> UseWiderVFIfCallVariantsPresent(392    "vectorizer-maximize-bandwidth-for-vector-calls", cl::init(true),393    cl::Hidden,394    cl::desc("Try wider VFs if they enable the use of vector variants"));395 396static cl::opt<bool> EnableEarlyExitVectorization(397    "enable-early-exit-vectorization", cl::init(true), cl::Hidden,398    cl::desc(399        "Enable vectorization of early exit loops with uncountable exits."));400 401static cl::opt<bool> ConsiderRegPressure(402    "vectorizer-consider-reg-pressure", cl::init(false), cl::Hidden,403    cl::desc("Discard VFs if their register pressure is too high."));404 405// Likelyhood of bypassing the vectorized loop because there are zero trips left406// after prolog. See `emitIterationCountCheck`.407static constexpr uint32_t MinItersBypassWeights[] = {1, 127};408 409/// A helper function that returns true if the given type is irregular. The410/// type is irregular if its allocated size doesn't equal the store size of an411/// element of the corresponding vector type.412static bool hasIrregularType(Type *Ty, const DataLayout &DL) {413  // Determine if an array of N elements of type Ty is "bitcast compatible"414  // with a <N x Ty> vector.415  // This is only true if there is no padding between the array elements.416  return DL.getTypeAllocSizeInBits(Ty) != DL.getTypeSizeInBits(Ty);417}418 419/// A version of ScalarEvolution::getSmallConstantTripCount that returns an420/// ElementCount to include loops whose trip count is a function of vscale.421static ElementCount getSmallConstantTripCount(ScalarEvolution *SE,422                                              const Loop *L) {423  if (unsigned ExpectedTC = SE->getSmallConstantTripCount(L))424    return ElementCount::getFixed(ExpectedTC);425 426  const SCEV *BTC = SE->getBackedgeTakenCount(L);427  if (isa<SCEVCouldNotCompute>(BTC))428    return ElementCount::getFixed(0);429 430  const SCEV *ExitCount = SE->getTripCountFromExitCount(BTC, BTC->getType(), L);431  if (isa<SCEVVScale>(ExitCount))432    return ElementCount::getScalable(1);433 434  const APInt *Scale;435  if (match(ExitCount, m_scev_Mul(m_scev_APInt(Scale), m_SCEVVScale())))436    if (cast<SCEVMulExpr>(ExitCount)->hasNoUnsignedWrap())437      if (Scale->getActiveBits() <= 32)438        return ElementCount::getScalable(Scale->getZExtValue());439 440  return ElementCount::getFixed(0);441}442 443/// Returns "best known" trip count, which is either a valid positive trip count444/// or std::nullopt when an estimate cannot be made (including when the trip445/// count would overflow), for the specified loop \p L as defined by the446/// following procedure:447///   1) Returns exact trip count if it is known.448///   2) Returns expected trip count according to profile data if any.449///   3) Returns upper bound estimate if known, and if \p CanUseConstantMax.450///   4) Returns std::nullopt if all of the above failed.451static std::optional<ElementCount>452getSmallBestKnownTC(PredicatedScalarEvolution &PSE, Loop *L,453                    bool CanUseConstantMax = true) {454  // Check if exact trip count is known.455  if (auto ExpectedTC = getSmallConstantTripCount(PSE.getSE(), L))456    return ExpectedTC;457 458  // Check if there is an expected trip count available from profile data.459  if (LoopVectorizeWithBlockFrequency)460    if (auto EstimatedTC = getLoopEstimatedTripCount(L))461      return ElementCount::getFixed(*EstimatedTC);462 463  if (!CanUseConstantMax)464    return std::nullopt;465 466  // Check if upper bound estimate is known.467  if (unsigned ExpectedTC = PSE.getSmallConstantMaxTripCount())468    return ElementCount::getFixed(ExpectedTC);469 470  return std::nullopt;471}472 473namespace {474// Forward declare GeneratedRTChecks.475class GeneratedRTChecks;476 477using SCEV2ValueTy = DenseMap<const SCEV *, Value *>;478} // namespace479 480namespace llvm {481 482AnalysisKey ShouldRunExtraVectorPasses::Key;483 484/// InnerLoopVectorizer vectorizes loops which contain only one basic485/// block to a specified vectorization factor (VF).486/// This class performs the widening of scalars into vectors, or multiple487/// scalars. This class also implements the following features:488/// * It inserts an epilogue loop for handling loops that don't have iteration489///   counts that are known to be a multiple of the vectorization factor.490/// * It handles the code generation for reduction variables.491/// * Scalarization (implementation using scalars) of un-vectorizable492///   instructions.493/// InnerLoopVectorizer does not perform any vectorization-legality494/// checks, and relies on the caller to check for the different legality495/// aspects. The InnerLoopVectorizer relies on the496/// LoopVectorizationLegality class to provide information about the induction497/// and reduction variables that were found to a given vectorization factor.498class InnerLoopVectorizer {499public:500  InnerLoopVectorizer(Loop *OrigLoop, PredicatedScalarEvolution &PSE,501                      LoopInfo *LI, DominatorTree *DT,502                      const TargetTransformInfo *TTI, AssumptionCache *AC,503                      ElementCount VecWidth, unsigned UnrollFactor,504                      LoopVectorizationCostModel *CM,505                      GeneratedRTChecks &RTChecks, VPlan &Plan)506      : OrigLoop(OrigLoop), PSE(PSE), LI(LI), DT(DT), TTI(TTI), AC(AC),507        VF(VecWidth), UF(UnrollFactor), Builder(PSE.getSE()->getContext()),508        Cost(CM), RTChecks(RTChecks), Plan(Plan),509        VectorPHVPBB(cast<VPBasicBlock>(510            Plan.getVectorLoopRegion()->getSinglePredecessor())) {}511 512  virtual ~InnerLoopVectorizer() = default;513 514  /// Creates a basic block for the scalar preheader. Both515  /// EpilogueVectorizerMainLoop and EpilogueVectorizerEpilogueLoop overwrite516  /// the method to create additional blocks and checks needed for epilogue517  /// vectorization.518  virtual BasicBlock *createVectorizedLoopSkeleton();519 520  /// Fix the vectorized code, taking care of header phi's, and more.521  void fixVectorizedLoop(VPTransformState &State);522 523  /// Fix the non-induction PHIs in \p Plan.524  void fixNonInductionPHIs(VPTransformState &State);525 526  /// Returns the original loop trip count.527  Value *getTripCount() const { return TripCount; }528 529  /// Used to set the trip count after ILV's construction and after the530  /// preheader block has been executed. Note that this always holds the trip531  /// count of the original loop for both main loop and epilogue vectorization.532  void setTripCount(Value *TC) { TripCount = TC; }533 534protected:535  friend class LoopVectorizationPlanner;536 537  /// Create and return a new IR basic block for the scalar preheader whose name538  /// is prefixed with \p Prefix.539  BasicBlock *createScalarPreheader(StringRef Prefix);540 541  /// Allow subclasses to override and print debug traces before/after vplan542  /// execution, when trace information is requested.543  virtual void printDebugTracesAtStart() {}544  virtual void printDebugTracesAtEnd() {}545 546  /// The original loop.547  Loop *OrigLoop;548 549  /// A wrapper around ScalarEvolution used to add runtime SCEV checks. Applies550  /// dynamic knowledge to simplify SCEV expressions and converts them to a551  /// more usable form.552  PredicatedScalarEvolution &PSE;553 554  /// Loop Info.555  LoopInfo *LI;556 557  /// Dominator Tree.558  DominatorTree *DT;559 560  /// Target Transform Info.561  const TargetTransformInfo *TTI;562 563  /// Assumption Cache.564  AssumptionCache *AC;565 566  /// The vectorization SIMD factor to use. Each vector will have this many567  /// vector elements.568  ElementCount VF;569 570  /// The vectorization unroll factor to use. Each scalar is vectorized to this571  /// many different vector instructions.572  unsigned UF;573 574  /// The builder that we use575  IRBuilder<> Builder;576 577  // --- Vectorization state ---578 579  /// Trip count of the original loop.580  Value *TripCount = nullptr;581 582  /// The profitablity analysis.583  LoopVectorizationCostModel *Cost;584 585  /// Structure to hold information about generated runtime checks, responsible586  /// for cleaning the checks, if vectorization turns out unprofitable.587  GeneratedRTChecks &RTChecks;588 589  VPlan &Plan;590 591  /// The vector preheader block of \p Plan, used as target for check blocks592  /// introduced during skeleton creation.593  VPBasicBlock *VectorPHVPBB;594};595 596/// Encapsulate information regarding vectorization of a loop and its epilogue.597/// This information is meant to be updated and used across two stages of598/// epilogue vectorization.599struct EpilogueLoopVectorizationInfo {600  ElementCount MainLoopVF = ElementCount::getFixed(0);601  unsigned MainLoopUF = 0;602  ElementCount EpilogueVF = ElementCount::getFixed(0);603  unsigned EpilogueUF = 0;604  BasicBlock *MainLoopIterationCountCheck = nullptr;605  BasicBlock *EpilogueIterationCountCheck = nullptr;606  Value *TripCount = nullptr;607  Value *VectorTripCount = nullptr;608  VPlan &EpiloguePlan;609 610  EpilogueLoopVectorizationInfo(ElementCount MVF, unsigned MUF,611                                ElementCount EVF, unsigned EUF,612                                VPlan &EpiloguePlan)613      : MainLoopVF(MVF), MainLoopUF(MUF), EpilogueVF(EVF), EpilogueUF(EUF),614        EpiloguePlan(EpiloguePlan) {615    assert(EUF == 1 &&616           "A high UF for the epilogue loop is likely not beneficial.");617  }618};619 620/// An extension of the inner loop vectorizer that creates a skeleton for a621/// vectorized loop that has its epilogue (residual) also vectorized.622/// The idea is to run the vplan on a given loop twice, firstly to setup the623/// skeleton and vectorize the main loop, and secondly to complete the skeleton624/// from the first step and vectorize the epilogue.  This is achieved by625/// deriving two concrete strategy classes from this base class and invoking626/// them in succession from the loop vectorizer planner.627class InnerLoopAndEpilogueVectorizer : public InnerLoopVectorizer {628public:629  InnerLoopAndEpilogueVectorizer(630      Loop *OrigLoop, PredicatedScalarEvolution &PSE, LoopInfo *LI,631      DominatorTree *DT, const TargetTransformInfo *TTI, AssumptionCache *AC,632      EpilogueLoopVectorizationInfo &EPI, LoopVectorizationCostModel *CM,633      GeneratedRTChecks &Checks, VPlan &Plan, ElementCount VecWidth,634      ElementCount MinProfitableTripCount, unsigned UnrollFactor)635      : InnerLoopVectorizer(OrigLoop, PSE, LI, DT, TTI, AC, VecWidth,636                            UnrollFactor, CM, Checks, Plan),637        EPI(EPI), MinProfitableTripCount(MinProfitableTripCount) {}638 639  /// Holds and updates state information required to vectorize the main loop640  /// and its epilogue in two separate passes. This setup helps us avoid641  /// regenerating and recomputing runtime safety checks. It also helps us to642  /// shorten the iteration-count-check path length for the cases where the643  /// iteration count of the loop is so small that the main vector loop is644  /// completely skipped.645  EpilogueLoopVectorizationInfo &EPI;646 647protected:648  ElementCount MinProfitableTripCount;649};650 651/// A specialized derived class of inner loop vectorizer that performs652/// vectorization of *main* loops in the process of vectorizing loops and their653/// epilogues.654class EpilogueVectorizerMainLoop : public InnerLoopAndEpilogueVectorizer {655public:656  EpilogueVectorizerMainLoop(Loop *OrigLoop, PredicatedScalarEvolution &PSE,657                             LoopInfo *LI, DominatorTree *DT,658                             const TargetTransformInfo *TTI,659                             AssumptionCache *AC,660                             EpilogueLoopVectorizationInfo &EPI,661                             LoopVectorizationCostModel *CM,662                             GeneratedRTChecks &Check, VPlan &Plan)663      : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TTI, AC, EPI, CM,664                                       Check, Plan, EPI.MainLoopVF,665                                       EPI.MainLoopVF, EPI.MainLoopUF) {}666  /// Implements the interface for creating a vectorized skeleton using the667  /// *main loop* strategy (i.e., the first pass of VPlan execution).668  BasicBlock *createVectorizedLoopSkeleton() final;669 670protected:671  /// Introduces a new VPIRBasicBlock for \p CheckIRBB to Plan between the672  /// vector preheader and its predecessor, also connecting the new block to the673  /// scalar preheader.674  void introduceCheckBlockInVPlan(BasicBlock *CheckIRBB);675 676  // Create a check to see if the main vector loop should be executed677  Value *createIterationCountCheck(BasicBlock *VectorPH, ElementCount VF,678                                   unsigned UF) const;679 680  /// Emits an iteration count bypass check once for the main loop (when \p681  /// ForEpilogue is false) and once for the epilogue loop (when \p682  /// ForEpilogue is true).683  BasicBlock *emitIterationCountCheck(BasicBlock *VectorPH, BasicBlock *Bypass,684                                      bool ForEpilogue);685  void printDebugTracesAtStart() override;686  void printDebugTracesAtEnd() override;687};688 689// A specialized derived class of inner loop vectorizer that performs690// vectorization of *epilogue* loops in the process of vectorizing loops and691// their epilogues.692class EpilogueVectorizerEpilogueLoop : public InnerLoopAndEpilogueVectorizer {693public:694  EpilogueVectorizerEpilogueLoop(Loop *OrigLoop, PredicatedScalarEvolution &PSE,695                                 LoopInfo *LI, DominatorTree *DT,696                                 const TargetTransformInfo *TTI,697                                 AssumptionCache *AC,698                                 EpilogueLoopVectorizationInfo &EPI,699                                 LoopVectorizationCostModel *CM,700                                 GeneratedRTChecks &Checks, VPlan &Plan)701      : InnerLoopAndEpilogueVectorizer(OrigLoop, PSE, LI, DT, TTI, AC, EPI, CM,702                                       Checks, Plan, EPI.EpilogueVF,703                                       EPI.EpilogueVF, EPI.EpilogueUF) {}704  /// Implements the interface for creating a vectorized skeleton using the705  /// *epilogue loop* strategy (i.e., the second pass of VPlan execution).706  BasicBlock *createVectorizedLoopSkeleton() final;707 708protected:709  void printDebugTracesAtStart() override;710  void printDebugTracesAtEnd() override;711};712} // end namespace llvm713 714/// Look for a meaningful debug location on the instruction or its operands.715static DebugLoc getDebugLocFromInstOrOperands(Instruction *I) {716  if (!I)717    return DebugLoc::getUnknown();718 719  DebugLoc Empty;720  if (I->getDebugLoc() != Empty)721    return I->getDebugLoc();722 723  for (Use &Op : I->operands()) {724    if (Instruction *OpInst = dyn_cast<Instruction>(Op))725      if (OpInst->getDebugLoc() != Empty)726        return OpInst->getDebugLoc();727  }728 729  return I->getDebugLoc();730}731 732/// Write a \p DebugMsg about vectorization to the debug output stream. If \p I733/// is passed, the message relates to that particular instruction.734#ifndef NDEBUG735static void debugVectorizationMessage(const StringRef Prefix,736                                      const StringRef DebugMsg,737                                      Instruction *I) {738  dbgs() << "LV: " << Prefix << DebugMsg;739  if (I != nullptr)740    dbgs() << " " << *I;741  else742    dbgs() << '.';743  dbgs() << '\n';744}745#endif746 747/// Create an analysis remark that explains why vectorization failed748///749/// \p PassName is the name of the pass (e.g. can be AlwaysPrint).  \p750/// RemarkName is the identifier for the remark.  If \p I is passed it is an751/// instruction that prevents vectorization.  Otherwise \p TheLoop is used for752/// the location of the remark. If \p DL is passed, use it as debug location for753/// the remark. \return the remark object that can be streamed to.754static OptimizationRemarkAnalysis755createLVAnalysis(const char *PassName, StringRef RemarkName, Loop *TheLoop,756                 Instruction *I, DebugLoc DL = {}) {757  BasicBlock *CodeRegion = I ? I->getParent() : TheLoop->getHeader();758  // If debug location is attached to the instruction, use it. Otherwise if DL759  // was not provided, use the loop's.760  if (I && I->getDebugLoc())761    DL = I->getDebugLoc();762  else if (!DL)763    DL = TheLoop->getStartLoc();764 765  return OptimizationRemarkAnalysis(PassName, RemarkName, DL, CodeRegion);766}767 768namespace llvm {769 770/// Return a value for Step multiplied by VF.771Value *createStepForVF(IRBuilderBase &B, Type *Ty, ElementCount VF,772                       int64_t Step) {773  assert(Ty->isIntegerTy() && "Expected an integer step");774  ElementCount VFxStep = VF.multiplyCoefficientBy(Step);775  assert(isPowerOf2_64(VF.getKnownMinValue()) && "must pass power-of-2 VF");776  if (VF.isScalable() && isPowerOf2_64(Step)) {777    return B.CreateShl(778        B.CreateVScale(Ty),779        ConstantInt::get(Ty, Log2_64(VFxStep.getKnownMinValue())), "", true);780  }781  return B.CreateElementCount(Ty, VFxStep);782}783 784/// Return the runtime value for VF.785Value *getRuntimeVF(IRBuilderBase &B, Type *Ty, ElementCount VF) {786  return B.CreateElementCount(Ty, VF);787}788 789void reportVectorizationFailure(const StringRef DebugMsg,790                                const StringRef OREMsg, const StringRef ORETag,791                                OptimizationRemarkEmitter *ORE, Loop *TheLoop,792                                Instruction *I) {793  LLVM_DEBUG(debugVectorizationMessage("Not vectorizing: ", DebugMsg, I));794  LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE);795  ORE->emit(796      createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop, I)797      << "loop not vectorized: " << OREMsg);798}799 800/// Reports an informative message: print \p Msg for debugging purposes as well801/// as an optimization remark. Uses either \p I as location of the remark, or802/// otherwise \p TheLoop. If \p DL is passed, use it as debug location for the803/// remark. If \p DL is passed, use it as debug location for the remark.804static void reportVectorizationInfo(const StringRef Msg, const StringRef ORETag,805                                    OptimizationRemarkEmitter *ORE,806                                    Loop *TheLoop, Instruction *I = nullptr,807                                    DebugLoc DL = {}) {808  LLVM_DEBUG(debugVectorizationMessage("", Msg, I));809  LoopVectorizeHints Hints(TheLoop, true /* doesn't matter */, *ORE);810  ORE->emit(createLVAnalysis(Hints.vectorizeAnalysisPassName(), ORETag, TheLoop,811                             I, DL)812            << Msg);813}814 815/// Report successful vectorization of the loop. In case an outer loop is816/// vectorized, prepend "outer" to the vectorization remark.817static void reportVectorization(OptimizationRemarkEmitter *ORE, Loop *TheLoop,818                                VectorizationFactor VF, unsigned IC) {819  LLVM_DEBUG(debugVectorizationMessage(820      "Vectorizing: ", TheLoop->isInnermost() ? "innermost loop" : "outer loop",821      nullptr));822  StringRef LoopType = TheLoop->isInnermost() ? "" : "outer ";823  ORE->emit([&]() {824    return OptimizationRemark(LV_NAME, "Vectorized", TheLoop->getStartLoc(),825                              TheLoop->getHeader())826           << "vectorized " << LoopType << "loop (vectorization width: "827           << ore::NV("VectorizationFactor", VF.Width)828           << ", interleaved count: " << ore::NV("InterleaveCount", IC) << ")";829  });830}831 832} // end namespace llvm833 834namespace llvm {835 836// Loop vectorization cost-model hints how the scalar epilogue loop should be837// lowered.838enum ScalarEpilogueLowering {839 840  // The default: allowing scalar epilogues.841  CM_ScalarEpilogueAllowed,842 843  // Vectorization with OptForSize: don't allow epilogues.844  CM_ScalarEpilogueNotAllowedOptSize,845 846  // A special case of vectorisation with OptForSize: loops with a very small847  // trip count are considered for vectorization under OptForSize, thereby848  // making sure the cost of their loop body is dominant, free of runtime849  // guards and scalar iteration overheads.850  CM_ScalarEpilogueNotAllowedLowTripLoop,851 852  // Loop hint predicate indicating an epilogue is undesired.853  CM_ScalarEpilogueNotNeededUsePredicate,854 855  // Directive indicating we must either tail fold or not vectorize856  CM_ScalarEpilogueNotAllowedUsePredicate857};858 859/// LoopVectorizationCostModel - estimates the expected speedups due to860/// vectorization.861/// In many cases vectorization is not profitable. This can happen because of862/// a number of reasons. In this class we mainly attempt to predict the863/// expected speedup/slowdowns due to the supported instruction set. We use the864/// TargetTransformInfo to query the different backends for the cost of865/// different operations.866class LoopVectorizationCostModel {867  friend class LoopVectorizationPlanner;868 869public:870  LoopVectorizationCostModel(ScalarEpilogueLowering SEL, Loop *L,871                             PredicatedScalarEvolution &PSE, LoopInfo *LI,872                             LoopVectorizationLegality *Legal,873                             const TargetTransformInfo &TTI,874                             const TargetLibraryInfo *TLI, DemandedBits *DB,875                             AssumptionCache *AC,876                             OptimizationRemarkEmitter *ORE, const Function *F,877                             const LoopVectorizeHints *Hints,878                             InterleavedAccessInfo &IAI, bool OptForSize)879      : ScalarEpilogueStatus(SEL), TheLoop(L), PSE(PSE), LI(LI), Legal(Legal),880        TTI(TTI), TLI(TLI), DB(DB), AC(AC), ORE(ORE), TheFunction(F),881        Hints(Hints), InterleaveInfo(IAI), OptForSize(OptForSize) {882    if (TTI.supportsScalableVectors() || ForceTargetSupportsScalableVectors)883      initializeVScaleForTuning();884    CostKind = F->hasMinSize() ? TTI::TCK_CodeSize : TTI::TCK_RecipThroughput;885  }886 887  /// \return An upper bound for the vectorization factors (both fixed and888  /// scalable). If the factors are 0, vectorization and interleaving should be889  /// avoided up front.890  FixedScalableVFPair computeMaxVF(ElementCount UserVF, unsigned UserIC);891 892  /// \return True if runtime checks are required for vectorization, and false893  /// otherwise.894  bool runtimeChecksRequired();895 896  /// Setup cost-based decisions for user vectorization factor.897  /// \return true if the UserVF is a feasible VF to be chosen.898  bool selectUserVectorizationFactor(ElementCount UserVF) {899    collectNonVectorizedAndSetWideningDecisions(UserVF);900    return expectedCost(UserVF).isValid();901  }902 903  /// \return True if maximizing vector bandwidth is enabled by the target or904  /// user options, for the given register kind.905  bool useMaxBandwidth(TargetTransformInfo::RegisterKind RegKind);906 907  /// \return True if register pressure should be considered for the given VF.908  bool shouldConsiderRegPressureForVF(ElementCount VF);909 910  /// \return The size (in bits) of the smallest and widest types in the code911  /// that needs to be vectorized. We ignore values that remain scalar such as912  /// 64 bit loop indices.913  std::pair<unsigned, unsigned> getSmallestAndWidestTypes();914 915  /// Memory access instruction may be vectorized in more than one way.916  /// Form of instruction after vectorization depends on cost.917  /// This function takes cost-based decisions for Load/Store instructions918  /// and collects them in a map. This decisions map is used for building919  /// the lists of loop-uniform and loop-scalar instructions.920  /// The calculated cost is saved with widening decision in order to921  /// avoid redundant calculations.922  void setCostBasedWideningDecision(ElementCount VF);923 924  /// A call may be vectorized in different ways depending on whether we have925  /// vectorized variants available and whether the target supports masking.926  /// This function analyzes all calls in the function at the supplied VF,927  /// makes a decision based on the costs of available options, and stores that928  /// decision in a map for use in planning and plan execution.929  void setVectorizedCallDecision(ElementCount VF);930 931  /// Collect values we want to ignore in the cost model.932  void collectValuesToIgnore();933 934  /// Collect all element types in the loop for which widening is needed.935  void collectElementTypesForWidening();936 937  /// Split reductions into those that happen in the loop, and those that happen938  /// outside. In loop reductions are collected into InLoopReductions.939  void collectInLoopReductions();940 941  /// Returns true if we should use strict in-order reductions for the given942  /// RdxDesc. This is true if the -enable-strict-reductions flag is passed,943  /// the IsOrdered flag of RdxDesc is set and we do not allow reordering944  /// of FP operations.945  bool useOrderedReductions(const RecurrenceDescriptor &RdxDesc) const {946    return !Hints->allowReordering() && RdxDesc.isOrdered();947  }948 949  /// \returns The smallest bitwidth each instruction can be represented with.950  /// The vector equivalents of these instructions should be truncated to this951  /// type.952  const MapVector<Instruction *, uint64_t> &getMinimalBitwidths() const {953    return MinBWs;954  }955 956  /// \returns True if it is more profitable to scalarize instruction \p I for957  /// vectorization factor \p VF.958  bool isProfitableToScalarize(Instruction *I, ElementCount VF) const {959    assert(VF.isVector() &&960           "Profitable to scalarize relevant only for VF > 1.");961    assert(962        TheLoop->isInnermost() &&963        "cost-model should not be used for outer loops (in VPlan-native path)");964 965    auto Scalars = InstsToScalarize.find(VF);966    assert(Scalars != InstsToScalarize.end() &&967           "VF not yet analyzed for scalarization profitability");968    return Scalars->second.contains(I);969  }970 971  /// Returns true if \p I is known to be uniform after vectorization.972  bool isUniformAfterVectorization(Instruction *I, ElementCount VF) const {973    assert(974        TheLoop->isInnermost() &&975        "cost-model should not be used for outer loops (in VPlan-native path)");976    // Pseudo probe needs to be duplicated for each unrolled iteration and977    // vector lane so that profiled loop trip count can be accurately978    // accumulated instead of being under counted.979    if (isa<PseudoProbeInst>(I))980      return false;981 982    if (VF.isScalar())983      return true;984 985    auto UniformsPerVF = Uniforms.find(VF);986    assert(UniformsPerVF != Uniforms.end() &&987           "VF not yet analyzed for uniformity");988    return UniformsPerVF->second.count(I);989  }990 991  /// Returns true if \p I is known to be scalar after vectorization.992  bool isScalarAfterVectorization(Instruction *I, ElementCount VF) const {993    assert(994        TheLoop->isInnermost() &&995        "cost-model should not be used for outer loops (in VPlan-native path)");996    if (VF.isScalar())997      return true;998 999    auto ScalarsPerVF = Scalars.find(VF);1000    assert(ScalarsPerVF != Scalars.end() &&1001           "Scalar values are not calculated for VF");1002    return ScalarsPerVF->second.count(I);1003  }1004 1005  /// \returns True if instruction \p I can be truncated to a smaller bitwidth1006  /// for vectorization factor \p VF.1007  bool canTruncateToMinimalBitwidth(Instruction *I, ElementCount VF) const {1008    // Truncs must truncate at most to their destination type.1009    if (isa_and_nonnull<TruncInst>(I) && MinBWs.contains(I) &&1010        I->getType()->getScalarSizeInBits() < MinBWs.lookup(I))1011      return false;1012    return VF.isVector() && MinBWs.contains(I) &&1013           !isProfitableToScalarize(I, VF) &&1014           !isScalarAfterVectorization(I, VF);1015  }1016 1017  /// Decision that was taken during cost calculation for memory instruction.1018  enum InstWidening {1019    CM_Unknown,1020    CM_Widen,         // For consecutive accesses with stride +1.1021    CM_Widen_Reverse, // For consecutive accesses with stride -1.1022    CM_Interleave,1023    CM_GatherScatter,1024    CM_Scalarize,1025    CM_VectorCall,1026    CM_IntrinsicCall1027  };1028 1029  /// Save vectorization decision \p W and \p Cost taken by the cost model for1030  /// instruction \p I and vector width \p VF.1031  void setWideningDecision(Instruction *I, ElementCount VF, InstWidening W,1032                           InstructionCost Cost) {1033    assert(VF.isVector() && "Expected VF >=2");1034    WideningDecisions[{I, VF}] = {W, Cost};1035  }1036 1037  /// Save vectorization decision \p W and \p Cost taken by the cost model for1038  /// interleaving group \p Grp and vector width \p VF.1039  void setWideningDecision(const InterleaveGroup<Instruction> *Grp,1040                           ElementCount VF, InstWidening W,1041                           InstructionCost Cost) {1042    assert(VF.isVector() && "Expected VF >=2");1043    /// Broadcast this decicion to all instructions inside the group.1044    /// When interleaving, the cost will only be assigned one instruction, the1045    /// insert position. For other cases, add the appropriate fraction of the1046    /// total cost to each instruction. This ensures accurate costs are used,1047    /// even if the insert position instruction is not used.1048    InstructionCost InsertPosCost = Cost;1049    InstructionCost OtherMemberCost = 0;1050    if (W != CM_Interleave)1051      OtherMemberCost = InsertPosCost = Cost / Grp->getNumMembers();1052    ;1053    for (unsigned Idx = 0; Idx < Grp->getFactor(); ++Idx) {1054      if (auto *I = Grp->getMember(Idx)) {1055        if (Grp->getInsertPos() == I)1056          WideningDecisions[{I, VF}] = {W, InsertPosCost};1057        else1058          WideningDecisions[{I, VF}] = {W, OtherMemberCost};1059      }1060    }1061  }1062 1063  /// Return the cost model decision for the given instruction \p I and vector1064  /// width \p VF. Return CM_Unknown if this instruction did not pass1065  /// through the cost modeling.1066  InstWidening getWideningDecision(Instruction *I, ElementCount VF) const {1067    assert(VF.isVector() && "Expected VF to be a vector VF");1068    assert(1069        TheLoop->isInnermost() &&1070        "cost-model should not be used for outer loops (in VPlan-native path)");1071 1072    std::pair<Instruction *, ElementCount> InstOnVF(I, VF);1073    auto Itr = WideningDecisions.find(InstOnVF);1074    if (Itr == WideningDecisions.end())1075      return CM_Unknown;1076    return Itr->second.first;1077  }1078 1079  /// Return the vectorization cost for the given instruction \p I and vector1080  /// width \p VF.1081  InstructionCost getWideningCost(Instruction *I, ElementCount VF) {1082    assert(VF.isVector() && "Expected VF >=2");1083    std::pair<Instruction *, ElementCount> InstOnVF(I, VF);1084    assert(WideningDecisions.contains(InstOnVF) &&1085           "The cost is not calculated");1086    return WideningDecisions[InstOnVF].second;1087  }1088 1089  struct CallWideningDecision {1090    InstWidening Kind;1091    Function *Variant;1092    Intrinsic::ID IID;1093    std::optional<unsigned> MaskPos;1094    InstructionCost Cost;1095  };1096 1097  void setCallWideningDecision(CallInst *CI, ElementCount VF, InstWidening Kind,1098                               Function *Variant, Intrinsic::ID IID,1099                               std::optional<unsigned> MaskPos,1100                               InstructionCost Cost) {1101    assert(!VF.isScalar() && "Expected vector VF");1102    CallWideningDecisions[{CI, VF}] = {Kind, Variant, IID, MaskPos, Cost};1103  }1104 1105  CallWideningDecision getCallWideningDecision(CallInst *CI,1106                                               ElementCount VF) const {1107    assert(!VF.isScalar() && "Expected vector VF");1108    auto I = CallWideningDecisions.find({CI, VF});1109    if (I == CallWideningDecisions.end())1110      return {CM_Unknown, nullptr, Intrinsic::not_intrinsic, std::nullopt, 0};1111    return I->second;1112  }1113 1114  /// Return True if instruction \p I is an optimizable truncate whose operand1115  /// is an induction variable. Such a truncate will be removed by adding a new1116  /// induction variable with the destination type.1117  bool isOptimizableIVTruncate(Instruction *I, ElementCount VF) {1118    // If the instruction is not a truncate, return false.1119    auto *Trunc = dyn_cast<TruncInst>(I);1120    if (!Trunc)1121      return false;1122 1123    // Get the source and destination types of the truncate.1124    Type *SrcTy = toVectorTy(Trunc->getSrcTy(), VF);1125    Type *DestTy = toVectorTy(Trunc->getDestTy(), VF);1126 1127    // If the truncate is free for the given types, return false. Replacing a1128    // free truncate with an induction variable would add an induction variable1129    // update instruction to each iteration of the loop. We exclude from this1130    // check the primary induction variable since it will need an update1131    // instruction regardless.1132    Value *Op = Trunc->getOperand(0);1133    if (Op != Legal->getPrimaryInduction() && TTI.isTruncateFree(SrcTy, DestTy))1134      return false;1135 1136    // If the truncated value is not an induction variable, return false.1137    return Legal->isInductionPhi(Op);1138  }1139 1140  /// Collects the instructions to scalarize for each predicated instruction in1141  /// the loop.1142  void collectInstsToScalarize(ElementCount VF);1143 1144  /// Collect values that will not be widened, including Uniforms, Scalars, and1145  /// Instructions to Scalarize for the given \p VF.1146  /// The sets depend on CM decision for Load/Store instructions1147  /// that may be vectorized as interleave, gather-scatter or scalarized.1148  /// Also make a decision on what to do about call instructions in the loop1149  /// at that VF -- scalarize, call a known vector routine, or call a1150  /// vector intrinsic.1151  void collectNonVectorizedAndSetWideningDecisions(ElementCount VF) {1152    // Do the analysis once.1153    if (VF.isScalar() || Uniforms.contains(VF))1154      return;1155    setCostBasedWideningDecision(VF);1156    collectLoopUniforms(VF);1157    setVectorizedCallDecision(VF);1158    collectLoopScalars(VF);1159    collectInstsToScalarize(VF);1160  }1161 1162  /// Returns true if the target machine supports masked store operation1163  /// for the given \p DataType and kind of access to \p Ptr.1164  bool isLegalMaskedStore(Type *DataType, Value *Ptr, Align Alignment,1165                          unsigned AddressSpace) const {1166    return Legal->isConsecutivePtr(DataType, Ptr) &&1167           TTI.isLegalMaskedStore(DataType, Alignment, AddressSpace);1168  }1169 1170  /// Returns true if the target machine supports masked load operation1171  /// for the given \p DataType and kind of access to \p Ptr.1172  bool isLegalMaskedLoad(Type *DataType, Value *Ptr, Align Alignment,1173                         unsigned AddressSpace) const {1174    return Legal->isConsecutivePtr(DataType, Ptr) &&1175           TTI.isLegalMaskedLoad(DataType, Alignment, AddressSpace);1176  }1177 1178  /// Returns true if the target machine can represent \p V as a masked gather1179  /// or scatter operation.1180  bool isLegalGatherOrScatter(Value *V, ElementCount VF) {1181    bool LI = isa<LoadInst>(V);1182    bool SI = isa<StoreInst>(V);1183    if (!LI && !SI)1184      return false;1185    auto *Ty = getLoadStoreType(V);1186    Align Align = getLoadStoreAlignment(V);1187    if (VF.isVector())1188      Ty = VectorType::get(Ty, VF);1189    return (LI && TTI.isLegalMaskedGather(Ty, Align)) ||1190           (SI && TTI.isLegalMaskedScatter(Ty, Align));1191  }1192 1193  /// Returns true if the target machine supports all of the reduction1194  /// variables found for the given VF.1195  bool canVectorizeReductions(ElementCount VF) const {1196    return (all_of(Legal->getReductionVars(), [&](auto &Reduction) -> bool {1197      const RecurrenceDescriptor &RdxDesc = Reduction.second;1198      return TTI.isLegalToVectorizeReduction(RdxDesc, VF);1199    }));1200  }1201 1202  /// Given costs for both strategies, return true if the scalar predication1203  /// lowering should be used for div/rem.  This incorporates an override1204  /// option so it is not simply a cost comparison.1205  bool isDivRemScalarWithPredication(InstructionCost ScalarCost,1206                                     InstructionCost SafeDivisorCost) const {1207    switch (ForceSafeDivisor) {1208    case cl::BOU_UNSET:1209      return ScalarCost < SafeDivisorCost;1210    case cl::BOU_TRUE:1211      return false;1212    case cl::BOU_FALSE:1213      return true;1214    }1215    llvm_unreachable("impossible case value");1216  }1217 1218  /// Returns true if \p I is an instruction which requires predication and1219  /// for which our chosen predication strategy is scalarization (i.e. we1220  /// don't have an alternate strategy such as masking available).1221  /// \p VF is the vectorization factor that will be used to vectorize \p I.1222  bool isScalarWithPredication(Instruction *I, ElementCount VF) const;1223 1224  /// Returns true if \p I is an instruction that needs to be predicated1225  /// at runtime.  The result is independent of the predication mechanism.1226  /// Superset of instructions that return true for isScalarWithPredication.1227  bool isPredicatedInst(Instruction *I) const;1228 1229  /// A helper function that returns how much we should divide the cost of a1230  /// predicated block by. Typically this is the reciprocal of the block1231  /// probability, i.e. if we return X we are assuming the predicated block will1232  /// execute once for every X iterations of the loop header so the block should1233  /// only contribute 1/X of its cost to the total cost calculation, but when1234  /// optimizing for code size it will just be 1 as code size costs don't depend1235  /// on execution probabilities.1236  ///1237  /// TODO: We should use actual block probability here, if available.1238  /// Currently, we always assume predicated blocks have a 50% chance of1239  /// executing, apart from blocks that are only predicated due to tail folding.1240  inline unsigned1241  getPredBlockCostDivisor(TargetTransformInfo::TargetCostKind CostKind,1242                          BasicBlock *BB) const {1243    // If a block wasn't originally predicated but was predicated due to1244    // e.g. tail folding, don't divide the cost. Tail folded loops may still be1245    // predicated in the final vector loop iteration, but for most loops that1246    // don't have low trip counts we can expect their probability to be close to1247    // zero.1248    if (!Legal->blockNeedsPredication(BB))1249      return 1;1250    return CostKind == TTI::TCK_CodeSize ? 1 : 2;1251  }1252 1253  /// Return the costs for our two available strategies for lowering a1254  /// div/rem operation which requires speculating at least one lane.1255  /// First result is for scalarization (will be invalid for scalable1256  /// vectors); second is for the safe-divisor strategy.1257  std::pair<InstructionCost, InstructionCost>1258  getDivRemSpeculationCost(Instruction *I,1259                           ElementCount VF) const;1260 1261  /// Returns true if \p I is a memory instruction with consecutive memory1262  /// access that can be widened.1263  bool memoryInstructionCanBeWidened(Instruction *I, ElementCount VF);1264 1265  /// Returns true if \p I is a memory instruction in an interleaved-group1266  /// of memory accesses that can be vectorized with wide vector loads/stores1267  /// and shuffles.1268  bool interleavedAccessCanBeWidened(Instruction *I, ElementCount VF) const;1269 1270  /// Check if \p Instr belongs to any interleaved access group.1271  bool isAccessInterleaved(Instruction *Instr) const {1272    return InterleaveInfo.isInterleaved(Instr);1273  }1274 1275  /// Get the interleaved access group that \p Instr belongs to.1276  const InterleaveGroup<Instruction> *1277  getInterleavedAccessGroup(Instruction *Instr) const {1278    return InterleaveInfo.getInterleaveGroup(Instr);1279  }1280 1281  /// Returns true if we're required to use a scalar epilogue for at least1282  /// the final iteration of the original loop.1283  bool requiresScalarEpilogue(bool IsVectorizing) const {1284    if (!isScalarEpilogueAllowed()) {1285      LLVM_DEBUG(dbgs() << "LV: Loop does not require scalar epilogue\n");1286      return false;1287    }1288    // If we might exit from anywhere but the latch and early exit vectorization1289    // is disabled, we must run the exiting iteration in scalar form.1290    if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch() &&1291        !(EnableEarlyExitVectorization && Legal->hasUncountableEarlyExit())) {1292      LLVM_DEBUG(dbgs() << "LV: Loop requires scalar epilogue: not exiting "1293                           "from latch block\n");1294      return true;1295    }1296    if (IsVectorizing && InterleaveInfo.requiresScalarEpilogue()) {1297      LLVM_DEBUG(dbgs() << "LV: Loop requires scalar epilogue: "1298                           "interleaved group requires scalar epilogue\n");1299      return true;1300    }1301    LLVM_DEBUG(dbgs() << "LV: Loop does not require scalar epilogue\n");1302    return false;1303  }1304 1305  /// Returns true if a scalar epilogue is not allowed due to optsize or a1306  /// loop hint annotation.1307  bool isScalarEpilogueAllowed() const {1308    return ScalarEpilogueStatus == CM_ScalarEpilogueAllowed;1309  }1310 1311  /// Returns true if tail-folding is preferred over a scalar epilogue.1312  bool preferPredicatedLoop() const {1313    return ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate ||1314           ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedUsePredicate;1315  }1316 1317  /// Returns the TailFoldingStyle that is best for the current loop.1318  TailFoldingStyle getTailFoldingStyle(bool IVUpdateMayOverflow = true) const {1319    if (!ChosenTailFoldingStyle)1320      return TailFoldingStyle::None;1321    return IVUpdateMayOverflow ? ChosenTailFoldingStyle->first1322                               : ChosenTailFoldingStyle->second;1323  }1324 1325  /// Selects and saves TailFoldingStyle for 2 options - if IV update may1326  /// overflow or not.1327  /// \param IsScalableVF true if scalable vector factors enabled.1328  /// \param UserIC User specific interleave count.1329  void setTailFoldingStyles(bool IsScalableVF, unsigned UserIC) {1330    assert(!ChosenTailFoldingStyle && "Tail folding must not be selected yet.");1331    if (!Legal->canFoldTailByMasking()) {1332      ChosenTailFoldingStyle = {TailFoldingStyle::None, TailFoldingStyle::None};1333      return;1334    }1335 1336    // Default to TTI preference, but allow command line override.1337    ChosenTailFoldingStyle = {1338        TTI.getPreferredTailFoldingStyle(/*IVUpdateMayOverflow=*/true),1339        TTI.getPreferredTailFoldingStyle(/*IVUpdateMayOverflow=*/false)};1340    if (ForceTailFoldingStyle.getNumOccurrences())1341      ChosenTailFoldingStyle = {ForceTailFoldingStyle.getValue(),1342                                ForceTailFoldingStyle.getValue()};1343 1344    if (ChosenTailFoldingStyle->first != TailFoldingStyle::DataWithEVL &&1345        ChosenTailFoldingStyle->second != TailFoldingStyle::DataWithEVL)1346      return;1347    // Override EVL styles if needed.1348    // FIXME: Investigate opportunity for fixed vector factor.1349    bool EVLIsLegal = UserIC <= 1 && IsScalableVF &&1350                      TTI.hasActiveVectorLength() && !EnableVPlanNativePath;1351    if (EVLIsLegal)1352      return;1353    // If for some reason EVL mode is unsupported, fallback to a scalar epilogue1354    // if it's allowed, or DataWithoutLaneMask otherwise.1355    if (ScalarEpilogueStatus == CM_ScalarEpilogueAllowed ||1356        ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate)1357      ChosenTailFoldingStyle = {TailFoldingStyle::None, TailFoldingStyle::None};1358    else1359      ChosenTailFoldingStyle = {TailFoldingStyle::DataWithoutLaneMask,1360                                TailFoldingStyle::DataWithoutLaneMask};1361 1362    LLVM_DEBUG(1363        dbgs() << "LV: Preference for VP intrinsics indicated. Will "1364                  "not try to generate VP Intrinsics "1365               << (UserIC > 11366                       ? "since interleave count specified is greater than 1.\n"1367                       : "due to non-interleaving reasons.\n"));1368  }1369 1370  /// Returns true if all loop blocks should be masked to fold tail loop.1371  bool foldTailByMasking() const {1372    // TODO: check if it is possible to check for None style independent of1373    // IVUpdateMayOverflow flag in getTailFoldingStyle.1374    return getTailFoldingStyle() != TailFoldingStyle::None;1375  }1376 1377  /// Returns true if the use of wide lane masks is requested and the loop is1378  /// using tail-folding with a lane mask for control flow.1379  bool useWideActiveLaneMask() const {1380    if (!EnableWideActiveLaneMask)1381      return false;1382 1383    TailFoldingStyle TF = getTailFoldingStyle();1384    return TF == TailFoldingStyle::DataAndControlFlow ||1385           TF == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck;1386  }1387 1388  /// Return maximum safe number of elements to be processed per vector1389  /// iteration, which do not prevent store-load forwarding and are safe with1390  /// regard to the memory dependencies. Required for EVL-based VPlans to1391  /// correctly calculate AVL (application vector length) as min(remaining AVL,1392  /// MaxSafeElements).1393  /// TODO: need to consider adjusting cost model to use this value as a1394  /// vectorization factor for EVL-based vectorization.1395  std::optional<unsigned> getMaxSafeElements() const { return MaxSafeElements; }1396 1397  /// Returns true if the instructions in this block requires predication1398  /// for any reason, e.g. because tail folding now requires a predicate1399  /// or because the block in the original loop was predicated.1400  bool blockNeedsPredicationForAnyReason(BasicBlock *BB) const {1401    return foldTailByMasking() || Legal->blockNeedsPredication(BB);1402  }1403 1404  /// Returns true if VP intrinsics with explicit vector length support should1405  /// be generated in the tail folded loop.1406  bool foldTailWithEVL() const {1407    return getTailFoldingStyle() == TailFoldingStyle::DataWithEVL;1408  }1409 1410  /// Returns true if the Phi is part of an inloop reduction.1411  bool isInLoopReduction(PHINode *Phi) const {1412    return InLoopReductions.contains(Phi);1413  }1414 1415  /// Returns true if the predicated reduction select should be used to set the1416  /// incoming value for the reduction phi.1417  bool usePredicatedReductionSelect() const {1418    // Force to use predicated reduction select since the EVL of the1419    // second-to-last iteration might not be VF*UF.1420    if (foldTailWithEVL())1421      return true;1422    return PreferPredicatedReductionSelect ||1423           TTI.preferPredicatedReductionSelect();1424  }1425 1426  /// Estimate cost of an intrinsic call instruction CI if it were vectorized1427  /// with factor VF.  Return the cost of the instruction, including1428  /// scalarization overhead if it's needed.1429  InstructionCost getVectorIntrinsicCost(CallInst *CI, ElementCount VF) const;1430 1431  /// Estimate cost of a call instruction CI if it were vectorized with factor1432  /// VF. Return the cost of the instruction, including scalarization overhead1433  /// if it's needed.1434  InstructionCost getVectorCallCost(CallInst *CI, ElementCount VF) const;1435 1436  /// Invalidates decisions already taken by the cost model.1437  void invalidateCostModelingDecisions() {1438    WideningDecisions.clear();1439    CallWideningDecisions.clear();1440    Uniforms.clear();1441    Scalars.clear();1442  }1443 1444  /// Returns the expected execution cost. The unit of the cost does1445  /// not matter because we use the 'cost' units to compare different1446  /// vector widths. The cost that is returned is *not* normalized by1447  /// the factor width.1448  InstructionCost expectedCost(ElementCount VF);1449 1450  bool hasPredStores() const { return NumPredStores > 0; }1451 1452  /// Returns true if epilogue vectorization is considered profitable, and1453  /// false otherwise.1454  /// \p VF is the vectorization factor chosen for the original loop.1455  /// \p Multiplier is an aditional scaling factor applied to VF before1456  /// comparing to EpilogueVectorizationMinVF.1457  bool isEpilogueVectorizationProfitable(const ElementCount VF,1458                                         const unsigned IC) const;1459 1460  /// Returns the execution time cost of an instruction for a given vector1461  /// width. Vector width of one means scalar.1462  InstructionCost getInstructionCost(Instruction *I, ElementCount VF);1463 1464  /// Return the cost of instructions in an inloop reduction pattern, if I is1465  /// part of that pattern.1466  std::optional<InstructionCost> getReductionPatternCost(Instruction *I,1467                                                         ElementCount VF,1468                                                         Type *VectorTy) const;1469 1470  /// Returns true if \p Op should be considered invariant and if it is1471  /// trivially hoistable.1472  bool shouldConsiderInvariant(Value *Op);1473 1474  /// Return the value of vscale used for tuning the cost model.1475  std::optional<unsigned> getVScaleForTuning() const { return VScaleForTuning; }1476 1477private:1478  unsigned NumPredStores = 0;1479 1480  /// Used to store the value of vscale used for tuning the cost model. It is1481  /// initialized during object construction.1482  std::optional<unsigned> VScaleForTuning;1483 1484  /// Initializes the value of vscale used for tuning the cost model. If1485  /// vscale_range.min == vscale_range.max then return vscale_range.max, else1486  /// return the value returned by the corresponding TTI method.1487  void initializeVScaleForTuning() {1488    const Function *Fn = TheLoop->getHeader()->getParent();1489    if (Fn->hasFnAttribute(Attribute::VScaleRange)) {1490      auto Attr = Fn->getFnAttribute(Attribute::VScaleRange);1491      auto Min = Attr.getVScaleRangeMin();1492      auto Max = Attr.getVScaleRangeMax();1493      if (Max && Min == Max) {1494        VScaleForTuning = Max;1495        return;1496      }1497    }1498 1499    VScaleForTuning = TTI.getVScaleForTuning();1500  }1501 1502  /// \return An upper bound for the vectorization factors for both1503  /// fixed and scalable vectorization, where the minimum-known number of1504  /// elements is a power-of-2 larger than zero. If scalable vectorization is1505  /// disabled or unsupported, then the scalable part will be equal to1506  /// ElementCount::getScalable(0).1507  FixedScalableVFPair computeFeasibleMaxVF(unsigned MaxTripCount,1508                                           ElementCount UserVF,1509                                           bool FoldTailByMasking);1510 1511  /// If \p VF > MaxTripcount, clamps it to the next lower VF that is <=1512  /// MaxTripCount.1513  ElementCount clampVFByMaxTripCount(ElementCount VF, unsigned MaxTripCount,1514                                     bool FoldTailByMasking) const;1515 1516  /// \return the maximized element count based on the targets vector1517  /// registers and the loop trip-count, but limited to a maximum safe VF.1518  /// This is a helper function of computeFeasibleMaxVF.1519  ElementCount getMaximizedVFForTarget(unsigned MaxTripCount,1520                                       unsigned SmallestType,1521                                       unsigned WidestType,1522                                       ElementCount MaxSafeVF,1523                                       bool FoldTailByMasking);1524 1525  /// Checks if scalable vectorization is supported and enabled. Caches the1526  /// result to avoid repeated debug dumps for repeated queries.1527  bool isScalableVectorizationAllowed();1528 1529  /// \return the maximum legal scalable VF, based on the safe max number1530  /// of elements.1531  ElementCount getMaxLegalScalableVF(unsigned MaxSafeElements);1532 1533  /// Calculate vectorization cost of memory instruction \p I.1534  InstructionCost getMemoryInstructionCost(Instruction *I, ElementCount VF);1535 1536  /// The cost computation for scalarized memory instruction.1537  InstructionCost getMemInstScalarizationCost(Instruction *I, ElementCount VF);1538 1539  /// The cost computation for interleaving group of memory instructions.1540  InstructionCost getInterleaveGroupCost(Instruction *I, ElementCount VF);1541 1542  /// The cost computation for Gather/Scatter instruction.1543  InstructionCost getGatherScatterCost(Instruction *I, ElementCount VF);1544 1545  /// The cost computation for widening instruction \p I with consecutive1546  /// memory access.1547  InstructionCost getConsecutiveMemOpCost(Instruction *I, ElementCount VF);1548 1549  /// The cost calculation for Load/Store instruction \p I with uniform pointer -1550  /// Load: scalar load + broadcast.1551  /// Store: scalar store + (loop invariant value stored? 0 : extract of last1552  /// element)1553  InstructionCost getUniformMemOpCost(Instruction *I, ElementCount VF);1554 1555  /// Estimate the overhead of scalarizing an instruction. This is a1556  /// convenience wrapper for the type-based getScalarizationOverhead API.1557  InstructionCost getScalarizationOverhead(Instruction *I,1558                                           ElementCount VF) const;1559 1560  /// Returns true if an artificially high cost for emulated masked memrefs1561  /// should be used.1562  bool useEmulatedMaskMemRefHack(Instruction *I, ElementCount VF);1563 1564  /// Map of scalar integer values to the smallest bitwidth they can be legally1565  /// represented as. The vector equivalents of these values should be truncated1566  /// to this type.1567  MapVector<Instruction *, uint64_t> MinBWs;1568 1569  /// A type representing the costs for instructions if they were to be1570  /// scalarized rather than vectorized. The entries are Instruction-Cost1571  /// pairs.1572  using ScalarCostsTy = MapVector<Instruction *, InstructionCost>;1573 1574  /// A set containing all BasicBlocks that are known to present after1575  /// vectorization as a predicated block.1576  DenseMap<ElementCount, SmallPtrSet<BasicBlock *, 4>>1577      PredicatedBBsAfterVectorization;1578 1579  /// Records whether it is allowed to have the original scalar loop execute at1580  /// least once. This may be needed as a fallback loop in case runtime1581  /// aliasing/dependence checks fail, or to handle the tail/remainder1582  /// iterations when the trip count is unknown or doesn't divide by the VF,1583  /// or as a peel-loop to handle gaps in interleave-groups.1584  /// Under optsize and when the trip count is very small we don't allow any1585  /// iterations to execute in the scalar loop.1586  ScalarEpilogueLowering ScalarEpilogueStatus = CM_ScalarEpilogueAllowed;1587 1588  /// Control finally chosen tail folding style. The first element is used if1589  /// the IV update may overflow, the second element - if it does not.1590  std::optional<std::pair<TailFoldingStyle, TailFoldingStyle>>1591      ChosenTailFoldingStyle;1592 1593  /// true if scalable vectorization is supported and enabled.1594  std::optional<bool> IsScalableVectorizationAllowed;1595 1596  /// Maximum safe number of elements to be processed per vector iteration,1597  /// which do not prevent store-load forwarding and are safe with regard to the1598  /// memory dependencies. Required for EVL-based veectorization, where this1599  /// value is used as the upper bound of the safe AVL.1600  std::optional<unsigned> MaxSafeElements;1601 1602  /// A map holding scalar costs for different vectorization factors. The1603  /// presence of a cost for an instruction in the mapping indicates that the1604  /// instruction will be scalarized when vectorizing with the associated1605  /// vectorization factor. The entries are VF-ScalarCostTy pairs.1606  MapVector<ElementCount, ScalarCostsTy> InstsToScalarize;1607 1608  /// Holds the instructions known to be uniform after vectorization.1609  /// The data is collected per VF.1610  DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Uniforms;1611 1612  /// Holds the instructions known to be scalar after vectorization.1613  /// The data is collected per VF.1614  DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> Scalars;1615 1616  /// Holds the instructions (address computations) that are forced to be1617  /// scalarized.1618  DenseMap<ElementCount, SmallPtrSet<Instruction *, 4>> ForcedScalars;1619 1620  /// PHINodes of the reductions that should be expanded in-loop.1621  SmallPtrSet<PHINode *, 4> InLoopReductions;1622 1623  /// A Map of inloop reduction operations and their immediate chain operand.1624  /// FIXME: This can be removed once reductions can be costed correctly in1625  /// VPlan. This was added to allow quick lookup of the inloop operations.1626  DenseMap<Instruction *, Instruction *> InLoopReductionImmediateChains;1627 1628  /// Returns the expected difference in cost from scalarizing the expression1629  /// feeding a predicated instruction \p PredInst. The instructions to1630  /// scalarize and their scalar costs are collected in \p ScalarCosts. A1631  /// non-negative return value implies the expression will be scalarized.1632  /// Currently, only single-use chains are considered for scalarization.1633  InstructionCost computePredInstDiscount(Instruction *PredInst,1634                                          ScalarCostsTy &ScalarCosts,1635                                          ElementCount VF);1636 1637  /// Collect the instructions that are uniform after vectorization. An1638  /// instruction is uniform if we represent it with a single scalar value in1639  /// the vectorized loop corresponding to each vector iteration. Examples of1640  /// uniform instructions include pointer operands of consecutive or1641  /// interleaved memory accesses. Note that although uniformity implies an1642  /// instruction will be scalar, the reverse is not true. In general, a1643  /// scalarized instruction will be represented by VF scalar values in the1644  /// vectorized loop, each corresponding to an iteration of the original1645  /// scalar loop.1646  void collectLoopUniforms(ElementCount VF);1647 1648  /// Collect the instructions that are scalar after vectorization. An1649  /// instruction is scalar if it is known to be uniform or will be scalarized1650  /// during vectorization. collectLoopScalars should only add non-uniform nodes1651  /// to the list if they are used by a load/store instruction that is marked as1652  /// CM_Scalarize. Non-uniform scalarized instructions will be represented by1653  /// VF values in the vectorized loop, each corresponding to an iteration of1654  /// the original scalar loop.1655  void collectLoopScalars(ElementCount VF);1656 1657  /// Keeps cost model vectorization decision and cost for instructions.1658  /// Right now it is used for memory instructions only.1659  using DecisionList = DenseMap<std::pair<Instruction *, ElementCount>,1660                                std::pair<InstWidening, InstructionCost>>;1661 1662  DecisionList WideningDecisions;1663 1664  using CallDecisionList =1665      DenseMap<std::pair<CallInst *, ElementCount>, CallWideningDecision>;1666 1667  CallDecisionList CallWideningDecisions;1668 1669  /// Returns true if \p V is expected to be vectorized and it needs to be1670  /// extracted.1671  bool needsExtract(Value *V, ElementCount VF) const {1672    Instruction *I = dyn_cast<Instruction>(V);1673    if (VF.isScalar() || !I || !TheLoop->contains(I) ||1674        TheLoop->isLoopInvariant(I) ||1675        getWideningDecision(I, VF) == CM_Scalarize ||1676        (isa<CallInst>(I) &&1677         getCallWideningDecision(cast<CallInst>(I), VF).Kind == CM_Scalarize))1678      return false;1679 1680    // Assume we can vectorize V (and hence we need extraction) if the1681    // scalars are not computed yet. This can happen, because it is called1682    // via getScalarizationOverhead from setCostBasedWideningDecision, before1683    // the scalars are collected. That should be a safe assumption in most1684    // cases, because we check if the operands have vectorizable types1685    // beforehand in LoopVectorizationLegality.1686    return !Scalars.contains(VF) || !isScalarAfterVectorization(I, VF);1687  };1688 1689  /// Returns a range containing only operands needing to be extracted.1690  SmallVector<Value *, 4> filterExtractingOperands(Instruction::op_range Ops,1691                                                   ElementCount VF) const {1692 1693    SmallPtrSet<const Value *, 4> UniqueOperands;1694    SmallVector<Value *, 4> Res;1695    for (Value *Op : Ops) {1696      if (isa<Constant>(Op) || !UniqueOperands.insert(Op).second ||1697          !needsExtract(Op, VF))1698        continue;1699      Res.push_back(Op);1700    }1701    return Res;1702  }1703 1704public:1705  /// The loop that we evaluate.1706  Loop *TheLoop;1707 1708  /// Predicated scalar evolution analysis.1709  PredicatedScalarEvolution &PSE;1710 1711  /// Loop Info analysis.1712  LoopInfo *LI;1713 1714  /// Vectorization legality.1715  LoopVectorizationLegality *Legal;1716 1717  /// Vector target information.1718  const TargetTransformInfo &TTI;1719 1720  /// Target Library Info.1721  const TargetLibraryInfo *TLI;1722 1723  /// Demanded bits analysis.1724  DemandedBits *DB;1725 1726  /// Assumption cache.1727  AssumptionCache *AC;1728 1729  /// Interface to emit optimization remarks.1730  OptimizationRemarkEmitter *ORE;1731 1732  const Function *TheFunction;1733 1734  /// Loop Vectorize Hint.1735  const LoopVectorizeHints *Hints;1736 1737  /// The interleave access information contains groups of interleaved accesses1738  /// with the same stride and close to each other.1739  InterleavedAccessInfo &InterleaveInfo;1740 1741  /// Values to ignore in the cost model.1742  SmallPtrSet<const Value *, 16> ValuesToIgnore;1743 1744  /// Values to ignore in the cost model when VF > 1.1745  SmallPtrSet<const Value *, 16> VecValuesToIgnore;1746 1747  /// All element types found in the loop.1748  SmallPtrSet<Type *, 16> ElementTypesInLoop;1749 1750  /// The kind of cost that we are calculating1751  TTI::TargetCostKind CostKind;1752 1753  /// Whether this loop should be optimized for size based on function attribute1754  /// or profile information.1755  bool OptForSize;1756 1757  /// The highest VF possible for this loop, without using MaxBandwidth.1758  FixedScalableVFPair MaxPermissibleVFWithoutMaxBW;1759};1760} // end namespace llvm1761 1762namespace {1763/// Helper struct to manage generating runtime checks for vectorization.1764///1765/// The runtime checks are created up-front in temporary blocks to allow better1766/// estimating the cost and un-linked from the existing IR. After deciding to1767/// vectorize, the checks are moved back. If deciding not to vectorize, the1768/// temporary blocks are completely removed.1769class GeneratedRTChecks {1770  /// Basic block which contains the generated SCEV checks, if any.1771  BasicBlock *SCEVCheckBlock = nullptr;1772 1773  /// The value representing the result of the generated SCEV checks. If it is1774  /// nullptr no SCEV checks have been generated.1775  Value *SCEVCheckCond = nullptr;1776 1777  /// Basic block which contains the generated memory runtime checks, if any.1778  BasicBlock *MemCheckBlock = nullptr;1779 1780  /// The value representing the result of the generated memory runtime checks.1781  /// If it is nullptr no memory runtime checks have been generated.1782  Value *MemRuntimeCheckCond = nullptr;1783 1784  DominatorTree *DT;1785  LoopInfo *LI;1786  TargetTransformInfo *TTI;1787 1788  SCEVExpander SCEVExp;1789  SCEVExpander MemCheckExp;1790 1791  bool CostTooHigh = false;1792 1793  Loop *OuterLoop = nullptr;1794 1795  PredicatedScalarEvolution &PSE;1796 1797  /// The kind of cost that we are calculating1798  TTI::TargetCostKind CostKind;1799 1800public:1801  GeneratedRTChecks(PredicatedScalarEvolution &PSE, DominatorTree *DT,1802                    LoopInfo *LI, TargetTransformInfo *TTI,1803                    const DataLayout &DL, TTI::TargetCostKind CostKind)1804      : DT(DT), LI(LI), TTI(TTI),1805        SCEVExp(*PSE.getSE(), DL, "scev.check", /*PreserveLCSSA=*/false),1806        MemCheckExp(*PSE.getSE(), DL, "scev.check", /*PreserveLCSSA=*/false),1807        PSE(PSE), CostKind(CostKind) {}1808 1809  /// Generate runtime checks in SCEVCheckBlock and MemCheckBlock, so we can1810  /// accurately estimate the cost of the runtime checks. The blocks are1811  /// un-linked from the IR and are added back during vector code generation. If1812  /// there is no vector code generation, the check blocks are removed1813  /// completely.1814  void create(Loop *L, const LoopAccessInfo &LAI,1815              const SCEVPredicate &UnionPred, ElementCount VF, unsigned IC) {1816 1817    // Hard cutoff to limit compile-time increase in case a very large number of1818    // runtime checks needs to be generated.1819    // TODO: Skip cutoff if the loop is guaranteed to execute, e.g. due to1820    // profile info.1821    CostTooHigh =1822        LAI.getNumRuntimePointerChecks() > VectorizeMemoryCheckThreshold;1823    if (CostTooHigh)1824      return;1825 1826    BasicBlock *LoopHeader = L->getHeader();1827    BasicBlock *Preheader = L->getLoopPreheader();1828 1829    // Use SplitBlock to create blocks for SCEV & memory runtime checks to1830    // ensure the blocks are properly added to LoopInfo & DominatorTree. Those1831    // may be used by SCEVExpander. The blocks will be un-linked from their1832    // predecessors and removed from LI & DT at the end of the function.1833    if (!UnionPred.isAlwaysTrue()) {1834      SCEVCheckBlock = SplitBlock(Preheader, Preheader->getTerminator(), DT, LI,1835                                  nullptr, "vector.scevcheck");1836 1837      SCEVCheckCond = SCEVExp.expandCodeForPredicate(1838          &UnionPred, SCEVCheckBlock->getTerminator());1839      if (isa<Constant>(SCEVCheckCond)) {1840        // Clean up directly after expanding the predicate to a constant, to1841        // avoid further expansions re-using anything left over from SCEVExp.1842        SCEVExpanderCleaner SCEVCleaner(SCEVExp);1843        SCEVCleaner.cleanup();1844      }1845    }1846 1847    const auto &RtPtrChecking = *LAI.getRuntimePointerChecking();1848    if (RtPtrChecking.Need) {1849      auto *Pred = SCEVCheckBlock ? SCEVCheckBlock : Preheader;1850      MemCheckBlock = SplitBlock(Pred, Pred->getTerminator(), DT, LI, nullptr,1851                                 "vector.memcheck");1852 1853      auto DiffChecks = RtPtrChecking.getDiffChecks();1854      if (DiffChecks) {1855        Value *RuntimeVF = nullptr;1856        MemRuntimeCheckCond = addDiffRuntimeChecks(1857            MemCheckBlock->getTerminator(), *DiffChecks, MemCheckExp,1858            [VF, &RuntimeVF](IRBuilderBase &B, unsigned Bits) {1859              if (!RuntimeVF)1860                RuntimeVF = getRuntimeVF(B, B.getIntNTy(Bits), VF);1861              return RuntimeVF;1862            },1863            IC);1864      } else {1865        MemRuntimeCheckCond = addRuntimeChecks(1866            MemCheckBlock->getTerminator(), L, RtPtrChecking.getChecks(),1867            MemCheckExp, VectorizerParams::HoistRuntimeChecks);1868      }1869      assert(MemRuntimeCheckCond &&1870             "no RT checks generated although RtPtrChecking "1871             "claimed checks are required");1872    }1873 1874    SCEVExp.eraseDeadInstructions(SCEVCheckCond);1875 1876    if (!MemCheckBlock && !SCEVCheckBlock)1877      return;1878 1879    // Unhook the temporary block with the checks, update various places1880    // accordingly.1881    if (SCEVCheckBlock)1882      SCEVCheckBlock->replaceAllUsesWith(Preheader);1883    if (MemCheckBlock)1884      MemCheckBlock->replaceAllUsesWith(Preheader);1885 1886    if (SCEVCheckBlock) {1887      SCEVCheckBlock->getTerminator()->moveBefore(1888          Preheader->getTerminator()->getIterator());1889      auto *UI = new UnreachableInst(Preheader->getContext(), SCEVCheckBlock);1890      UI->setDebugLoc(DebugLoc::getTemporary());1891      Preheader->getTerminator()->eraseFromParent();1892    }1893    if (MemCheckBlock) {1894      MemCheckBlock->getTerminator()->moveBefore(1895          Preheader->getTerminator()->getIterator());1896      auto *UI = new UnreachableInst(Preheader->getContext(), MemCheckBlock);1897      UI->setDebugLoc(DebugLoc::getTemporary());1898      Preheader->getTerminator()->eraseFromParent();1899    }1900 1901    DT->changeImmediateDominator(LoopHeader, Preheader);1902    if (MemCheckBlock) {1903      DT->eraseNode(MemCheckBlock);1904      LI->removeBlock(MemCheckBlock);1905    }1906    if (SCEVCheckBlock) {1907      DT->eraseNode(SCEVCheckBlock);1908      LI->removeBlock(SCEVCheckBlock);1909    }1910 1911    // Outer loop is used as part of the later cost calculations.1912    OuterLoop = L->getParentLoop();1913  }1914 1915  InstructionCost getCost() {1916    if (SCEVCheckBlock || MemCheckBlock)1917      LLVM_DEBUG(dbgs() << "Calculating cost of runtime checks:\n");1918 1919    if (CostTooHigh) {1920      InstructionCost Cost;1921      Cost.setInvalid();1922      LLVM_DEBUG(dbgs() << "  number of checks exceeded threshold\n");1923      return Cost;1924    }1925 1926    InstructionCost RTCheckCost = 0;1927    if (SCEVCheckBlock)1928      for (Instruction &I : *SCEVCheckBlock) {1929        if (SCEVCheckBlock->getTerminator() == &I)1930          continue;1931        InstructionCost C = TTI->getInstructionCost(&I, CostKind);1932        LLVM_DEBUG(dbgs() << "  " << C << "  for " << I << "\n");1933        RTCheckCost += C;1934      }1935    if (MemCheckBlock) {1936      InstructionCost MemCheckCost = 0;1937      for (Instruction &I : *MemCheckBlock) {1938        if (MemCheckBlock->getTerminator() == &I)1939          continue;1940        InstructionCost C = TTI->getInstructionCost(&I, CostKind);1941        LLVM_DEBUG(dbgs() << "  " << C << "  for " << I << "\n");1942        MemCheckCost += C;1943      }1944 1945      // If the runtime memory checks are being created inside an outer loop1946      // we should find out if these checks are outer loop invariant. If so,1947      // the checks will likely be hoisted out and so the effective cost will1948      // reduce according to the outer loop trip count.1949      if (OuterLoop) {1950        ScalarEvolution *SE = MemCheckExp.getSE();1951        // TODO: If profitable, we could refine this further by analysing every1952        // individual memory check, since there could be a mixture of loop1953        // variant and invariant checks that mean the final condition is1954        // variant.1955        const SCEV *Cond = SE->getSCEV(MemRuntimeCheckCond);1956        if (SE->isLoopInvariant(Cond, OuterLoop)) {1957          // It seems reasonable to assume that we can reduce the effective1958          // cost of the checks even when we know nothing about the trip1959          // count. Assume that the outer loop executes at least twice.1960          unsigned BestTripCount = 2;1961 1962          // Get the best known TC estimate.1963          if (auto EstimatedTC = getSmallBestKnownTC(1964                  PSE, OuterLoop, /* CanUseConstantMax = */ false))1965            if (EstimatedTC->isFixed())1966              BestTripCount = EstimatedTC->getFixedValue();1967 1968          InstructionCost NewMemCheckCost = MemCheckCost / BestTripCount;1969 1970          // Let's ensure the cost is always at least 1.1971          NewMemCheckCost = std::max(NewMemCheckCost.getValue(),1972                                     (InstructionCost::CostType)1);1973 1974          if (BestTripCount > 1)1975            LLVM_DEBUG(dbgs()1976                       << "We expect runtime memory checks to be hoisted "1977                       << "out of the outer loop. Cost reduced from "1978                       << MemCheckCost << " to " << NewMemCheckCost << '\n');1979 1980          MemCheckCost = NewMemCheckCost;1981        }1982      }1983 1984      RTCheckCost += MemCheckCost;1985    }1986 1987    if (SCEVCheckBlock || MemCheckBlock)1988      LLVM_DEBUG(dbgs() << "Total cost of runtime checks: " << RTCheckCost1989                        << "\n");1990 1991    return RTCheckCost;1992  }1993 1994  /// Remove the created SCEV & memory runtime check blocks & instructions, if1995  /// unused.1996  ~GeneratedRTChecks() {1997    SCEVExpanderCleaner SCEVCleaner(SCEVExp);1998    SCEVExpanderCleaner MemCheckCleaner(MemCheckExp);1999    bool SCEVChecksUsed = !SCEVCheckBlock || !pred_empty(SCEVCheckBlock);2000    bool MemChecksUsed = !MemCheckBlock || !pred_empty(MemCheckBlock);2001    if (SCEVChecksUsed)2002      SCEVCleaner.markResultUsed();2003 2004    if (MemChecksUsed) {2005      MemCheckCleaner.markResultUsed();2006    } else {2007      auto &SE = *MemCheckExp.getSE();2008      // Memory runtime check generation creates compares that use expanded2009      // values. Remove them before running the SCEVExpanderCleaners.2010      for (auto &I : make_early_inc_range(reverse(*MemCheckBlock))) {2011        if (MemCheckExp.isInsertedInstruction(&I))2012          continue;2013        SE.forgetValue(&I);2014        I.eraseFromParent();2015      }2016    }2017    MemCheckCleaner.cleanup();2018    SCEVCleaner.cleanup();2019 2020    if (!SCEVChecksUsed)2021      SCEVCheckBlock->eraseFromParent();2022    if (!MemChecksUsed)2023      MemCheckBlock->eraseFromParent();2024  }2025 2026  /// Retrieves the SCEVCheckCond and SCEVCheckBlock that were generated as IR2027  /// outside VPlan.2028  std::pair<Value *, BasicBlock *> getSCEVChecks() const {2029    using namespace llvm::PatternMatch;2030    if (!SCEVCheckCond || match(SCEVCheckCond, m_ZeroInt()))2031      return {nullptr, nullptr};2032 2033    return {SCEVCheckCond, SCEVCheckBlock};2034  }2035 2036  /// Retrieves the MemCheckCond and MemCheckBlock that were generated as IR2037  /// outside VPlan.2038  std::pair<Value *, BasicBlock *> getMemRuntimeChecks() const {2039    using namespace llvm::PatternMatch;2040    if (MemRuntimeCheckCond && match(MemRuntimeCheckCond, m_ZeroInt()))2041      return {nullptr, nullptr};2042    return {MemRuntimeCheckCond, MemCheckBlock};2043  }2044 2045  /// Return true if any runtime checks have been added2046  bool hasChecks() const {2047    return getSCEVChecks().first || getMemRuntimeChecks().first;2048  }2049};2050} // namespace2051 2052static bool useActiveLaneMask(TailFoldingStyle Style) {2053  return Style == TailFoldingStyle::Data ||2054         Style == TailFoldingStyle::DataAndControlFlow ||2055         Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck;2056}2057 2058static bool useActiveLaneMaskForControlFlow(TailFoldingStyle Style) {2059  return Style == TailFoldingStyle::DataAndControlFlow ||2060         Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck;2061}2062 2063// Return true if \p OuterLp is an outer loop annotated with hints for explicit2064// vectorization. The loop needs to be annotated with #pragma omp simd2065// simdlen(#) or #pragma clang vectorize(enable) vectorize_width(#). If the2066// vector length information is not provided, vectorization is not considered2067// explicit. Interleave hints are not allowed either. These limitations will be2068// relaxed in the future.2069// Please, note that we are currently forced to abuse the pragma 'clang2070// vectorize' semantics. This pragma provides *auto-vectorization hints*2071// (i.e., LV must check that vectorization is legal) whereas pragma 'omp simd'2072// provides *explicit vectorization hints* (LV can bypass legal checks and2073// assume that vectorization is legal). However, both hints are implemented2074// using the same metadata (llvm.loop.vectorize, processed by2075// LoopVectorizeHints). This will be fixed in the future when the native IR2076// representation for pragma 'omp simd' is introduced.2077static bool isExplicitVecOuterLoop(Loop *OuterLp,2078                                   OptimizationRemarkEmitter *ORE) {2079  assert(!OuterLp->isInnermost() && "This is not an outer loop");2080  LoopVectorizeHints Hints(OuterLp, true /*DisableInterleaving*/, *ORE);2081 2082  // Only outer loops with an explicit vectorization hint are supported.2083  // Unannotated outer loops are ignored.2084  if (Hints.getForce() == LoopVectorizeHints::FK_Undefined)2085    return false;2086 2087  Function *Fn = OuterLp->getHeader()->getParent();2088  if (!Hints.allowVectorization(Fn, OuterLp,2089                                true /*VectorizeOnlyWhenForced*/)) {2090    LLVM_DEBUG(dbgs() << "LV: Loop hints prevent outer loop vectorization.\n");2091    return false;2092  }2093 2094  if (Hints.getInterleave() > 1) {2095    // TODO: Interleave support is future work.2096    LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Interleave is not supported for "2097                         "outer loops.\n");2098    Hints.emitRemarkWithHints();2099    return false;2100  }2101 2102  return true;2103}2104 2105static void collectSupportedLoops(Loop &L, LoopInfo *LI,2106                                  OptimizationRemarkEmitter *ORE,2107                                  SmallVectorImpl<Loop *> &V) {2108  // Collect inner loops and outer loops without irreducible control flow. For2109  // now, only collect outer loops that have explicit vectorization hints. If we2110  // are stress testing the VPlan H-CFG construction, we collect the outermost2111  // loop of every loop nest.2112  if (L.isInnermost() || VPlanBuildStressTest ||2113      (EnableVPlanNativePath && isExplicitVecOuterLoop(&L, ORE))) {2114    LoopBlocksRPO RPOT(&L);2115    RPOT.perform(LI);2116    if (!containsIrreducibleCFG<const BasicBlock *>(RPOT, *LI)) {2117      V.push_back(&L);2118      // TODO: Collect inner loops inside marked outer loops in case2119      // vectorization fails for the outer loop. Do not invoke2120      // 'containsIrreducibleCFG' again for inner loops when the outer loop is2121      // already known to be reducible. We can use an inherited attribute for2122      // that.2123      return;2124    }2125  }2126  for (Loop *InnerL : L)2127    collectSupportedLoops(*InnerL, LI, ORE, V);2128}2129 2130//===----------------------------------------------------------------------===//2131// Implementation of LoopVectorizationLegality, InnerLoopVectorizer and2132// LoopVectorizationCostModel and LoopVectorizationPlanner.2133//===----------------------------------------------------------------------===//2134 2135/// Compute the transformed value of Index at offset StartValue using step2136/// StepValue.2137/// For integer induction, returns StartValue + Index * StepValue.2138/// For pointer induction, returns StartValue[Index * StepValue].2139/// FIXME: The newly created binary instructions should contain nsw/nuw2140/// flags, which can be found from the original scalar operations.2141static Value *2142emitTransformedIndex(IRBuilderBase &B, Value *Index, Value *StartValue,2143                     Value *Step,2144                     InductionDescriptor::InductionKind InductionKind,2145                     const BinaryOperator *InductionBinOp) {2146  using namespace llvm::PatternMatch;2147  Type *StepTy = Step->getType();2148  Value *CastedIndex = StepTy->isIntegerTy()2149                           ? B.CreateSExtOrTrunc(Index, StepTy)2150                           : B.CreateCast(Instruction::SIToFP, Index, StepTy);2151  if (CastedIndex != Index) {2152    CastedIndex->setName(CastedIndex->getName() + ".cast");2153    Index = CastedIndex;2154  }2155 2156  // Note: the IR at this point is broken. We cannot use SE to create any new2157  // SCEV and then expand it, hoping that SCEV's simplification will give us2158  // a more optimal code. Unfortunately, attempt of doing so on invalid IR may2159  // lead to various SCEV crashes. So all we can do is to use builder and rely2160  // on InstCombine for future simplifications. Here we handle some trivial2161  // cases only.2162  auto CreateAdd = [&B](Value *X, Value *Y) {2163    assert(X->getType() == Y->getType() && "Types don't match!");2164    if (match(X, m_ZeroInt()))2165      return Y;2166    if (match(Y, m_ZeroInt()))2167      return X;2168    return B.CreateAdd(X, Y);2169  };2170 2171  // We allow X to be a vector type, in which case Y will potentially be2172  // splatted into a vector with the same element count.2173  auto CreateMul = [&B](Value *X, Value *Y) {2174    assert(X->getType()->getScalarType() == Y->getType() &&2175           "Types don't match!");2176    if (match(X, m_One()))2177      return Y;2178    if (match(Y, m_One()))2179      return X;2180    VectorType *XVTy = dyn_cast<VectorType>(X->getType());2181    if (XVTy && !isa<VectorType>(Y->getType()))2182      Y = B.CreateVectorSplat(XVTy->getElementCount(), Y);2183    return B.CreateMul(X, Y);2184  };2185 2186  switch (InductionKind) {2187  case InductionDescriptor::IK_IntInduction: {2188    assert(!isa<VectorType>(Index->getType()) &&2189           "Vector indices not supported for integer inductions yet");2190    assert(Index->getType() == StartValue->getType() &&2191           "Index type does not match StartValue type");2192    if (isa<ConstantInt>(Step) && cast<ConstantInt>(Step)->isMinusOne())2193      return B.CreateSub(StartValue, Index);2194    auto *Offset = CreateMul(Index, Step);2195    return CreateAdd(StartValue, Offset);2196  }2197  case InductionDescriptor::IK_PtrInduction:2198    return B.CreatePtrAdd(StartValue, CreateMul(Index, Step));2199  case InductionDescriptor::IK_FpInduction: {2200    assert(!isa<VectorType>(Index->getType()) &&2201           "Vector indices not supported for FP inductions yet");2202    assert(Step->getType()->isFloatingPointTy() && "Expected FP Step value");2203    assert(InductionBinOp &&2204           (InductionBinOp->getOpcode() == Instruction::FAdd ||2205            InductionBinOp->getOpcode() == Instruction::FSub) &&2206           "Original bin op should be defined for FP induction");2207 2208    Value *MulExp = B.CreateFMul(Step, Index);2209    return B.CreateBinOp(InductionBinOp->getOpcode(), StartValue, MulExp,2210                         "induction");2211  }2212  case InductionDescriptor::IK_NoInduction:2213    return nullptr;2214  }2215  llvm_unreachable("invalid enum");2216}2217 2218static std::optional<unsigned> getMaxVScale(const Function &F,2219                                            const TargetTransformInfo &TTI) {2220  if (std::optional<unsigned> MaxVScale = TTI.getMaxVScale())2221    return MaxVScale;2222 2223  if (F.hasFnAttribute(Attribute::VScaleRange))2224    return F.getFnAttribute(Attribute::VScaleRange).getVScaleRangeMax();2225 2226  return std::nullopt;2227}2228 2229/// For the given VF and UF and maximum trip count computed for the loop, return2230/// whether the induction variable might overflow in the vectorized loop. If not,2231/// then we know a runtime overflow check always evaluates to false and can be2232/// removed.2233static bool isIndvarOverflowCheckKnownFalse(2234    const LoopVectorizationCostModel *Cost,2235    ElementCount VF, std::optional<unsigned> UF = std::nullopt) {2236  // Always be conservative if we don't know the exact unroll factor.2237  unsigned MaxUF = UF ? *UF : Cost->TTI.getMaxInterleaveFactor(VF);2238 2239  IntegerType *IdxTy = Cost->Legal->getWidestInductionType();2240  APInt MaxUIntTripCount = IdxTy->getMask();2241 2242  // We know the runtime overflow check is known false iff the (max) trip-count2243  // is known and (max) trip-count + (VF * UF) does not overflow in the type of2244  // the vector loop induction variable.2245  if (unsigned TC = Cost->PSE.getSmallConstantMaxTripCount()) {2246    uint64_t MaxVF = VF.getKnownMinValue();2247    if (VF.isScalable()) {2248      std::optional<unsigned> MaxVScale =2249          getMaxVScale(*Cost->TheFunction, Cost->TTI);2250      if (!MaxVScale)2251        return false;2252      MaxVF *= *MaxVScale;2253    }2254 2255    return (MaxUIntTripCount - TC).ugt(MaxVF * MaxUF);2256  }2257 2258  return false;2259}2260 2261// Return whether we allow using masked interleave-groups (for dealing with2262// strided loads/stores that reside in predicated blocks, or for dealing2263// with gaps).2264static bool useMaskedInterleavedAccesses(const TargetTransformInfo &TTI) {2265  // If an override option has been passed in for interleaved accesses, use it.2266  if (EnableMaskedInterleavedMemAccesses.getNumOccurrences() > 0)2267    return EnableMaskedInterleavedMemAccesses;2268 2269  return TTI.enableMaskedInterleavedAccessVectorization();2270}2271 2272void EpilogueVectorizerMainLoop::introduceCheckBlockInVPlan(2273    BasicBlock *CheckIRBB) {2274  // Note: The block with the minimum trip-count check is already connected2275  // during earlier VPlan construction.2276  VPBlockBase *ScalarPH = Plan.getScalarPreheader();2277  VPBlockBase *PreVectorPH = VectorPHVPBB->getSinglePredecessor();2278  assert(PreVectorPH->getNumSuccessors() == 2 && "Expected 2 successors");2279  assert(PreVectorPH->getSuccessors()[0] == ScalarPH && "Unexpected successor");2280  VPIRBasicBlock *CheckVPIRBB = Plan.createVPIRBasicBlock(CheckIRBB);2281  VPBlockUtils::insertOnEdge(PreVectorPH, VectorPHVPBB, CheckVPIRBB);2282  PreVectorPH = CheckVPIRBB;2283  VPBlockUtils::connectBlocks(PreVectorPH, ScalarPH);2284  PreVectorPH->swapSuccessors();2285 2286  // We just connected a new block to the scalar preheader. Update all2287  // VPPhis by adding an incoming value for it, replicating the last value.2288  unsigned NumPredecessors = ScalarPH->getNumPredecessors();2289  for (VPRecipeBase &R : cast<VPBasicBlock>(ScalarPH)->phis()) {2290    assert(isa<VPPhi>(&R) && "Phi expected to be VPPhi");2291    assert(cast<VPPhi>(&R)->getNumIncoming() == NumPredecessors - 1 &&2292           "must have incoming values for all operands");2293    R.addOperand(R.getOperand(NumPredecessors - 2));2294  }2295}2296 2297Value *EpilogueVectorizerMainLoop::createIterationCountCheck(2298    BasicBlock *VectorPH, ElementCount VF, unsigned UF) const {2299  // Generate code to check if the loop's trip count is less than VF * UF, or2300  // equal to it in case a scalar epilogue is required; this implies that the2301  // vector trip count is zero. This check also covers the case where adding one2302  // to the backedge-taken count overflowed leading to an incorrect trip count2303  // of zero. In this case we will also jump to the scalar loop.2304  auto P = Cost->requiresScalarEpilogue(VF.isVector()) ? ICmpInst::ICMP_ULE2305                                                       : ICmpInst::ICMP_ULT;2306 2307  // Reuse existing vector loop preheader for TC checks.2308  // Note that new preheader block is generated for vector loop.2309  BasicBlock *const TCCheckBlock = VectorPH;2310  IRBuilder<InstSimplifyFolder> Builder(2311      TCCheckBlock->getContext(),2312      InstSimplifyFolder(TCCheckBlock->getDataLayout()));2313  Builder.SetInsertPoint(TCCheckBlock->getTerminator());2314 2315  // If tail is to be folded, vector loop takes care of all iterations.2316  Value *Count = getTripCount();2317  Type *CountTy = Count->getType();2318  Value *CheckMinIters = Builder.getFalse();2319  auto CreateStep = [&]() -> Value * {2320    // Create step with max(MinProTripCount, UF * VF).2321    if (UF * VF.getKnownMinValue() >= MinProfitableTripCount.getKnownMinValue())2322      return createStepForVF(Builder, CountTy, VF, UF);2323 2324    Value *MinProfTC =2325        Builder.CreateElementCount(CountTy, MinProfitableTripCount);2326    if (!VF.isScalable())2327      return MinProfTC;2328    return Builder.CreateBinaryIntrinsic(2329        Intrinsic::umax, MinProfTC, createStepForVF(Builder, CountTy, VF, UF));2330  };2331 2332  TailFoldingStyle Style = Cost->getTailFoldingStyle();2333  if (Style == TailFoldingStyle::None) {2334    Value *Step = CreateStep();2335    ScalarEvolution &SE = *PSE.getSE();2336    // TODO: Emit unconditional branch to vector preheader instead of2337    // conditional branch with known condition.2338    const SCEV *TripCountSCEV = SE.applyLoopGuards(SE.getSCEV(Count), OrigLoop);2339    // Check if the trip count is < the step.2340    if (SE.isKnownPredicate(P, TripCountSCEV, SE.getSCEV(Step))) {2341      // TODO: Ensure step is at most the trip count when determining max VF and2342      // UF, w/o tail folding.2343      CheckMinIters = Builder.getTrue();2344    } else if (!SE.isKnownPredicate(CmpInst::getInversePredicate(P),2345                                    TripCountSCEV, SE.getSCEV(Step))) {2346      // Generate the minimum iteration check only if we cannot prove the2347      // check is known to be true, or known to be false.2348      CheckMinIters = Builder.CreateICmp(P, Count, Step, "min.iters.check");2349    } // else step known to be < trip count, use CheckMinIters preset to false.2350  } else if (VF.isScalable() && !TTI->isVScaleKnownToBeAPowerOfTwo() &&2351             !isIndvarOverflowCheckKnownFalse(Cost, VF, UF) &&2352             Style != TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck) {2353    // vscale is not necessarily a power-of-2, which means we cannot guarantee2354    // an overflow to zero when updating induction variables and so an2355    // additional overflow check is required before entering the vector loop.2356 2357    // Get the maximum unsigned value for the type.2358    Value *MaxUIntTripCount =2359        ConstantInt::get(CountTy, cast<IntegerType>(CountTy)->getMask());2360    Value *LHS = Builder.CreateSub(MaxUIntTripCount, Count);2361 2362    // Don't execute the vector loop if (UMax - n) < (VF * UF).2363    CheckMinIters = Builder.CreateICmp(ICmpInst::ICMP_ULT, LHS, CreateStep());2364  }2365  return CheckMinIters;2366}2367 2368/// Replace \p VPBB with a VPIRBasicBlock wrapping \p IRBB. All recipes from \p2369/// VPBB are moved to the end of the newly created VPIRBasicBlock. All2370/// predecessors and successors of VPBB, if any, are rewired to the new2371/// VPIRBasicBlock. If \p VPBB may be unreachable, \p Plan must be passed.2372static VPIRBasicBlock *replaceVPBBWithIRVPBB(VPBasicBlock *VPBB,2373                                             BasicBlock *IRBB,2374                                             VPlan *Plan = nullptr) {2375  if (!Plan)2376    Plan = VPBB->getPlan();2377  VPIRBasicBlock *IRVPBB = Plan->createVPIRBasicBlock(IRBB);2378  auto IP = IRVPBB->begin();2379  for (auto &R : make_early_inc_range(VPBB->phis()))2380    R.moveBefore(*IRVPBB, IP);2381 2382  for (auto &R :2383       make_early_inc_range(make_range(VPBB->getFirstNonPhi(), VPBB->end())))2384    R.moveBefore(*IRVPBB, IRVPBB->end());2385 2386  VPBlockUtils::reassociateBlocks(VPBB, IRVPBB);2387  // VPBB is now dead and will be cleaned up when the plan gets destroyed.2388  return IRVPBB;2389}2390 2391BasicBlock *InnerLoopVectorizer::createScalarPreheader(StringRef Prefix) {2392  BasicBlock *VectorPH = OrigLoop->getLoopPreheader();2393  assert(VectorPH && "Invalid loop structure");2394  assert((OrigLoop->getUniqueLatchExitBlock() ||2395          Cost->requiresScalarEpilogue(VF.isVector())) &&2396         "loops not exiting via the latch without required epilogue?");2397 2398  // NOTE: The Plan's scalar preheader VPBB isn't replaced with a VPIRBasicBlock2399  // wrapping the newly created scalar preheader here at the moment, because the2400  // Plan's scalar preheader may be unreachable at this point. Instead it is2401  // replaced in executePlan.2402  return SplitBlock(VectorPH, VectorPH->getTerminator(), DT, LI, nullptr,2403                    Twine(Prefix) + "scalar.ph");2404}2405 2406/// Return the expanded step for \p ID using \p ExpandedSCEVs to look up SCEV2407/// expansion results.2408static Value *getExpandedStep(const InductionDescriptor &ID,2409                              const SCEV2ValueTy &ExpandedSCEVs) {2410  const SCEV *Step = ID.getStep();2411  if (auto *C = dyn_cast<SCEVConstant>(Step))2412    return C->getValue();2413  if (auto *U = dyn_cast<SCEVUnknown>(Step))2414    return U->getValue();2415  Value *V = ExpandedSCEVs.lookup(Step);2416  assert(V && "SCEV must be expanded at this point");2417  return V;2418}2419 2420/// Knowing that loop \p L executes a single vector iteration, add instructions2421/// that will get simplified and thus should not have any cost to \p2422/// InstsToIgnore.2423static void addFullyUnrolledInstructionsToIgnore(2424    Loop *L, const LoopVectorizationLegality::InductionList &IL,2425    SmallPtrSetImpl<Instruction *> &InstsToIgnore) {2426  auto *Cmp = L->getLatchCmpInst();2427  if (Cmp)2428    InstsToIgnore.insert(Cmp);2429  for (const auto &KV : IL) {2430    // Extract the key by hand so that it can be used in the lambda below.  Note2431    // that captured structured bindings are a C++20 extension.2432    const PHINode *IV = KV.first;2433 2434    // Get next iteration value of the induction variable.2435    Instruction *IVInst =2436        cast<Instruction>(IV->getIncomingValueForBlock(L->getLoopLatch()));2437    if (all_of(IVInst->users(),2438               [&](const User *U) { return U == IV || U == Cmp; }))2439      InstsToIgnore.insert(IVInst);2440  }2441}2442 2443BasicBlock *InnerLoopVectorizer::createVectorizedLoopSkeleton() {2444  // Create a new IR basic block for the scalar preheader.2445  BasicBlock *ScalarPH = createScalarPreheader("");2446  return ScalarPH->getSinglePredecessor();2447}2448 2449namespace {2450 2451struct CSEDenseMapInfo {2452  static bool canHandle(const Instruction *I) {2453    return isa<InsertElementInst>(I) || isa<ExtractElementInst>(I) ||2454           isa<ShuffleVectorInst>(I) || isa<GetElementPtrInst>(I);2455  }2456 2457  static inline Instruction *getEmptyKey() {2458    return DenseMapInfo<Instruction *>::getEmptyKey();2459  }2460 2461  static inline Instruction *getTombstoneKey() {2462    return DenseMapInfo<Instruction *>::getTombstoneKey();2463  }2464 2465  static unsigned getHashValue(const Instruction *I) {2466    assert(canHandle(I) && "Unknown instruction!");2467    return hash_combine(I->getOpcode(),2468                        hash_combine_range(I->operand_values()));2469  }2470 2471  static bool isEqual(const Instruction *LHS, const Instruction *RHS) {2472    if (LHS == getEmptyKey() || RHS == getEmptyKey() ||2473        LHS == getTombstoneKey() || RHS == getTombstoneKey())2474      return LHS == RHS;2475    return LHS->isIdenticalTo(RHS);2476  }2477};2478 2479} // end anonymous namespace2480 2481/// FIXME: This legacy common-subexpression-elimination routine is scheduled for2482/// removal, in favor of the VPlan-based one.2483static void legacyCSE(BasicBlock *BB) {2484  // Perform simple cse.2485  SmallDenseMap<Instruction *, Instruction *, 4, CSEDenseMapInfo> CSEMap;2486  for (Instruction &In : llvm::make_early_inc_range(*BB)) {2487    if (!CSEDenseMapInfo::canHandle(&In))2488      continue;2489 2490    // Check if we can replace this instruction with any of the2491    // visited instructions.2492    if (Instruction *V = CSEMap.lookup(&In)) {2493      In.replaceAllUsesWith(V);2494      In.eraseFromParent();2495      continue;2496    }2497 2498    CSEMap[&In] = &In;2499  }2500}2501 2502/// This function attempts to return a value that represents the ElementCount2503/// at runtime. For fixed-width VFs we know this precisely at compile2504/// time, but for scalable VFs we calculate it based on an estimate of the2505/// vscale value.2506static unsigned estimateElementCount(ElementCount VF,2507                                     std::optional<unsigned> VScale) {2508  unsigned EstimatedVF = VF.getKnownMinValue();2509  if (VF.isScalable())2510    if (VScale)2511      EstimatedVF *= *VScale;2512  assert(EstimatedVF >= 1 && "Estimated VF shouldn't be less than 1");2513  return EstimatedVF;2514}2515 2516InstructionCost2517LoopVectorizationCostModel::getVectorCallCost(CallInst *CI,2518                                              ElementCount VF) const {2519  // We only need to calculate a cost if the VF is scalar; for actual vectors2520  // we should already have a pre-calculated cost at each VF.2521  if (!VF.isScalar())2522    return getCallWideningDecision(CI, VF).Cost;2523 2524  Type *RetTy = CI->getType();2525  if (RecurrenceDescriptor::isFMulAddIntrinsic(CI))2526    if (auto RedCost = getReductionPatternCost(CI, VF, RetTy))2527      return *RedCost;2528 2529  SmallVector<Type *, 4> Tys;2530  for (auto &ArgOp : CI->args())2531    Tys.push_back(ArgOp->getType());2532 2533  InstructionCost ScalarCallCost =2534      TTI.getCallInstrCost(CI->getCalledFunction(), RetTy, Tys, CostKind);2535 2536  // If this is an intrinsic we may have a lower cost for it.2537  if (getVectorIntrinsicIDForCall(CI, TLI)) {2538    InstructionCost IntrinsicCost = getVectorIntrinsicCost(CI, VF);2539    return std::min(ScalarCallCost, IntrinsicCost);2540  }2541  return ScalarCallCost;2542}2543 2544static Type *maybeVectorizeType(Type *Ty, ElementCount VF) {2545  if (VF.isScalar() || !canVectorizeTy(Ty))2546    return Ty;2547  return toVectorizedTy(Ty, VF);2548}2549 2550InstructionCost2551LoopVectorizationCostModel::getVectorIntrinsicCost(CallInst *CI,2552                                                   ElementCount VF) const {2553  Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);2554  assert(ID && "Expected intrinsic call!");2555  Type *RetTy = maybeVectorizeType(CI->getType(), VF);2556  FastMathFlags FMF;2557  if (auto *FPMO = dyn_cast<FPMathOperator>(CI))2558    FMF = FPMO->getFastMathFlags();2559 2560  SmallVector<const Value *> Arguments(CI->args());2561  FunctionType *FTy = CI->getCalledFunction()->getFunctionType();2562  SmallVector<Type *> ParamTys;2563  std::transform(FTy->param_begin(), FTy->param_end(),2564                 std::back_inserter(ParamTys),2565                 [&](Type *Ty) { return maybeVectorizeType(Ty, VF); });2566 2567  IntrinsicCostAttributes CostAttrs(ID, RetTy, Arguments, ParamTys, FMF,2568                                    dyn_cast<IntrinsicInst>(CI),2569                                    InstructionCost::getInvalid(), TLI);2570  return TTI.getIntrinsicInstrCost(CostAttrs, CostKind);2571}2572 2573void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) {2574  // Fix widened non-induction PHIs by setting up the PHI operands.2575  fixNonInductionPHIs(State);2576 2577  // Don't apply optimizations below when no (vector) loop remains, as they all2578  // require one at the moment.2579  VPBasicBlock *HeaderVPBB =2580      vputils::getFirstLoopHeader(*State.Plan, State.VPDT);2581  if (!HeaderVPBB)2582    return;2583 2584  BasicBlock *HeaderBB = State.CFG.VPBB2IRBB[HeaderVPBB];2585 2586  // Remove redundant induction instructions.2587  legacyCSE(HeaderBB);2588}2589 2590void InnerLoopVectorizer::fixNonInductionPHIs(VPTransformState &State) {2591  auto Iter = vp_depth_first_shallow(Plan.getEntry());2592  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) {2593    for (VPRecipeBase &P : VPBB->phis()) {2594      VPWidenPHIRecipe *VPPhi = dyn_cast<VPWidenPHIRecipe>(&P);2595      if (!VPPhi)2596        continue;2597      PHINode *NewPhi = cast<PHINode>(State.get(VPPhi));2598      // Make sure the builder has a valid insert point.2599      Builder.SetInsertPoint(NewPhi);2600      for (const auto &[Inc, VPBB] : VPPhi->incoming_values_and_blocks())2601        NewPhi->addIncoming(State.get(Inc), State.CFG.VPBB2IRBB[VPBB]);2602    }2603  }2604}2605 2606void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) {2607  // We should not collect Scalars more than once per VF. Right now, this2608  // function is called from collectUniformsAndScalars(), which already does2609  // this check. Collecting Scalars for VF=1 does not make any sense.2610  assert(VF.isVector() && !Scalars.contains(VF) &&2611         "This function should not be visited twice for the same VF");2612 2613  // This avoids any chances of creating a REPLICATE recipe during planning2614  // since that would result in generation of scalarized code during execution,2615  // which is not supported for scalable vectors.2616  if (VF.isScalable()) {2617    Scalars[VF].insert_range(Uniforms[VF]);2618    return;2619  }2620 2621  SmallSetVector<Instruction *, 8> Worklist;2622 2623  // These sets are used to seed the analysis with pointers used by memory2624  // accesses that will remain scalar.2625  SmallSetVector<Instruction *, 8> ScalarPtrs;2626  SmallPtrSet<Instruction *, 8> PossibleNonScalarPtrs;2627  auto *Latch = TheLoop->getLoopLatch();2628 2629  // A helper that returns true if the use of Ptr by MemAccess will be scalar.2630  // The pointer operands of loads and stores will be scalar as long as the2631  // memory access is not a gather or scatter operation. The value operand of a2632  // store will remain scalar if the store is scalarized.2633  auto IsScalarUse = [&](Instruction *MemAccess, Value *Ptr) {2634    InstWidening WideningDecision = getWideningDecision(MemAccess, VF);2635    assert(WideningDecision != CM_Unknown &&2636           "Widening decision should be ready at this moment");2637    if (auto *Store = dyn_cast<StoreInst>(MemAccess))2638      if (Ptr == Store->getValueOperand())2639        return WideningDecision == CM_Scalarize;2640    assert(Ptr == getLoadStorePointerOperand(MemAccess) &&2641           "Ptr is neither a value or pointer operand");2642    return WideningDecision != CM_GatherScatter;2643  };2644 2645  // A helper that returns true if the given value is a getelementptr2646  // instruction contained in the loop.2647  auto IsLoopVaryingGEP = [&](Value *V) {2648    return isa<GetElementPtrInst>(V) && !TheLoop->isLoopInvariant(V);2649  };2650 2651  // A helper that evaluates a memory access's use of a pointer. If the use will2652  // be a scalar use and the pointer is only used by memory accesses, we place2653  // the pointer in ScalarPtrs. Otherwise, the pointer is placed in2654  // PossibleNonScalarPtrs.2655  auto EvaluatePtrUse = [&](Instruction *MemAccess, Value *Ptr) {2656    // We only care about bitcast and getelementptr instructions contained in2657    // the loop.2658    if (!IsLoopVaryingGEP(Ptr))2659      return;2660 2661    // If the pointer has already been identified as scalar (e.g., if it was2662    // also identified as uniform), there's nothing to do.2663    auto *I = cast<Instruction>(Ptr);2664    if (Worklist.count(I))2665      return;2666 2667    // If the use of the pointer will be a scalar use, and all users of the2668    // pointer are memory accesses, place the pointer in ScalarPtrs. Otherwise,2669    // place the pointer in PossibleNonScalarPtrs.2670    if (IsScalarUse(MemAccess, Ptr) &&2671        all_of(I->users(), IsaPred<LoadInst, StoreInst>))2672      ScalarPtrs.insert(I);2673    else2674      PossibleNonScalarPtrs.insert(I);2675  };2676 2677  // We seed the scalars analysis with three classes of instructions: (1)2678  // instructions marked uniform-after-vectorization and (2) bitcast,2679  // getelementptr and (pointer) phi instructions used by memory accesses2680  // requiring a scalar use.2681  //2682  // (1) Add to the worklist all instructions that have been identified as2683  // uniform-after-vectorization.2684  Worklist.insert_range(Uniforms[VF]);2685 2686  // (2) Add to the worklist all bitcast and getelementptr instructions used by2687  // memory accesses requiring a scalar use. The pointer operands of loads and2688  // stores will be scalar unless the operation is a gather or scatter.2689  // The value operand of a store will remain scalar if the store is scalarized.2690  for (auto *BB : TheLoop->blocks())2691    for (auto &I : *BB) {2692      if (auto *Load = dyn_cast<LoadInst>(&I)) {2693        EvaluatePtrUse(Load, Load->getPointerOperand());2694      } else if (auto *Store = dyn_cast<StoreInst>(&I)) {2695        EvaluatePtrUse(Store, Store->getPointerOperand());2696        EvaluatePtrUse(Store, Store->getValueOperand());2697      }2698    }2699  for (auto *I : ScalarPtrs)2700    if (!PossibleNonScalarPtrs.count(I)) {2701      LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *I << "\n");2702      Worklist.insert(I);2703    }2704 2705  // Insert the forced scalars.2706  // FIXME: Currently VPWidenPHIRecipe() often creates a dead vector2707  // induction variable when the PHI user is scalarized.2708  auto ForcedScalar = ForcedScalars.find(VF);2709  if (ForcedScalar != ForcedScalars.end())2710    for (auto *I : ForcedScalar->second) {2711      LLVM_DEBUG(dbgs() << "LV: Found (forced) scalar instruction: " << *I << "\n");2712      Worklist.insert(I);2713    }2714 2715  // Expand the worklist by looking through any bitcasts and getelementptr2716  // instructions we've already identified as scalar. This is similar to the2717  // expansion step in collectLoopUniforms(); however, here we're only2718  // expanding to include additional bitcasts and getelementptr instructions.2719  unsigned Idx = 0;2720  while (Idx != Worklist.size()) {2721    Instruction *Dst = Worklist[Idx++];2722    if (!IsLoopVaryingGEP(Dst->getOperand(0)))2723      continue;2724    auto *Src = cast<Instruction>(Dst->getOperand(0));2725    if (llvm::all_of(Src->users(), [&](User *U) -> bool {2726          auto *J = cast<Instruction>(U);2727          return !TheLoop->contains(J) || Worklist.count(J) ||2728                 ((isa<LoadInst>(J) || isa<StoreInst>(J)) &&2729                  IsScalarUse(J, Src));2730        })) {2731      Worklist.insert(Src);2732      LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Src << "\n");2733    }2734  }2735 2736  // An induction variable will remain scalar if all users of the induction2737  // variable and induction variable update remain scalar.2738  for (const auto &Induction : Legal->getInductionVars()) {2739    auto *Ind = Induction.first;2740    auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));2741 2742    // If tail-folding is applied, the primary induction variable will be used2743    // to feed a vector compare.2744    if (Ind == Legal->getPrimaryInduction() && foldTailByMasking())2745      continue;2746 2747    // Returns true if \p Indvar is a pointer induction that is used directly by2748    // load/store instruction \p I.2749    auto IsDirectLoadStoreFromPtrIndvar = [&](Instruction *Indvar,2750                                              Instruction *I) {2751      return Induction.second.getKind() ==2752                 InductionDescriptor::IK_PtrInduction &&2753             (isa<LoadInst>(I) || isa<StoreInst>(I)) &&2754             Indvar == getLoadStorePointerOperand(I) && IsScalarUse(I, Indvar);2755    };2756 2757    // Determine if all users of the induction variable are scalar after2758    // vectorization.2759    bool ScalarInd = all_of(Ind->users(), [&](User *U) -> bool {2760      auto *I = cast<Instruction>(U);2761      return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||2762             IsDirectLoadStoreFromPtrIndvar(Ind, I);2763    });2764    if (!ScalarInd)2765      continue;2766 2767    // If the induction variable update is a fixed-order recurrence, neither the2768    // induction variable or its update should be marked scalar after2769    // vectorization.2770    auto *IndUpdatePhi = dyn_cast<PHINode>(IndUpdate);2771    if (IndUpdatePhi && Legal->isFixedOrderRecurrence(IndUpdatePhi))2772      continue;2773 2774    // Determine if all users of the induction variable update instruction are2775    // scalar after vectorization.2776    bool ScalarIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {2777      auto *I = cast<Instruction>(U);2778      return I == Ind || !TheLoop->contains(I) || Worklist.count(I) ||2779             IsDirectLoadStoreFromPtrIndvar(IndUpdate, I);2780    });2781    if (!ScalarIndUpdate)2782      continue;2783 2784    // The induction variable and its update instruction will remain scalar.2785    Worklist.insert(Ind);2786    Worklist.insert(IndUpdate);2787    LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *Ind << "\n");2788    LLVM_DEBUG(dbgs() << "LV: Found scalar instruction: " << *IndUpdate2789                      << "\n");2790  }2791 2792  Scalars[VF].insert_range(Worklist);2793}2794 2795bool LoopVectorizationCostModel::isScalarWithPredication(2796    Instruction *I, ElementCount VF) const {2797  if (!isPredicatedInst(I))2798    return false;2799 2800  // Do we have a non-scalar lowering for this predicated2801  // instruction? No - it is scalar with predication.2802  switch(I->getOpcode()) {2803  default:2804    return true;2805  case Instruction::Call:2806    if (VF.isScalar())2807      return true;2808    return getCallWideningDecision(cast<CallInst>(I), VF).Kind == CM_Scalarize;2809  case Instruction::Load:2810  case Instruction::Store: {2811    auto *Ptr = getLoadStorePointerOperand(I);2812    auto *Ty = getLoadStoreType(I);2813    unsigned AS = getLoadStoreAddressSpace(I);2814    Type *VTy = Ty;2815    if (VF.isVector())2816      VTy = VectorType::get(Ty, VF);2817    const Align Alignment = getLoadStoreAlignment(I);2818    return isa<LoadInst>(I) ? !(isLegalMaskedLoad(Ty, Ptr, Alignment, AS) ||2819                                TTI.isLegalMaskedGather(VTy, Alignment))2820                            : !(isLegalMaskedStore(Ty, Ptr, Alignment, AS) ||2821                                TTI.isLegalMaskedScatter(VTy, Alignment));2822  }2823  case Instruction::UDiv:2824  case Instruction::SDiv:2825  case Instruction::SRem:2826  case Instruction::URem: {2827    // We have the option to use the safe-divisor idiom to avoid predication.2828    // The cost based decision here will always select safe-divisor for2829    // scalable vectors as scalarization isn't legal.2830    const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF);2831    return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost);2832  }2833  }2834}2835 2836// TODO: Fold into LoopVectorizationLegality::isMaskRequired.2837bool LoopVectorizationCostModel::isPredicatedInst(Instruction *I) const {2838  // TODO: We can use the loop-preheader as context point here and get2839  // context sensitive reasoning for isSafeToSpeculativelyExecute.2840  if (isSafeToSpeculativelyExecute(I) ||2841      (isa<LoadInst, StoreInst, CallInst>(I) && !Legal->isMaskRequired(I)) ||2842      isa<BranchInst, SwitchInst, PHINode, AllocaInst>(I))2843    return false;2844 2845  // If the instruction was executed conditionally in the original scalar loop,2846  // predication is needed with a mask whose lanes are all possibly inactive.2847  if (Legal->blockNeedsPredication(I->getParent()))2848    return true;2849 2850  // If we're not folding the tail by masking, predication is unnecessary.2851  if (!foldTailByMasking())2852    return false;2853 2854  // All that remain are instructions with side-effects originally executed in2855  // the loop unconditionally, but now execute under a tail-fold mask (only)2856  // having at least one active lane (the first). If the side-effects of the2857  // instruction are invariant, executing it w/o (the tail-folding) mask is safe2858  // - it will cause the same side-effects as when masked.2859  switch(I->getOpcode()) {2860  default:2861    llvm_unreachable(2862        "instruction should have been considered by earlier checks");2863  case Instruction::Call:2864    // Side-effects of a Call are assumed to be non-invariant, needing a2865    // (fold-tail) mask.2866    assert(Legal->isMaskRequired(I) &&2867           "should have returned earlier for calls not needing a mask");2868    return true;2869  case Instruction::Load:2870    // If the address is loop invariant no predication is needed.2871    return !Legal->isInvariant(getLoadStorePointerOperand(I));2872  case Instruction::Store: {2873    // For stores, we need to prove both speculation safety (which follows from2874    // the same argument as loads), but also must prove the value being stored2875    // is correct.  The easiest form of the later is to require that all values2876    // stored are the same.2877    return !(Legal->isInvariant(getLoadStorePointerOperand(I)) &&2878             TheLoop->isLoopInvariant(cast<StoreInst>(I)->getValueOperand()));2879  }2880  case Instruction::UDiv:2881  case Instruction::SDiv:2882  case Instruction::SRem:2883  case Instruction::URem:2884    // If the divisor is loop-invariant no predication is needed.2885    return !Legal->isInvariant(I->getOperand(1));2886  }2887}2888 2889std::pair<InstructionCost, InstructionCost>2890LoopVectorizationCostModel::getDivRemSpeculationCost(Instruction *I,2891                                                    ElementCount VF) const {2892  assert(I->getOpcode() == Instruction::UDiv ||2893         I->getOpcode() == Instruction::SDiv ||2894         I->getOpcode() == Instruction::SRem ||2895         I->getOpcode() == Instruction::URem);2896  assert(!isSafeToSpeculativelyExecute(I));2897 2898  // Scalarization isn't legal for scalable vector types2899  InstructionCost ScalarizationCost = InstructionCost::getInvalid();2900  if (!VF.isScalable()) {2901    // Get the scalarization cost and scale this amount by the probability of2902    // executing the predicated block. If the instruction is not predicated,2903    // we fall through to the next case.2904    ScalarizationCost = 0;2905 2906    // These instructions have a non-void type, so account for the phi nodes2907    // that we will create. This cost is likely to be zero. The phi node2908    // cost, if any, should be scaled by the block probability because it2909    // models a copy at the end of each predicated block.2910    ScalarizationCost +=2911        VF.getFixedValue() * TTI.getCFInstrCost(Instruction::PHI, CostKind);2912 2913    // The cost of the non-predicated instruction.2914    ScalarizationCost +=2915        VF.getFixedValue() *2916        TTI.getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind);2917 2918    // The cost of insertelement and extractelement instructions needed for2919    // scalarization.2920    ScalarizationCost += getScalarizationOverhead(I, VF);2921 2922    // Scale the cost by the probability of executing the predicated blocks.2923    // This assumes the predicated block for each vector lane is equally2924    // likely.2925    ScalarizationCost =2926        ScalarizationCost / getPredBlockCostDivisor(CostKind, I->getParent());2927  }2928 2929  InstructionCost SafeDivisorCost = 0;2930  auto *VecTy = toVectorTy(I->getType(), VF);2931  // The cost of the select guard to ensure all lanes are well defined2932  // after we speculate above any internal control flow.2933  SafeDivisorCost +=2934      TTI.getCmpSelInstrCost(Instruction::Select, VecTy,2935                             toVectorTy(Type::getInt1Ty(I->getContext()), VF),2936                             CmpInst::BAD_ICMP_PREDICATE, CostKind);2937 2938  SmallVector<const Value *, 4> Operands(I->operand_values());2939  SafeDivisorCost += TTI.getArithmeticInstrCost(2940      I->getOpcode(), VecTy, CostKind,2941      {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None},2942      {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None},2943      Operands, I);2944  return {ScalarizationCost, SafeDivisorCost};2945}2946 2947bool LoopVectorizationCostModel::interleavedAccessCanBeWidened(2948    Instruction *I, ElementCount VF) const {2949  assert(isAccessInterleaved(I) && "Expecting interleaved access.");2950  assert(getWideningDecision(I, VF) == CM_Unknown &&2951         "Decision should not be set yet.");2952  auto *Group = getInterleavedAccessGroup(I);2953  assert(Group && "Must have a group.");2954  unsigned InterleaveFactor = Group->getFactor();2955 2956  // If the instruction's allocated size doesn't equal its type size, it2957  // requires padding and will be scalarized.2958  auto &DL = I->getDataLayout();2959  auto *ScalarTy = getLoadStoreType(I);2960  if (hasIrregularType(ScalarTy, DL))2961    return false;2962 2963  // For scalable vectors, the interleave factors must be <= 8 since we require2964  // the (de)interleaveN intrinsics instead of shufflevectors.2965  if (VF.isScalable() && InterleaveFactor > 8)2966    return false;2967 2968  // If the group involves a non-integral pointer, we may not be able to2969  // losslessly cast all values to a common type.2970  bool ScalarNI = DL.isNonIntegralPointerType(ScalarTy);2971  for (unsigned Idx = 0; Idx < InterleaveFactor; Idx++) {2972    Instruction *Member = Group->getMember(Idx);2973    if (!Member)2974      continue;2975    auto *MemberTy = getLoadStoreType(Member);2976    bool MemberNI = DL.isNonIntegralPointerType(MemberTy);2977    // Don't coerce non-integral pointers to integers or vice versa.2978    if (MemberNI != ScalarNI)2979      // TODO: Consider adding special nullptr value case here2980      return false;2981    if (MemberNI && ScalarNI &&2982        ScalarTy->getPointerAddressSpace() !=2983            MemberTy->getPointerAddressSpace())2984      return false;2985  }2986 2987  // Check if masking is required.2988  // A Group may need masking for one of two reasons: it resides in a block that2989  // needs predication, or it was decided to use masking to deal with gaps2990  // (either a gap at the end of a load-access that may result in a speculative2991  // load, or any gaps in a store-access).2992  bool PredicatedAccessRequiresMasking =2993      blockNeedsPredicationForAnyReason(I->getParent()) &&2994      Legal->isMaskRequired(I);2995  bool LoadAccessWithGapsRequiresEpilogMasking =2996      isa<LoadInst>(I) && Group->requiresScalarEpilogue() &&2997      !isScalarEpilogueAllowed();2998  bool StoreAccessWithGapsRequiresMasking =2999      isa<StoreInst>(I) && !Group->isFull();3000  if (!PredicatedAccessRequiresMasking &&3001      !LoadAccessWithGapsRequiresEpilogMasking &&3002      !StoreAccessWithGapsRequiresMasking)3003    return true;3004 3005  // If masked interleaving is required, we expect that the user/target had3006  // enabled it, because otherwise it either wouldn't have been created or3007  // it should have been invalidated by the CostModel.3008  assert(useMaskedInterleavedAccesses(TTI) &&3009         "Masked interleave-groups for predicated accesses are not enabled.");3010 3011  if (Group->isReverse())3012    return false;3013 3014  // TODO: Support interleaved access that requires a gap mask for scalable VFs.3015  bool NeedsMaskForGaps = LoadAccessWithGapsRequiresEpilogMasking ||3016                          StoreAccessWithGapsRequiresMasking;3017  if (VF.isScalable() && NeedsMaskForGaps)3018    return false;3019 3020  auto *Ty = getLoadStoreType(I);3021  const Align Alignment = getLoadStoreAlignment(I);3022  unsigned AS = getLoadStoreAddressSpace(I);3023  return isa<LoadInst>(I) ? TTI.isLegalMaskedLoad(Ty, Alignment, AS)3024                          : TTI.isLegalMaskedStore(Ty, Alignment, AS);3025}3026 3027bool LoopVectorizationCostModel::memoryInstructionCanBeWidened(3028    Instruction *I, ElementCount VF) {3029  // Get and ensure we have a valid memory instruction.3030  assert((isa<LoadInst, StoreInst>(I)) && "Invalid memory instruction");3031 3032  auto *Ptr = getLoadStorePointerOperand(I);3033  auto *ScalarTy = getLoadStoreType(I);3034 3035  // In order to be widened, the pointer should be consecutive, first of all.3036  if (!Legal->isConsecutivePtr(ScalarTy, Ptr))3037    return false;3038 3039  // If the instruction is a store located in a predicated block, it will be3040  // scalarized.3041  if (isScalarWithPredication(I, VF))3042    return false;3043 3044  // If the instruction's allocated size doesn't equal it's type size, it3045  // requires padding and will be scalarized.3046  auto &DL = I->getDataLayout();3047  if (hasIrregularType(ScalarTy, DL))3048    return false;3049 3050  return true;3051}3052 3053void LoopVectorizationCostModel::collectLoopUniforms(ElementCount VF) {3054  // We should not collect Uniforms more than once per VF. Right now,3055  // this function is called from collectUniformsAndScalars(), which3056  // already does this check. Collecting Uniforms for VF=1 does not make any3057  // sense.3058 3059  assert(VF.isVector() && !Uniforms.contains(VF) &&3060         "This function should not be visited twice for the same VF");3061 3062  // Visit the list of Uniforms. If we find no uniform value, we won't3063  // analyze again.  Uniforms.count(VF) will return 1.3064  Uniforms[VF].clear();3065 3066  // Now we know that the loop is vectorizable!3067  // Collect instructions inside the loop that will remain uniform after3068  // vectorization.3069 3070  // Global values, params and instructions outside of current loop are out of3071  // scope.3072  auto IsOutOfScope = [&](Value *V) -> bool {3073    Instruction *I = dyn_cast<Instruction>(V);3074    return (!I || !TheLoop->contains(I));3075  };3076 3077  // Worklist containing uniform instructions demanding lane 0.3078  SetVector<Instruction *> Worklist;3079 3080  // Add uniform instructions demanding lane 0 to the worklist. Instructions3081  // that require predication must not be considered uniform after3082  // vectorization, because that would create an erroneous replicating region3083  // where only a single instance out of VF should be formed.3084  auto AddToWorklistIfAllowed = [&](Instruction *I) -> void {3085    if (IsOutOfScope(I)) {3086      LLVM_DEBUG(dbgs() << "LV: Found not uniform due to scope: "3087                        << *I << "\n");3088      return;3089    }3090    if (isPredicatedInst(I)) {3091      LLVM_DEBUG(3092          dbgs() << "LV: Found not uniform due to requiring predication: " << *I3093                 << "\n");3094      return;3095    }3096    LLVM_DEBUG(dbgs() << "LV: Found uniform instruction: " << *I << "\n");3097    Worklist.insert(I);3098  };3099 3100  // Start with the conditional branches exiting the loop. If the branch3101  // condition is an instruction contained in the loop that is only used by the3102  // branch, it is uniform. Note conditions from uncountable early exits are not3103  // uniform.3104  SmallVector<BasicBlock *> Exiting;3105  TheLoop->getExitingBlocks(Exiting);3106  for (BasicBlock *E : Exiting) {3107    if (Legal->hasUncountableEarlyExit() && TheLoop->getLoopLatch() != E)3108      continue;3109    auto *Cmp = dyn_cast<Instruction>(E->getTerminator()->getOperand(0));3110    if (Cmp && TheLoop->contains(Cmp) && Cmp->hasOneUse())3111      AddToWorklistIfAllowed(Cmp);3112  }3113 3114  auto PrevVF = VF.divideCoefficientBy(2);3115  // Return true if all lanes perform the same memory operation, and we can3116  // thus choose to execute only one.3117  auto IsUniformMemOpUse = [&](Instruction *I) {3118    // If the value was already known to not be uniform for the previous3119    // (smaller VF), it cannot be uniform for the larger VF.3120    if (PrevVF.isVector()) {3121      auto Iter = Uniforms.find(PrevVF);3122      if (Iter != Uniforms.end() && !Iter->second.contains(I))3123        return false;3124    }3125    if (!Legal->isUniformMemOp(*I, VF))3126      return false;3127    if (isa<LoadInst>(I))3128      // Loading the same address always produces the same result - at least3129      // assuming aliasing and ordering which have already been checked.3130      return true;3131    // Storing the same value on every iteration.3132    return TheLoop->isLoopInvariant(cast<StoreInst>(I)->getValueOperand());3133  };3134 3135  auto IsUniformDecision = [&](Instruction *I, ElementCount VF) {3136    InstWidening WideningDecision = getWideningDecision(I, VF);3137    assert(WideningDecision != CM_Unknown &&3138           "Widening decision should be ready at this moment");3139 3140    if (IsUniformMemOpUse(I))3141      return true;3142 3143    return (WideningDecision == CM_Widen ||3144            WideningDecision == CM_Widen_Reverse ||3145            WideningDecision == CM_Interleave);3146  };3147 3148  // Returns true if Ptr is the pointer operand of a memory access instruction3149  // I, I is known to not require scalarization, and the pointer is not also3150  // stored.3151  auto IsVectorizedMemAccessUse = [&](Instruction *I, Value *Ptr) -> bool {3152    if (isa<StoreInst>(I) && I->getOperand(0) == Ptr)3153      return false;3154    return getLoadStorePointerOperand(I) == Ptr &&3155           (IsUniformDecision(I, VF) || Legal->isInvariant(Ptr));3156  };3157 3158  // Holds a list of values which are known to have at least one uniform use.3159  // Note that there may be other uses which aren't uniform.  A "uniform use"3160  // here is something which only demands lane 0 of the unrolled iterations;3161  // it does not imply that all lanes produce the same value (e.g. this is not3162  // the usual meaning of uniform)3163  SetVector<Value *> HasUniformUse;3164 3165  // Scan the loop for instructions which are either a) known to have only3166  // lane 0 demanded or b) are uses which demand only lane 0 of their operand.3167  for (auto *BB : TheLoop->blocks())3168    for (auto &I : *BB) {3169      if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I)) {3170        switch (II->getIntrinsicID()) {3171        case Intrinsic::sideeffect:3172        case Intrinsic::experimental_noalias_scope_decl:3173        case Intrinsic::assume:3174        case Intrinsic::lifetime_start:3175        case Intrinsic::lifetime_end:3176          if (TheLoop->hasLoopInvariantOperands(&I))3177            AddToWorklistIfAllowed(&I);3178          break;3179        default:3180          break;3181        }3182      }3183 3184      if (auto *EVI = dyn_cast<ExtractValueInst>(&I)) {3185        if (IsOutOfScope(EVI->getAggregateOperand())) {3186          AddToWorklistIfAllowed(EVI);3187          continue;3188        }3189        // Only ExtractValue instructions where the aggregate value comes from a3190        // call are allowed to be non-uniform.3191        assert(isa<CallInst>(EVI->getAggregateOperand()) &&3192               "Expected aggregate value to be call return value");3193      }3194 3195      // If there's no pointer operand, there's nothing to do.3196      auto *Ptr = getLoadStorePointerOperand(&I);3197      if (!Ptr)3198        continue;3199 3200      // If the pointer can be proven to be uniform, always add it to the3201      // worklist.3202      if (isa<Instruction>(Ptr) && Legal->isUniform(Ptr, VF))3203        AddToWorklistIfAllowed(cast<Instruction>(Ptr));3204 3205      if (IsUniformMemOpUse(&I))3206        AddToWorklistIfAllowed(&I);3207 3208      if (IsVectorizedMemAccessUse(&I, Ptr))3209        HasUniformUse.insert(Ptr);3210    }3211 3212  // Add to the worklist any operands which have *only* uniform (e.g. lane 03213  // demanding) users.  Since loops are assumed to be in LCSSA form, this3214  // disallows uses outside the loop as well.3215  for (auto *V : HasUniformUse) {3216    if (IsOutOfScope(V))3217      continue;3218    auto *I = cast<Instruction>(V);3219    bool UsersAreMemAccesses = all_of(I->users(), [&](User *U) -> bool {3220      auto *UI = cast<Instruction>(U);3221      return TheLoop->contains(UI) && IsVectorizedMemAccessUse(UI, V);3222    });3223    if (UsersAreMemAccesses)3224      AddToWorklistIfAllowed(I);3225  }3226 3227  // Expand Worklist in topological order: whenever a new instruction3228  // is added , its users should be already inside Worklist.  It ensures3229  // a uniform instruction will only be used by uniform instructions.3230  unsigned Idx = 0;3231  while (Idx != Worklist.size()) {3232    Instruction *I = Worklist[Idx++];3233 3234    for (auto *OV : I->operand_values()) {3235      // isOutOfScope operands cannot be uniform instructions.3236      if (IsOutOfScope(OV))3237        continue;3238      // First order recurrence Phi's should typically be considered3239      // non-uniform.3240      auto *OP = dyn_cast<PHINode>(OV);3241      if (OP && Legal->isFixedOrderRecurrence(OP))3242        continue;3243      // If all the users of the operand are uniform, then add the3244      // operand into the uniform worklist.3245      auto *OI = cast<Instruction>(OV);3246      if (llvm::all_of(OI->users(), [&](User *U) -> bool {3247            auto *J = cast<Instruction>(U);3248            return Worklist.count(J) || IsVectorizedMemAccessUse(J, OI);3249          }))3250        AddToWorklistIfAllowed(OI);3251    }3252  }3253 3254  // For an instruction to be added into Worklist above, all its users inside3255  // the loop should also be in Worklist. However, this condition cannot be3256  // true for phi nodes that form a cyclic dependence. We must process phi3257  // nodes separately. An induction variable will remain uniform if all users3258  // of the induction variable and induction variable update remain uniform.3259  // The code below handles both pointer and non-pointer induction variables.3260  BasicBlock *Latch = TheLoop->getLoopLatch();3261  for (const auto &Induction : Legal->getInductionVars()) {3262    auto *Ind = Induction.first;3263    auto *IndUpdate = cast<Instruction>(Ind->getIncomingValueForBlock(Latch));3264 3265    // Determine if all users of the induction variable are uniform after3266    // vectorization.3267    bool UniformInd = all_of(Ind->users(), [&](User *U) -> bool {3268      auto *I = cast<Instruction>(U);3269      return I == IndUpdate || !TheLoop->contains(I) || Worklist.count(I) ||3270             IsVectorizedMemAccessUse(I, Ind);3271    });3272    if (!UniformInd)3273      continue;3274 3275    // Determine if all users of the induction variable update instruction are3276    // uniform after vectorization.3277    bool UniformIndUpdate = all_of(IndUpdate->users(), [&](User *U) -> bool {3278      auto *I = cast<Instruction>(U);3279      return I == Ind || Worklist.count(I) ||3280             IsVectorizedMemAccessUse(I, IndUpdate);3281    });3282    if (!UniformIndUpdate)3283      continue;3284 3285    // The induction variable and its update instruction will remain uniform.3286    AddToWorklistIfAllowed(Ind);3287    AddToWorklistIfAllowed(IndUpdate);3288  }3289 3290  Uniforms[VF].insert_range(Worklist);3291}3292 3293bool LoopVectorizationCostModel::runtimeChecksRequired() {3294  LLVM_DEBUG(dbgs() << "LV: Performing code size checks.\n");3295 3296  if (Legal->getRuntimePointerChecking()->Need) {3297    reportVectorizationFailure("Runtime ptr check is required with -Os/-Oz",3298        "runtime pointer checks needed. Enable vectorization of this "3299        "loop with '#pragma clang loop vectorize(enable)' when "3300        "compiling with -Os/-Oz",3301        "CantVersionLoopWithOptForSize", ORE, TheLoop);3302    return true;3303  }3304 3305  if (!PSE.getPredicate().isAlwaysTrue()) {3306    reportVectorizationFailure("Runtime SCEV check is required with -Os/-Oz",3307        "runtime SCEV checks needed. Enable vectorization of this "3308        "loop with '#pragma clang loop vectorize(enable)' when "3309        "compiling with -Os/-Oz",3310        "CantVersionLoopWithOptForSize", ORE, TheLoop);3311    return true;3312  }3313 3314  // FIXME: Avoid specializing for stride==1 instead of bailing out.3315  if (!Legal->getLAI()->getSymbolicStrides().empty()) {3316    reportVectorizationFailure("Runtime stride check for small trip count",3317        "runtime stride == 1 checks needed. Enable vectorization of "3318        "this loop without such check by compiling with -Os/-Oz",3319        "CantVersionLoopWithOptForSize", ORE, TheLoop);3320    return true;3321  }3322 3323  return false;3324}3325 3326bool LoopVectorizationCostModel::isScalableVectorizationAllowed() {3327  if (IsScalableVectorizationAllowed)3328    return *IsScalableVectorizationAllowed;3329 3330  IsScalableVectorizationAllowed = false;3331  if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors)3332    return false;3333 3334  if (Hints->isScalableVectorizationDisabled()) {3335    reportVectorizationInfo("Scalable vectorization is explicitly disabled",3336                            "ScalableVectorizationDisabled", ORE, TheLoop);3337    return false;3338  }3339 3340  LLVM_DEBUG(dbgs() << "LV: Scalable vectorization is available\n");3341 3342  auto MaxScalableVF = ElementCount::getScalable(3343      std::numeric_limits<ElementCount::ScalarTy>::max());3344 3345  // Test that the loop-vectorizer can legalize all operations for this MaxVF.3346  // FIXME: While for scalable vectors this is currently sufficient, this should3347  // be replaced by a more detailed mechanism that filters out specific VFs,3348  // instead of invalidating vectorization for a whole set of VFs based on the3349  // MaxVF.3350 3351  // Disable scalable vectorization if the loop contains unsupported reductions.3352  if (!canVectorizeReductions(MaxScalableVF)) {3353    reportVectorizationInfo(3354        "Scalable vectorization not supported for the reduction "3355        "operations found in this loop.",3356        "ScalableVFUnfeasible", ORE, TheLoop);3357    return false;3358  }3359 3360  // Disable scalable vectorization if the loop contains any instructions3361  // with element types not supported for scalable vectors.3362  if (any_of(ElementTypesInLoop, [&](Type *Ty) {3363        return !Ty->isVoidTy() &&3364               !this->TTI.isElementTypeLegalForScalableVector(Ty);3365      })) {3366    reportVectorizationInfo("Scalable vectorization is not supported "3367                            "for all element types found in this loop.",3368                            "ScalableVFUnfeasible", ORE, TheLoop);3369    return false;3370  }3371 3372  if (!Legal->isSafeForAnyVectorWidth() && !getMaxVScale(*TheFunction, TTI)) {3373    reportVectorizationInfo("The target does not provide maximum vscale value "3374                            "for safe distance analysis.",3375                            "ScalableVFUnfeasible", ORE, TheLoop);3376    return false;3377  }3378 3379  IsScalableVectorizationAllowed = true;3380  return true;3381}3382 3383ElementCount3384LoopVectorizationCostModel::getMaxLegalScalableVF(unsigned MaxSafeElements) {3385  if (!isScalableVectorizationAllowed())3386    return ElementCount::getScalable(0);3387 3388  auto MaxScalableVF = ElementCount::getScalable(3389      std::numeric_limits<ElementCount::ScalarTy>::max());3390  if (Legal->isSafeForAnyVectorWidth())3391    return MaxScalableVF;3392 3393  std::optional<unsigned> MaxVScale = getMaxVScale(*TheFunction, TTI);3394  // Limit MaxScalableVF by the maximum safe dependence distance.3395  MaxScalableVF = ElementCount::getScalable(MaxSafeElements / *MaxVScale);3396 3397  if (!MaxScalableVF)3398    reportVectorizationInfo(3399        "Max legal vector width too small, scalable vectorization "3400        "unfeasible.",3401        "ScalableVFUnfeasible", ORE, TheLoop);3402 3403  return MaxScalableVF;3404}3405 3406FixedScalableVFPair LoopVectorizationCostModel::computeFeasibleMaxVF(3407    unsigned MaxTripCount, ElementCount UserVF, bool FoldTailByMasking) {3408  MinBWs = computeMinimumValueSizes(TheLoop->getBlocks(), *DB, &TTI);3409  unsigned SmallestType, WidestType;3410  std::tie(SmallestType, WidestType) = getSmallestAndWidestTypes();3411 3412  // Get the maximum safe dependence distance in bits computed by LAA.3413  // It is computed by MaxVF * sizeOf(type) * 8, where type is taken from3414  // the memory accesses that is most restrictive (involved in the smallest3415  // dependence distance).3416  unsigned MaxSafeElementsPowerOf2 =3417      bit_floor(Legal->getMaxSafeVectorWidthInBits() / WidestType);3418  if (!Legal->isSafeForAnyStoreLoadForwardDistances()) {3419    unsigned SLDist = Legal->getMaxStoreLoadForwardSafeDistanceInBits();3420    MaxSafeElementsPowerOf2 =3421        std::min(MaxSafeElementsPowerOf2, SLDist / WidestType);3422  }3423  auto MaxSafeFixedVF = ElementCount::getFixed(MaxSafeElementsPowerOf2);3424  auto MaxSafeScalableVF = getMaxLegalScalableVF(MaxSafeElementsPowerOf2);3425 3426  if (!Legal->isSafeForAnyVectorWidth())3427    this->MaxSafeElements = MaxSafeElementsPowerOf2;3428 3429  LLVM_DEBUG(dbgs() << "LV: The max safe fixed VF is: " << MaxSafeFixedVF3430                    << ".\n");3431  LLVM_DEBUG(dbgs() << "LV: The max safe scalable VF is: " << MaxSafeScalableVF3432                    << ".\n");3433 3434  // First analyze the UserVF, fall back if the UserVF should be ignored.3435  if (UserVF) {3436    auto MaxSafeUserVF =3437        UserVF.isScalable() ? MaxSafeScalableVF : MaxSafeFixedVF;3438 3439    if (ElementCount::isKnownLE(UserVF, MaxSafeUserVF)) {3440      // If `VF=vscale x N` is safe, then so is `VF=N`3441      if (UserVF.isScalable())3442        return FixedScalableVFPair(3443            ElementCount::getFixed(UserVF.getKnownMinValue()), UserVF);3444 3445      return UserVF;3446    }3447 3448    assert(ElementCount::isKnownGT(UserVF, MaxSafeUserVF));3449 3450    // Only clamp if the UserVF is not scalable. If the UserVF is scalable, it3451    // is better to ignore the hint and let the compiler choose a suitable VF.3452    if (!UserVF.isScalable()) {3453      LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF3454                        << " is unsafe, clamping to max safe VF="3455                        << MaxSafeFixedVF << ".\n");3456      ORE->emit([&]() {3457        return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor",3458                                          TheLoop->getStartLoc(),3459                                          TheLoop->getHeader())3460               << "User-specified vectorization factor "3461               << ore::NV("UserVectorizationFactor", UserVF)3462               << " is unsafe, clamping to maximum safe vectorization factor "3463               << ore::NV("VectorizationFactor", MaxSafeFixedVF);3464      });3465      return MaxSafeFixedVF;3466    }3467 3468    if (!TTI.supportsScalableVectors() && !ForceTargetSupportsScalableVectors) {3469      LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF3470                        << " is ignored because scalable vectors are not "3471                           "available.\n");3472      ORE->emit([&]() {3473        return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor",3474                                          TheLoop->getStartLoc(),3475                                          TheLoop->getHeader())3476               << "User-specified vectorization factor "3477               << ore::NV("UserVectorizationFactor", UserVF)3478               << " is ignored because the target does not support scalable "3479                  "vectors. The compiler will pick a more suitable value.";3480      });3481    } else {3482      LLVM_DEBUG(dbgs() << "LV: User VF=" << UserVF3483                        << " is unsafe. Ignoring scalable UserVF.\n");3484      ORE->emit([&]() {3485        return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationFactor",3486                                          TheLoop->getStartLoc(),3487                                          TheLoop->getHeader())3488               << "User-specified vectorization factor "3489               << ore::NV("UserVectorizationFactor", UserVF)3490               << " is unsafe. Ignoring the hint to let the compiler pick a "3491                  "more suitable value.";3492      });3493    }3494  }3495 3496  LLVM_DEBUG(dbgs() << "LV: The Smallest and Widest types: " << SmallestType3497                    << " / " << WidestType << " bits.\n");3498 3499  FixedScalableVFPair Result(ElementCount::getFixed(1),3500                             ElementCount::getScalable(0));3501  if (auto MaxVF =3502          getMaximizedVFForTarget(MaxTripCount, SmallestType, WidestType,3503                                  MaxSafeFixedVF, FoldTailByMasking))3504    Result.FixedVF = MaxVF;3505 3506  if (auto MaxVF =3507          getMaximizedVFForTarget(MaxTripCount, SmallestType, WidestType,3508                                  MaxSafeScalableVF, FoldTailByMasking))3509    if (MaxVF.isScalable()) {3510      Result.ScalableVF = MaxVF;3511      LLVM_DEBUG(dbgs() << "LV: Found feasible scalable VF = " << MaxVF3512                        << "\n");3513    }3514 3515  return Result;3516}3517 3518FixedScalableVFPair3519LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) {3520  if (Legal->getRuntimePointerChecking()->Need && TTI.hasBranchDivergence()) {3521    // TODO: It may be useful to do since it's still likely to be dynamically3522    // uniform if the target can skip.3523    reportVectorizationFailure(3524        "Not inserting runtime ptr check for divergent target",3525        "runtime pointer checks needed. Not enabled for divergent target",3526        "CantVersionLoopWithDivergentTarget", ORE, TheLoop);3527    return FixedScalableVFPair::getNone();3528  }3529 3530  ScalarEvolution *SE = PSE.getSE();3531  ElementCount TC = getSmallConstantTripCount(SE, TheLoop);3532  unsigned MaxTC = PSE.getSmallConstantMaxTripCount();3533  LLVM_DEBUG(dbgs() << "LV: Found trip count: " << TC << '\n');3534  if (TC != ElementCount::getFixed(MaxTC))3535    LLVM_DEBUG(dbgs() << "LV: Found maximum trip count: " << MaxTC << '\n');3536  if (TC.isScalar()) {3537    reportVectorizationFailure("Single iteration (non) loop",3538        "loop trip count is one, irrelevant for vectorization",3539        "SingleIterationLoop", ORE, TheLoop);3540    return FixedScalableVFPair::getNone();3541  }3542 3543  // If BTC matches the widest induction type and is -1 then the trip count3544  // computation will wrap to 0 and the vector trip count will be 0. Do not try3545  // to vectorize.3546  const SCEV *BTC = SE->getBackedgeTakenCount(TheLoop);3547  if (!isa<SCEVCouldNotCompute>(BTC) &&3548      BTC->getType()->getScalarSizeInBits() >=3549          Legal->getWidestInductionType()->getScalarSizeInBits() &&3550      SE->isKnownPredicate(CmpInst::ICMP_EQ, BTC,3551                           SE->getMinusOne(BTC->getType()))) {3552    reportVectorizationFailure(3553        "Trip count computation wrapped",3554        "backedge-taken count is -1, loop trip count wrapped to 0",3555        "TripCountWrapped", ORE, TheLoop);3556    return FixedScalableVFPair::getNone();3557  }3558 3559  switch (ScalarEpilogueStatus) {3560  case CM_ScalarEpilogueAllowed:3561    return computeFeasibleMaxVF(MaxTC, UserVF, false);3562  case CM_ScalarEpilogueNotAllowedUsePredicate:3563    [[fallthrough]];3564  case CM_ScalarEpilogueNotNeededUsePredicate:3565    LLVM_DEBUG(3566        dbgs() << "LV: vector predicate hint/switch found.\n"3567               << "LV: Not allowing scalar epilogue, creating predicated "3568               << "vector loop.\n");3569    break;3570  case CM_ScalarEpilogueNotAllowedLowTripLoop:3571    // fallthrough as a special case of OptForSize3572  case CM_ScalarEpilogueNotAllowedOptSize:3573    if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedOptSize)3574      LLVM_DEBUG(3575          dbgs() << "LV: Not allowing scalar epilogue due to -Os/-Oz.\n");3576    else3577      LLVM_DEBUG(dbgs() << "LV: Not allowing scalar epilogue due to low trip "3578                        << "count.\n");3579 3580    // Bail if runtime checks are required, which are not good when optimising3581    // for size.3582    if (runtimeChecksRequired())3583      return FixedScalableVFPair::getNone();3584 3585    break;3586  }3587 3588  // Now try the tail folding3589 3590  // Invalidate interleave groups that require an epilogue if we can't mask3591  // the interleave-group.3592  if (!useMaskedInterleavedAccesses(TTI)) {3593    assert(WideningDecisions.empty() && Uniforms.empty() && Scalars.empty() &&3594           "No decisions should have been taken at this point");3595    // Note: There is no need to invalidate any cost modeling decisions here, as3596    // none were taken so far.3597    InterleaveInfo.invalidateGroupsRequiringScalarEpilogue();3598  }3599 3600  FixedScalableVFPair MaxFactors = computeFeasibleMaxVF(MaxTC, UserVF, true);3601 3602  // Avoid tail folding if the trip count is known to be a multiple of any VF3603  // we choose.3604  std::optional<unsigned> MaxPowerOf2RuntimeVF =3605      MaxFactors.FixedVF.getFixedValue();3606  if (MaxFactors.ScalableVF) {3607    std::optional<unsigned> MaxVScale = getMaxVScale(*TheFunction, TTI);3608    if (MaxVScale && TTI.isVScaleKnownToBeAPowerOfTwo()) {3609      MaxPowerOf2RuntimeVF = std::max<unsigned>(3610          *MaxPowerOf2RuntimeVF,3611          *MaxVScale * MaxFactors.ScalableVF.getKnownMinValue());3612    } else3613      MaxPowerOf2RuntimeVF = std::nullopt; // Stick with tail-folding for now.3614  }3615 3616  auto NoScalarEpilogueNeeded = [this, &UserIC](unsigned MaxVF) {3617    // Return false if the loop is neither a single-latch-exit loop nor an3618    // early-exit loop as tail-folding is not supported in that case.3619    if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch() &&3620        !Legal->hasUncountableEarlyExit())3621      return false;3622    unsigned MaxVFtimesIC = UserIC ? MaxVF * UserIC : MaxVF;3623    ScalarEvolution *SE = PSE.getSE();3624    // Calling getSymbolicMaxBackedgeTakenCount enables support for loops3625    // with uncountable exits. For countable loops, the symbolic maximum must3626    // remain identical to the known back-edge taken count.3627    const SCEV *BackedgeTakenCount = PSE.getSymbolicMaxBackedgeTakenCount();3628    assert((Legal->hasUncountableEarlyExit() ||3629            BackedgeTakenCount == PSE.getBackedgeTakenCount()) &&3630           "Invalid loop count");3631    const SCEV *ExitCount = SE->getAddExpr(3632        BackedgeTakenCount, SE->getOne(BackedgeTakenCount->getType()));3633    const SCEV *Rem = SE->getURemExpr(3634        SE->applyLoopGuards(ExitCount, TheLoop),3635        SE->getConstant(BackedgeTakenCount->getType(), MaxVFtimesIC));3636    return Rem->isZero();3637  };3638 3639  if (MaxPowerOf2RuntimeVF > 0u) {3640    assert((UserVF.isNonZero() || isPowerOf2_32(*MaxPowerOf2RuntimeVF)) &&3641           "MaxFixedVF must be a power of 2");3642    if (NoScalarEpilogueNeeded(*MaxPowerOf2RuntimeVF)) {3643      // Accept MaxFixedVF if we do not have a tail.3644      LLVM_DEBUG(dbgs() << "LV: No tail will remain for any chosen VF.\n");3645      return MaxFactors;3646    }3647  }3648 3649  auto ExpectedTC = getSmallBestKnownTC(PSE, TheLoop);3650  if (ExpectedTC && ExpectedTC->isFixed() &&3651      ExpectedTC->getFixedValue() <=3652          TTI.getMinTripCountTailFoldingThreshold()) {3653    if (MaxPowerOf2RuntimeVF > 0u) {3654      // If we have a low-trip-count, and the fixed-width VF is known to divide3655      // the trip count but the scalable factor does not, use the fixed-width3656      // factor in preference to allow the generation of a non-predicated loop.3657      if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedLowTripLoop &&3658          NoScalarEpilogueNeeded(MaxFactors.FixedVF.getFixedValue())) {3659        LLVM_DEBUG(dbgs() << "LV: Picking a fixed-width so that no tail will "3660                             "remain for any chosen VF.\n");3661        MaxFactors.ScalableVF = ElementCount::getScalable(0);3662        return MaxFactors;3663      }3664    }3665 3666    reportVectorizationFailure(3667        "The trip count is below the minial threshold value.",3668        "loop trip count is too low, avoiding vectorization", "LowTripCount",3669        ORE, TheLoop);3670    return FixedScalableVFPair::getNone();3671  }3672 3673  // If we don't know the precise trip count, or if the trip count that we3674  // found modulo the vectorization factor is not zero, try to fold the tail3675  // by masking.3676  // FIXME: look for a smaller MaxVF that does divide TC rather than masking.3677  bool ContainsScalableVF = MaxFactors.ScalableVF.isNonZero();3678  setTailFoldingStyles(ContainsScalableVF, UserIC);3679  if (foldTailByMasking()) {3680    if (getTailFoldingStyle() == TailFoldingStyle::DataWithEVL) {3681      LLVM_DEBUG(3682          dbgs()3683          << "LV: tail is folded with EVL, forcing unroll factor to be 1. Will "3684             "try to generate VP Intrinsics with scalable vector "3685             "factors only.\n");3686      // Tail folded loop using VP intrinsics restricts the VF to be scalable3687      // for now.3688      // TODO: extend it for fixed vectors, if required.3689      assert(ContainsScalableVF && "Expected scalable vector factor.");3690 3691      MaxFactors.FixedVF = ElementCount::getFixed(1);3692    }3693    return MaxFactors;3694  }3695 3696  // If there was a tail-folding hint/switch, but we can't fold the tail by3697  // masking, fallback to a vectorization with a scalar epilogue.3698  if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) {3699    LLVM_DEBUG(dbgs() << "LV: Cannot fold tail by masking: vectorize with a "3700                         "scalar epilogue instead.\n");3701    ScalarEpilogueStatus = CM_ScalarEpilogueAllowed;3702    return MaxFactors;3703  }3704 3705  if (ScalarEpilogueStatus == CM_ScalarEpilogueNotAllowedUsePredicate) {3706    LLVM_DEBUG(dbgs() << "LV: Can't fold tail by masking: don't vectorize\n");3707    return FixedScalableVFPair::getNone();3708  }3709 3710  if (TC.isZero()) {3711    reportVectorizationFailure(3712        "unable to calculate the loop count due to complex control flow",3713        "UnknownLoopCountComplexCFG", ORE, TheLoop);3714    return FixedScalableVFPair::getNone();3715  }3716 3717  reportVectorizationFailure(3718      "Cannot optimize for size and vectorize at the same time.",3719      "cannot optimize for size and vectorize at the same time. "3720      "Enable vectorization of this loop with '#pragma clang loop "3721      "vectorize(enable)' when compiling with -Os/-Oz",3722      "NoTailLoopWithOptForSize", ORE, TheLoop);3723  return FixedScalableVFPair::getNone();3724}3725 3726bool LoopVectorizationCostModel::shouldConsiderRegPressureForVF(3727    ElementCount VF) {3728  if (ConsiderRegPressure.getNumOccurrences())3729    return ConsiderRegPressure;3730 3731  // TODO: We should eventually consider register pressure for all targets. The3732  // TTI hook is temporary whilst target-specific issues are being fixed.3733  if (TTI.shouldConsiderVectorizationRegPressure())3734    return true;3735 3736  if (!useMaxBandwidth(VF.isScalable()3737                           ? TargetTransformInfo::RGK_ScalableVector3738                           : TargetTransformInfo::RGK_FixedWidthVector))3739    return false;3740  // Only calculate register pressure for VFs enabled by MaxBandwidth.3741  return ElementCount::isKnownGT(3742      VF, VF.isScalable() ? MaxPermissibleVFWithoutMaxBW.ScalableVF3743                          : MaxPermissibleVFWithoutMaxBW.FixedVF);3744}3745 3746bool LoopVectorizationCostModel::useMaxBandwidth(3747    TargetTransformInfo::RegisterKind RegKind) {3748  return MaximizeBandwidth || (MaximizeBandwidth.getNumOccurrences() == 0 &&3749                               (TTI.shouldMaximizeVectorBandwidth(RegKind) ||3750                                (UseWiderVFIfCallVariantsPresent &&3751                                 Legal->hasVectorCallVariants())));3752}3753 3754ElementCount LoopVectorizationCostModel::clampVFByMaxTripCount(3755    ElementCount VF, unsigned MaxTripCount, bool FoldTailByMasking) const {3756  unsigned EstimatedVF = VF.getKnownMinValue();3757  if (VF.isScalable() && TheFunction->hasFnAttribute(Attribute::VScaleRange)) {3758    auto Attr = TheFunction->getFnAttribute(Attribute::VScaleRange);3759    auto Min = Attr.getVScaleRangeMin();3760    EstimatedVF *= Min;3761  }3762 3763  // When a scalar epilogue is required, at least one iteration of the scalar3764  // loop has to execute. Adjust MaxTripCount accordingly to avoid picking a3765  // max VF that results in a dead vector loop.3766  if (MaxTripCount > 0 && requiresScalarEpilogue(true))3767    MaxTripCount -= 1;3768 3769  if (MaxTripCount && MaxTripCount <= EstimatedVF &&3770      (!FoldTailByMasking || isPowerOf2_32(MaxTripCount))) {3771    // If upper bound loop trip count (TC) is known at compile time there is no3772    // point in choosing VF greater than TC (as done in the loop below). Select3773    // maximum power of two which doesn't exceed TC. If VF is3774    // scalable, we only fall back on a fixed VF when the TC is less than or3775    // equal to the known number of lanes.3776    auto ClampedUpperTripCount = llvm::bit_floor(MaxTripCount);3777    LLVM_DEBUG(dbgs() << "LV: Clamping the MaxVF to maximum power of two not "3778                         "exceeding the constant trip count: "3779                      << ClampedUpperTripCount << "\n");3780    return ElementCount::get(ClampedUpperTripCount,3781                             FoldTailByMasking ? VF.isScalable() : false);3782  }3783  return VF;3784}3785 3786ElementCount LoopVectorizationCostModel::getMaximizedVFForTarget(3787    unsigned MaxTripCount, unsigned SmallestType, unsigned WidestType,3788    ElementCount MaxSafeVF, bool FoldTailByMasking) {3789  bool ComputeScalableMaxVF = MaxSafeVF.isScalable();3790  const TypeSize WidestRegister = TTI.getRegisterBitWidth(3791      ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector3792                           : TargetTransformInfo::RGK_FixedWidthVector);3793 3794  // Convenience function to return the minimum of two ElementCounts.3795  auto MinVF = [](const ElementCount &LHS, const ElementCount &RHS) {3796    assert((LHS.isScalable() == RHS.isScalable()) &&3797           "Scalable flags must match");3798    return ElementCount::isKnownLT(LHS, RHS) ? LHS : RHS;3799  };3800 3801  // Ensure MaxVF is a power of 2; the dependence distance bound may not be.3802  // Note that both WidestRegister and WidestType may not be a powers of 2.3803  auto MaxVectorElementCount = ElementCount::get(3804      llvm::bit_floor(WidestRegister.getKnownMinValue() / WidestType),3805      ComputeScalableMaxVF);3806  MaxVectorElementCount = MinVF(MaxVectorElementCount, MaxSafeVF);3807  LLVM_DEBUG(dbgs() << "LV: The Widest register safe to use is: "3808                    << (MaxVectorElementCount * WidestType) << " bits.\n");3809 3810  if (!MaxVectorElementCount) {3811    LLVM_DEBUG(dbgs() << "LV: The target has no "3812                      << (ComputeScalableMaxVF ? "scalable" : "fixed")3813                      << " vector registers.\n");3814    return ElementCount::getFixed(1);3815  }3816 3817  ElementCount MaxVF = clampVFByMaxTripCount(MaxVectorElementCount,3818                                             MaxTripCount, FoldTailByMasking);3819  // If the MaxVF was already clamped, there's no point in trying to pick a3820  // larger one.3821  if (MaxVF != MaxVectorElementCount)3822    return MaxVF;3823 3824  TargetTransformInfo::RegisterKind RegKind =3825      ComputeScalableMaxVF ? TargetTransformInfo::RGK_ScalableVector3826                           : TargetTransformInfo::RGK_FixedWidthVector;3827 3828  if (MaxVF.isScalable())3829    MaxPermissibleVFWithoutMaxBW.ScalableVF = MaxVF;3830  else3831    MaxPermissibleVFWithoutMaxBW.FixedVF = MaxVF;3832 3833  if (useMaxBandwidth(RegKind)) {3834    auto MaxVectorElementCountMaxBW = ElementCount::get(3835        llvm::bit_floor(WidestRegister.getKnownMinValue() / SmallestType),3836        ComputeScalableMaxVF);3837    MaxVF = MinVF(MaxVectorElementCountMaxBW, MaxSafeVF);3838 3839    if (ElementCount MinVF =3840            TTI.getMinimumVF(SmallestType, ComputeScalableMaxVF)) {3841      if (ElementCount::isKnownLT(MaxVF, MinVF)) {3842        LLVM_DEBUG(dbgs() << "LV: Overriding calculated MaxVF(" << MaxVF3843                          << ") with target's minimum: " << MinVF << '\n');3844        MaxVF = MinVF;3845      }3846    }3847 3848    MaxVF = clampVFByMaxTripCount(MaxVF, MaxTripCount, FoldTailByMasking);3849 3850    if (MaxVectorElementCount != MaxVF) {3851      // Invalidate any widening decisions we might have made, in case the loop3852      // requires prediction (decided later), but we have already made some3853      // load/store widening decisions.3854      invalidateCostModelingDecisions();3855    }3856  }3857  return MaxVF;3858}3859 3860bool LoopVectorizationPlanner::isMoreProfitable(const VectorizationFactor &A,3861                                                const VectorizationFactor &B,3862                                                const unsigned MaxTripCount,3863                                                bool HasTail,3864                                                bool IsEpilogue) const {3865  InstructionCost CostA = A.Cost;3866  InstructionCost CostB = B.Cost;3867 3868  // Improve estimate for the vector width if it is scalable.3869  unsigned EstimatedWidthA = A.Width.getKnownMinValue();3870  unsigned EstimatedWidthB = B.Width.getKnownMinValue();3871  if (std::optional<unsigned> VScale = CM.getVScaleForTuning()) {3872    if (A.Width.isScalable())3873      EstimatedWidthA *= *VScale;3874    if (B.Width.isScalable())3875      EstimatedWidthB *= *VScale;3876  }3877 3878  // When optimizing for size choose whichever is smallest, which will be the3879  // one with the smallest cost for the whole loop. On a tie pick the larger3880  // vector width, on the assumption that throughput will be greater.3881  if (CM.CostKind == TTI::TCK_CodeSize)3882    return CostA < CostB ||3883           (CostA == CostB && EstimatedWidthA > EstimatedWidthB);3884 3885  // Assume vscale may be larger than 1 (or the value being tuned for),3886  // so that scalable vectorization is slightly favorable over fixed-width3887  // vectorization.3888  bool PreferScalable = !TTI.preferFixedOverScalableIfEqualCost(IsEpilogue) &&3889                        A.Width.isScalable() && !B.Width.isScalable();3890 3891  auto CmpFn = [PreferScalable](const InstructionCost &LHS,3892                                const InstructionCost &RHS) {3893    return PreferScalable ? LHS <= RHS : LHS < RHS;3894  };3895 3896  // To avoid the need for FP division:3897  //      (CostA / EstimatedWidthA) < (CostB / EstimatedWidthB)3898  // <=>  (CostA * EstimatedWidthB) < (CostB * EstimatedWidthA)3899  if (!MaxTripCount)3900    return CmpFn(CostA * EstimatedWidthB, CostB * EstimatedWidthA);3901 3902  auto GetCostForTC = [MaxTripCount, HasTail](unsigned VF,3903                                              InstructionCost VectorCost,3904                                              InstructionCost ScalarCost) {3905    // If the trip count is a known (possibly small) constant, the trip count3906    // will be rounded up to an integer number of iterations under3907    // FoldTailByMasking. The total cost in that case will be3908    // VecCost*ceil(TripCount/VF). When not folding the tail, the total3909    // cost will be VecCost*floor(TC/VF) + ScalarCost*(TC%VF). There will be3910    // some extra overheads, but for the purpose of comparing the costs of3911    // different VFs we can use this to compare the total loop-body cost3912    // expected after vectorization.3913    if (HasTail)3914      return VectorCost * (MaxTripCount / VF) +3915             ScalarCost * (MaxTripCount % VF);3916    return VectorCost * divideCeil(MaxTripCount, VF);3917  };3918 3919  auto RTCostA = GetCostForTC(EstimatedWidthA, CostA, A.ScalarCost);3920  auto RTCostB = GetCostForTC(EstimatedWidthB, CostB, B.ScalarCost);3921  return CmpFn(RTCostA, RTCostB);3922}3923 3924bool LoopVectorizationPlanner::isMoreProfitable(const VectorizationFactor &A,3925                                                const VectorizationFactor &B,3926                                                bool HasTail,3927                                                bool IsEpilogue) const {3928  const unsigned MaxTripCount = PSE.getSmallConstantMaxTripCount();3929  return LoopVectorizationPlanner::isMoreProfitable(A, B, MaxTripCount, HasTail,3930                                                    IsEpilogue);3931}3932 3933void LoopVectorizationPlanner::emitInvalidCostRemarks(3934    OptimizationRemarkEmitter *ORE) {3935  using RecipeVFPair = std::pair<VPRecipeBase *, ElementCount>;3936  SmallVector<RecipeVFPair> InvalidCosts;3937  for (const auto &Plan : VPlans) {3938    for (ElementCount VF : Plan->vectorFactors()) {3939      // The VPlan-based cost model is designed for computing vector cost.3940      // Querying VPlan-based cost model with a scarlar VF will cause some3941      // errors because we expect the VF is vector for most of the widen3942      // recipes.3943      if (VF.isScalar())3944        continue;3945 3946      VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind,3947                            *CM.PSE.getSE(), OrigLoop);3948      precomputeCosts(*Plan, VF, CostCtx);3949      auto Iter = vp_depth_first_deep(Plan->getVectorLoopRegion()->getEntry());3950      for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) {3951        for (auto &R : *VPBB) {3952          if (!R.cost(VF, CostCtx).isValid())3953            InvalidCosts.emplace_back(&R, VF);3954        }3955      }3956    }3957  }3958  if (InvalidCosts.empty())3959    return;3960 3961  // Emit a report of VFs with invalid costs in the loop.3962 3963  // Group the remarks per recipe, keeping the recipe order from InvalidCosts.3964  DenseMap<VPRecipeBase *, unsigned> Numbering;3965  unsigned I = 0;3966  for (auto &Pair : InvalidCosts)3967    if (Numbering.try_emplace(Pair.first, I).second)3968      ++I;3969 3970  // Sort the list, first on recipe(number) then on VF.3971  sort(InvalidCosts, [&Numbering](RecipeVFPair &A, RecipeVFPair &B) {3972    unsigned NA = Numbering[A.first];3973    unsigned NB = Numbering[B.first];3974    if (NA != NB)3975      return NA < NB;3976    return ElementCount::isKnownLT(A.second, B.second);3977  });3978 3979  // For a list of ordered recipe-VF pairs:3980  //   [(load, VF1), (load, VF2), (store, VF1)]3981  // group the recipes together to emit separate remarks for:3982  //   load  (VF1, VF2)3983  //   store (VF1)3984  auto Tail = ArrayRef<RecipeVFPair>(InvalidCosts);3985  auto Subset = ArrayRef<RecipeVFPair>();3986  do {3987    if (Subset.empty())3988      Subset = Tail.take_front(1);3989 3990    VPRecipeBase *R = Subset.front().first;3991 3992    unsigned Opcode =3993        TypeSwitch<const VPRecipeBase *, unsigned>(R)3994            .Case<VPHeaderPHIRecipe>(3995                [](const auto *R) { return Instruction::PHI; })3996            .Case<VPWidenSelectRecipe>(3997                [](const auto *R) { return Instruction::Select; })3998            .Case<VPWidenStoreRecipe>(3999                [](const auto *R) { return Instruction::Store; })4000            .Case<VPWidenLoadRecipe>(4001                [](const auto *R) { return Instruction::Load; })4002            .Case<VPWidenCallRecipe, VPWidenIntrinsicRecipe>(4003                [](const auto *R) { return Instruction::Call; })4004            .Case<VPInstruction, VPWidenRecipe, VPReplicateRecipe,4005                  VPWidenCastRecipe>(4006                [](const auto *R) { return R->getOpcode(); })4007            .Case<VPInterleaveRecipe>([](const VPInterleaveRecipe *R) {4008              return R->getStoredValues().empty() ? Instruction::Load4009                                                  : Instruction::Store;4010            })4011            .Case<VPReductionRecipe>([](const auto *R) {4012              return RecurrenceDescriptor::getOpcode(R->getRecurrenceKind());4013            });4014 4015    // If the next recipe is different, or if there are no other pairs,4016    // emit a remark for the collated subset. e.g.4017    //   [(load, VF1), (load, VF2))]4018    // to emit:4019    //  remark: invalid costs for 'load' at VF=(VF1, VF2)4020    if (Subset == Tail || Tail[Subset.size()].first != R) {4021      std::string OutString;4022      raw_string_ostream OS(OutString);4023      assert(!Subset.empty() && "Unexpected empty range");4024      OS << "Recipe with invalid costs prevented vectorization at VF=(";4025      for (const auto &Pair : Subset)4026        OS << (Pair.second == Subset.front().second ? "" : ", ") << Pair.second;4027      OS << "):";4028      if (Opcode == Instruction::Call) {4029        StringRef Name = "";4030        if (auto *Int = dyn_cast<VPWidenIntrinsicRecipe>(R)) {4031          Name = Int->getIntrinsicName();4032        } else {4033          auto *WidenCall = dyn_cast<VPWidenCallRecipe>(R);4034          Function *CalledFn =4035              WidenCall ? WidenCall->getCalledScalarFunction()4036                        : cast<Function>(R->getOperand(R->getNumOperands() - 1)4037                                             ->getLiveInIRValue());4038          Name = CalledFn->getName();4039        }4040        OS << " call to " << Name;4041      } else4042        OS << " " << Instruction::getOpcodeName(Opcode);4043      reportVectorizationInfo(OutString, "InvalidCost", ORE, OrigLoop, nullptr,4044                              R->getDebugLoc());4045      Tail = Tail.drop_front(Subset.size());4046      Subset = {};4047    } else4048      // Grow the subset by one element4049      Subset = Tail.take_front(Subset.size() + 1);4050  } while (!Tail.empty());4051}4052 4053/// Check if any recipe of \p Plan will generate a vector value, which will be4054/// assigned a vector register.4055static bool willGenerateVectors(VPlan &Plan, ElementCount VF,4056                                const TargetTransformInfo &TTI) {4057  assert(VF.isVector() && "Checking a scalar VF?");4058  VPTypeAnalysis TypeInfo(Plan);4059  DenseSet<VPRecipeBase *> EphemeralRecipes;4060  collectEphemeralRecipesForVPlan(Plan, EphemeralRecipes);4061  // Set of already visited types.4062  DenseSet<Type *> Visited;4063  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(4064           vp_depth_first_shallow(Plan.getVectorLoopRegion()->getEntry()))) {4065    for (VPRecipeBase &R : *VPBB) {4066      if (EphemeralRecipes.contains(&R))4067        continue;4068      // Continue early if the recipe is considered to not produce a vector4069      // result. Note that this includes VPInstruction where some opcodes may4070      // produce a vector, to preserve existing behavior as VPInstructions model4071      // aspects not directly mapped to existing IR instructions.4072      switch (R.getVPDefID()) {4073      case VPDef::VPDerivedIVSC:4074      case VPDef::VPScalarIVStepsSC:4075      case VPDef::VPReplicateSC:4076      case VPDef::VPInstructionSC:4077      case VPDef::VPCanonicalIVPHISC:4078      case VPDef::VPVectorPointerSC:4079      case VPDef::VPVectorEndPointerSC:4080      case VPDef::VPExpandSCEVSC:4081      case VPDef::VPEVLBasedIVPHISC:4082      case VPDef::VPPredInstPHISC:4083      case VPDef::VPBranchOnMaskSC:4084        continue;4085      case VPDef::VPReductionSC:4086      case VPDef::VPActiveLaneMaskPHISC:4087      case VPDef::VPWidenCallSC:4088      case VPDef::VPWidenCanonicalIVSC:4089      case VPDef::VPWidenCastSC:4090      case VPDef::VPWidenGEPSC:4091      case VPDef::VPWidenIntrinsicSC:4092      case VPDef::VPWidenSC:4093      case VPDef::VPWidenSelectSC:4094      case VPDef::VPBlendSC:4095      case VPDef::VPFirstOrderRecurrencePHISC:4096      case VPDef::VPHistogramSC:4097      case VPDef::VPWidenPHISC:4098      case VPDef::VPWidenIntOrFpInductionSC:4099      case VPDef::VPWidenPointerInductionSC:4100      case VPDef::VPReductionPHISC:4101      case VPDef::VPInterleaveEVLSC:4102      case VPDef::VPInterleaveSC:4103      case VPDef::VPWidenLoadEVLSC:4104      case VPDef::VPWidenLoadSC:4105      case VPDef::VPWidenStoreEVLSC:4106      case VPDef::VPWidenStoreSC:4107        break;4108      default:4109        llvm_unreachable("unhandled recipe");4110      }4111 4112      auto WillGenerateTargetVectors = [&TTI, VF](Type *VectorTy) {4113        unsigned NumLegalParts = TTI.getNumberOfParts(VectorTy);4114        if (!NumLegalParts)4115          return false;4116        if (VF.isScalable()) {4117          // <vscale x 1 x iN> is assumed to be profitable over iN because4118          // scalable registers are a distinct register class from scalar4119          // ones. If we ever find a target which wants to lower scalable4120          // vectors back to scalars, we'll need to update this code to4121          // explicitly ask TTI about the register class uses for each part.4122          return NumLegalParts <= VF.getKnownMinValue();4123        }4124        // Two or more elements that share a register - are vectorized.4125        return NumLegalParts < VF.getFixedValue();4126      };4127 4128      // If no def nor is a store, e.g., branches, continue - no value to check.4129      if (R.getNumDefinedValues() == 0 &&4130          !isa<VPWidenStoreRecipe, VPWidenStoreEVLRecipe, VPInterleaveBase>(&R))4131        continue;4132      // For multi-def recipes, currently only interleaved loads, suffice to4133      // check first def only.4134      // For stores check their stored value; for interleaved stores suffice4135      // the check first stored value only. In all cases this is the second4136      // operand.4137      VPValue *ToCheck =4138          R.getNumDefinedValues() >= 1 ? R.getVPValue(0) : R.getOperand(1);4139      Type *ScalarTy = TypeInfo.inferScalarType(ToCheck);4140      if (!Visited.insert({ScalarTy}).second)4141        continue;4142      Type *WideTy = toVectorizedTy(ScalarTy, VF);4143      if (any_of(getContainedTypes(WideTy), WillGenerateTargetVectors))4144        return true;4145    }4146  }4147 4148  return false;4149}4150 4151static bool hasReplicatorRegion(VPlan &Plan) {4152  return any_of(VPBlockUtils::blocksOnly<VPRegionBlock>(vp_depth_first_shallow(4153                    Plan.getVectorLoopRegion()->getEntry())),4154                [](auto *VPRB) { return VPRB->isReplicator(); });4155}4156 4157#ifndef NDEBUG4158VectorizationFactor LoopVectorizationPlanner::selectVectorizationFactor() {4159  InstructionCost ExpectedCost = CM.expectedCost(ElementCount::getFixed(1));4160  LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ExpectedCost << ".\n");4161  assert(ExpectedCost.isValid() && "Unexpected invalid cost for scalar loop");4162  assert(4163      any_of(VPlans,4164             [](std::unique_ptr<VPlan> &P) { return P->hasScalarVFOnly(); }) &&4165      "Expected Scalar VF to be a candidate");4166 4167  const VectorizationFactor ScalarCost(ElementCount::getFixed(1), ExpectedCost,4168                                       ExpectedCost);4169  VectorizationFactor ChosenFactor = ScalarCost;4170 4171  bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled;4172  if (ForceVectorization &&4173      (VPlans.size() > 1 || !VPlans[0]->hasScalarVFOnly())) {4174    // Ignore scalar width, because the user explicitly wants vectorization.4175    // Initialize cost to max so that VF = 2 is, at least, chosen during cost4176    // evaluation.4177    ChosenFactor.Cost = InstructionCost::getMax();4178  }4179 4180  for (auto &P : VPlans) {4181    ArrayRef<ElementCount> VFs(P->vectorFactors().begin(),4182                               P->vectorFactors().end());4183 4184    SmallVector<VPRegisterUsage, 8> RUs;4185    if (any_of(VFs, [this](ElementCount VF) {4186          return CM.shouldConsiderRegPressureForVF(VF);4187        }))4188      RUs = calculateRegisterUsageForPlan(*P, VFs, TTI, CM.ValuesToIgnore);4189 4190    for (unsigned I = 0; I < VFs.size(); I++) {4191      ElementCount VF = VFs[I];4192      // The cost for scalar VF=1 is already calculated, so ignore it.4193      if (VF.isScalar())4194        continue;4195 4196      /// If the register pressure needs to be considered for VF,4197      /// don't consider the VF as valid if it exceeds the number4198      /// of registers for the target.4199      if (CM.shouldConsiderRegPressureForVF(VF) &&4200          RUs[I].exceedsMaxNumRegs(TTI, ForceTargetNumVectorRegs))4201        continue;4202 4203      InstructionCost C = CM.expectedCost(VF);4204 4205      // Add on other costs that are modelled in VPlan, but not in the legacy4206      // cost model.4207      VPCostContext CostCtx(CM.TTI, *CM.TLI, *P, CM, CM.CostKind,4208                            *CM.PSE.getSE(), OrigLoop);4209      VPRegionBlock *VectorRegion = P->getVectorLoopRegion();4210      assert(VectorRegion && "Expected to have a vector region!");4211      for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(4212               vp_depth_first_shallow(VectorRegion->getEntry()))) {4213        for (VPRecipeBase &R : *VPBB) {4214          auto *VPI = dyn_cast<VPInstruction>(&R);4215          if (!VPI)4216            continue;4217          switch (VPI->getOpcode()) {4218          // Selects are only modelled in the legacy cost model for safe4219          // divisors.4220          case Instruction::Select: {4221            if (auto *WR =4222                    dyn_cast_or_null<VPWidenRecipe>(VPI->getSingleUser())) {4223              switch (WR->getOpcode()) {4224              case Instruction::UDiv:4225              case Instruction::SDiv:4226              case Instruction::URem:4227              case Instruction::SRem:4228                continue;4229              default:4230                break;4231              }4232            }4233            C += VPI->cost(VF, CostCtx);4234            break;4235          }4236          case VPInstruction::ActiveLaneMask: {4237            unsigned Multiplier =4238                cast<ConstantInt>(VPI->getOperand(2)->getLiveInIRValue())4239                    ->getZExtValue();4240            C += VPI->cost(VF * Multiplier, CostCtx);4241            break;4242          }4243          case VPInstruction::ExplicitVectorLength:4244            C += VPI->cost(VF, CostCtx);4245            break;4246          default:4247            break;4248          }4249        }4250      }4251 4252      VectorizationFactor Candidate(VF, C, ScalarCost.ScalarCost);4253      unsigned Width =4254          estimateElementCount(Candidate.Width, CM.getVScaleForTuning());4255      LLVM_DEBUG(dbgs() << "LV: Vector loop of width " << VF4256                        << " costs: " << (Candidate.Cost / Width));4257      if (VF.isScalable())4258        LLVM_DEBUG(dbgs() << " (assuming a minimum vscale of "4259                          << CM.getVScaleForTuning().value_or(1) << ")");4260      LLVM_DEBUG(dbgs() << ".\n");4261 4262      if (!ForceVectorization && !willGenerateVectors(*P, VF, TTI)) {4263        LLVM_DEBUG(4264            dbgs()4265            << "LV: Not considering vector loop of width " << VF4266            << " because it will not generate any vector instructions.\n");4267        continue;4268      }4269 4270      if (CM.OptForSize && !ForceVectorization && hasReplicatorRegion(*P)) {4271        LLVM_DEBUG(4272            dbgs()4273            << "LV: Not considering vector loop of width " << VF4274            << " because it would cause replicated blocks to be generated,"4275            << " which isn't allowed when optimizing for size.\n");4276        continue;4277      }4278 4279      if (isMoreProfitable(Candidate, ChosenFactor, P->hasScalarTail()))4280        ChosenFactor = Candidate;4281    }4282  }4283 4284  if (!EnableCondStoresVectorization && CM.hasPredStores()) {4285    reportVectorizationFailure(4286        "There are conditional stores.",4287        "store that is conditionally executed prevents vectorization",4288        "ConditionalStore", ORE, OrigLoop);4289    ChosenFactor = ScalarCost;4290  }4291 4292  LLVM_DEBUG(if (ForceVectorization && !ChosenFactor.Width.isScalar() &&4293                 !isMoreProfitable(ChosenFactor, ScalarCost,4294                                   !CM.foldTailByMasking())) dbgs()4295             << "LV: Vectorization seems to be not beneficial, "4296             << "but was forced by a user.\n");4297  return ChosenFactor;4298}4299#endif4300 4301bool LoopVectorizationPlanner::isCandidateForEpilogueVectorization(4302    ElementCount VF) const {4303  // Cross iteration phis such as fixed-order recurrences and FMaxNum/FMinNum4304  // reductions need special handling and are currently unsupported.4305  if (any_of(OrigLoop->getHeader()->phis(), [&](PHINode &Phi) {4306        if (!Legal->isReductionVariable(&Phi))4307          return Legal->isFixedOrderRecurrence(&Phi);4308        return RecurrenceDescriptor::isFPMinMaxNumRecurrenceKind(4309            Legal->getRecurrenceDescriptor(&Phi).getRecurrenceKind());4310      }))4311    return false;4312 4313  // Phis with uses outside of the loop require special handling and are4314  // currently unsupported.4315  for (const auto &Entry : Legal->getInductionVars()) {4316    // Look for uses of the value of the induction at the last iteration.4317    Value *PostInc =4318        Entry.first->getIncomingValueForBlock(OrigLoop->getLoopLatch());4319    for (User *U : PostInc->users())4320      if (!OrigLoop->contains(cast<Instruction>(U)))4321        return false;4322    // Look for uses of penultimate value of the induction.4323    for (User *U : Entry.first->users())4324      if (!OrigLoop->contains(cast<Instruction>(U)))4325        return false;4326  }4327 4328  // Epilogue vectorization code has not been auditted to ensure it handles4329  // non-latch exits properly.  It may be fine, but it needs auditted and4330  // tested.4331  // TODO: Add support for loops with an early exit.4332  if (OrigLoop->getExitingBlock() != OrigLoop->getLoopLatch())4333    return false;4334 4335  return true;4336}4337 4338bool LoopVectorizationCostModel::isEpilogueVectorizationProfitable(4339    const ElementCount VF, const unsigned IC) const {4340  // FIXME: We need a much better cost-model to take different parameters such4341  // as register pressure, code size increase and cost of extra branches into4342  // account. For now we apply a very crude heuristic and only consider loops4343  // with vectorization factors larger than a certain value.4344 4345  // Allow the target to opt out entirely.4346  if (!TTI.preferEpilogueVectorization())4347    return false;4348 4349  // We also consider epilogue vectorization unprofitable for targets that don't4350  // consider interleaving beneficial (eg. MVE).4351  if (TTI.getMaxInterleaveFactor(VF) <= 1)4352    return false;4353 4354  unsigned MinVFThreshold = EpilogueVectorizationMinVF.getNumOccurrences() > 04355                                ? EpilogueVectorizationMinVF4356                                : TTI.getEpilogueVectorizationMinVF();4357  return estimateElementCount(VF * IC, VScaleForTuning) >= MinVFThreshold;4358}4359 4360VectorizationFactor LoopVectorizationPlanner::selectEpilogueVectorizationFactor(4361    const ElementCount MainLoopVF, unsigned IC) {4362  VectorizationFactor Result = VectorizationFactor::Disabled();4363  if (!EnableEpilogueVectorization) {4364    LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is disabled.\n");4365    return Result;4366  }4367 4368  if (!CM.isScalarEpilogueAllowed()) {4369    LLVM_DEBUG(dbgs() << "LEV: Unable to vectorize epilogue because no "4370                         "epilogue is allowed.\n");4371    return Result;4372  }4373 4374  // Not really a cost consideration, but check for unsupported cases here to4375  // simplify the logic.4376  if (!isCandidateForEpilogueVectorization(MainLoopVF)) {4377    LLVM_DEBUG(dbgs() << "LEV: Unable to vectorize epilogue because the loop "4378                         "is not a supported candidate.\n");4379    return Result;4380  }4381 4382  if (EpilogueVectorizationForceVF > 1) {4383    LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization factor is forced.\n");4384    ElementCount ForcedEC = ElementCount::getFixed(EpilogueVectorizationForceVF);4385    if (hasPlanWithVF(ForcedEC))4386      return {ForcedEC, 0, 0};4387 4388    LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization forced factor is not "4389                         "viable.\n");4390    return Result;4391  }4392 4393  if (OrigLoop->getHeader()->getParent()->hasOptSize()) {4394    LLVM_DEBUG(4395        dbgs() << "LEV: Epilogue vectorization skipped due to opt for size.\n");4396    return Result;4397  }4398 4399  if (!CM.isEpilogueVectorizationProfitable(MainLoopVF, IC)) {4400    LLVM_DEBUG(dbgs() << "LEV: Epilogue vectorization is not profitable for "4401                         "this loop\n");4402    return Result;4403  }4404 4405  // If MainLoopVF = vscale x 2, and vscale is expected to be 4, then we know4406  // the main loop handles 8 lanes per iteration. We could still benefit from4407  // vectorizing the epilogue loop with VF=4.4408  ElementCount EstimatedRuntimeVF = ElementCount::getFixed(4409      estimateElementCount(MainLoopVF, CM.getVScaleForTuning()));4410 4411  ScalarEvolution &SE = *PSE.getSE();4412  Type *TCType = Legal->getWidestInductionType();4413  const SCEV *RemainingIterations = nullptr;4414  unsigned MaxTripCount = 0;4415  const SCEV *TC =4416      vputils::getSCEVExprForVPValue(getPlanFor(MainLoopVF).getTripCount(), SE);4417  assert(!isa<SCEVCouldNotCompute>(TC) && "Trip count SCEV must be computable");4418  const SCEV *KnownMinTC;4419  bool ScalableTC = match(TC, m_scev_c_Mul(m_SCEV(KnownMinTC), m_SCEVVScale()));4420  bool ScalableRemIter = false;4421  // Use versions of TC and VF in which both are either scalable or fixed.4422  if (ScalableTC == MainLoopVF.isScalable()) {4423    ScalableRemIter = ScalableTC;4424    RemainingIterations =4425        SE.getURemExpr(TC, SE.getElementCount(TCType, MainLoopVF * IC));4426  } else if (ScalableTC) {4427    const SCEV *EstimatedTC = SE.getMulExpr(4428        KnownMinTC,4429        SE.getConstant(TCType, CM.getVScaleForTuning().value_or(1)));4430    RemainingIterations = SE.getURemExpr(4431        EstimatedTC, SE.getElementCount(TCType, MainLoopVF * IC));4432  } else4433    RemainingIterations =4434        SE.getURemExpr(TC, SE.getElementCount(TCType, EstimatedRuntimeVF * IC));4435 4436  // No iterations left to process in the epilogue.4437  if (RemainingIterations->isZero())4438    return Result;4439 4440  if (MainLoopVF.isFixed()) {4441    MaxTripCount = MainLoopVF.getFixedValue() * IC - 1;4442    if (SE.isKnownPredicate(CmpInst::ICMP_ULT, RemainingIterations,4443                            SE.getConstant(TCType, MaxTripCount))) {4444      MaxTripCount = SE.getUnsignedRangeMax(RemainingIterations).getZExtValue();4445    }4446    LLVM_DEBUG(dbgs() << "LEV: Maximum Trip Count for Epilogue: "4447                      << MaxTripCount << "\n");4448  }4449 4450  auto SkipVF = [&](const SCEV *VF, const SCEV *RemIter) -> bool {4451    return SE.isKnownPredicate(CmpInst::ICMP_UGT, VF, RemIter);4452  };4453  for (auto &NextVF : ProfitableVFs) {4454    // Skip candidate VFs without a corresponding VPlan.4455    if (!hasPlanWithVF(NextVF.Width))4456      continue;4457 4458    // Skip candidate VFs with widths >= the (estimated) runtime VF (scalable4459    // vectors) or > the VF of the main loop (fixed vectors).4460    if ((!NextVF.Width.isScalable() && MainLoopVF.isScalable() &&4461         ElementCount::isKnownGE(NextVF.Width, EstimatedRuntimeVF)) ||4462        (NextVF.Width.isScalable() &&4463         ElementCount::isKnownGE(NextVF.Width, MainLoopVF)) ||4464        (!NextVF.Width.isScalable() && !MainLoopVF.isScalable() &&4465         ElementCount::isKnownGT(NextVF.Width, MainLoopVF)))4466      continue;4467 4468    // If NextVF is greater than the number of remaining iterations, the4469    // epilogue loop would be dead. Skip such factors.4470    // TODO: We should also consider comparing against a scalable4471    // RemainingIterations when SCEV be able to evaluate non-canonical4472    // vscale-based expressions.4473    if (!ScalableRemIter) {4474      // Handle the case where NextVF and RemainingIterations are in different4475      // numerical spaces.4476      ElementCount EC = NextVF.Width;4477      if (NextVF.Width.isScalable())4478        EC = ElementCount::getFixed(4479            estimateElementCount(NextVF.Width, CM.getVScaleForTuning()));4480      if (SkipVF(SE.getElementCount(TCType, EC), RemainingIterations))4481        continue;4482    }4483 4484    if (Result.Width.isScalar() ||4485        isMoreProfitable(NextVF, Result, MaxTripCount, !CM.foldTailByMasking(),4486                         /*IsEpilogue*/ true))4487      Result = NextVF;4488  }4489 4490  if (Result != VectorizationFactor::Disabled())4491    LLVM_DEBUG(dbgs() << "LEV: Vectorizing epilogue loop with VF = "4492                      << Result.Width << "\n");4493  return Result;4494}4495 4496std::pair<unsigned, unsigned>4497LoopVectorizationCostModel::getSmallestAndWidestTypes() {4498  unsigned MinWidth = -1U;4499  unsigned MaxWidth = 8;4500  const DataLayout &DL = TheFunction->getDataLayout();4501  // For in-loop reductions, no element types are added to ElementTypesInLoop4502  // if there are no loads/stores in the loop. In this case, check through the4503  // reduction variables to determine the maximum width.4504  if (ElementTypesInLoop.empty() && !Legal->getReductionVars().empty()) {4505    for (const auto &PhiDescriptorPair : Legal->getReductionVars()) {4506      const RecurrenceDescriptor &RdxDesc = PhiDescriptorPair.second;4507      // When finding the min width used by the recurrence we need to account4508      // for casts on the input operands of the recurrence.4509      MinWidth = std::min(4510          MinWidth,4511          std::min(RdxDesc.getMinWidthCastToRecurrenceTypeInBits(),4512                   RdxDesc.getRecurrenceType()->getScalarSizeInBits()));4513      MaxWidth = std::max(MaxWidth,4514                          RdxDesc.getRecurrenceType()->getScalarSizeInBits());4515    }4516  } else {4517    for (Type *T : ElementTypesInLoop) {4518      MinWidth = std::min<unsigned>(4519          MinWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedValue());4520      MaxWidth = std::max<unsigned>(4521          MaxWidth, DL.getTypeSizeInBits(T->getScalarType()).getFixedValue());4522    }4523  }4524  return {MinWidth, MaxWidth};4525}4526 4527void LoopVectorizationCostModel::collectElementTypesForWidening() {4528  ElementTypesInLoop.clear();4529  // For each block.4530  for (BasicBlock *BB : TheLoop->blocks()) {4531    // For each instruction in the loop.4532    for (Instruction &I : BB->instructionsWithoutDebug()) {4533      Type *T = I.getType();4534 4535      // Skip ignored values.4536      if (ValuesToIgnore.count(&I))4537        continue;4538 4539      // Only examine Loads, Stores and PHINodes.4540      if (!isa<LoadInst>(I) && !isa<StoreInst>(I) && !isa<PHINode>(I))4541        continue;4542 4543      // Examine PHI nodes that are reduction variables. Update the type to4544      // account for the recurrence type.4545      if (auto *PN = dyn_cast<PHINode>(&I)) {4546        if (!Legal->isReductionVariable(PN))4547          continue;4548        const RecurrenceDescriptor &RdxDesc =4549            Legal->getRecurrenceDescriptor(PN);4550        if (PreferInLoopReductions || useOrderedReductions(RdxDesc) ||4551            TTI.preferInLoopReduction(RdxDesc.getRecurrenceKind(),4552                                      RdxDesc.getRecurrenceType()))4553          continue;4554        T = RdxDesc.getRecurrenceType();4555      }4556 4557      // Examine the stored values.4558      if (auto *ST = dyn_cast<StoreInst>(&I))4559        T = ST->getValueOperand()->getType();4560 4561      assert(T->isSized() &&4562             "Expected the load/store/recurrence type to be sized");4563 4564      ElementTypesInLoop.insert(T);4565    }4566  }4567}4568 4569unsigned4570LoopVectorizationPlanner::selectInterleaveCount(VPlan &Plan, ElementCount VF,4571                                                InstructionCost LoopCost) {4572  // -- The interleave heuristics --4573  // We interleave the loop in order to expose ILP and reduce the loop overhead.4574  // There are many micro-architectural considerations that we can't predict4575  // at this level. For example, frontend pressure (on decode or fetch) due to4576  // code size, or the number and capabilities of the execution ports.4577  //4578  // We use the following heuristics to select the interleave count:4579  // 1. If the code has reductions, then we interleave to break the cross4580  // iteration dependency.4581  // 2. If the loop is really small, then we interleave to reduce the loop4582  // overhead.4583  // 3. We don't interleave if we think that we will spill registers to memory4584  // due to the increased register pressure.4585 4586  // Only interleave tail-folded loops if wide lane masks are requested, as the4587  // overhead of multiple instructions to calculate the predicate is likely4588  // not beneficial. If a scalar epilogue is not allowed for any other reason,4589  // do not interleave.4590  if (!CM.isScalarEpilogueAllowed() &&4591      !(CM.preferPredicatedLoop() && CM.useWideActiveLaneMask()))4592    return 1;4593 4594  if (any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(),4595             IsaPred<VPEVLBasedIVPHIRecipe>)) {4596    LLVM_DEBUG(dbgs() << "LV: Preference for VP intrinsics indicated. "4597                         "Unroll factor forced to be 1.\n");4598    return 1;4599  }4600 4601  // We used the distance for the interleave count.4602  if (!Legal->isSafeForAnyVectorWidth())4603    return 1;4604 4605  // We don't attempt to perform interleaving for loops with uncountable early4606  // exits because the VPInstruction::AnyOf code cannot currently handle4607  // multiple parts.4608  if (Plan.hasEarlyExit())4609    return 1;4610 4611  const bool HasReductions =4612      any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(),4613             IsaPred<VPReductionPHIRecipe>);4614 4615  // If we did not calculate the cost for VF (because the user selected the VF)4616  // then we calculate the cost of VF here.4617  if (LoopCost == 0) {4618    if (VF.isScalar())4619      LoopCost = CM.expectedCost(VF);4620    else4621      LoopCost = cost(Plan, VF);4622    assert(LoopCost.isValid() && "Expected to have chosen a VF with valid cost");4623 4624    // Loop body is free and there is no need for interleaving.4625    if (LoopCost == 0)4626      return 1;4627  }4628 4629  VPRegisterUsage R =4630      calculateRegisterUsageForPlan(Plan, {VF}, TTI, CM.ValuesToIgnore)[0];4631  // We divide by these constants so assume that we have at least one4632  // instruction that uses at least one register.4633  for (auto &Pair : R.MaxLocalUsers) {4634    Pair.second = std::max(Pair.second, 1U);4635  }4636 4637  // We calculate the interleave count using the following formula.4638  // Subtract the number of loop invariants from the number of available4639  // registers. These registers are used by all of the interleaved instances.4640  // Next, divide the remaining registers by the number of registers that is4641  // required by the loop, in order to estimate how many parallel instances4642  // fit without causing spills. All of this is rounded down if necessary to be4643  // a power of two. We want power of two interleave count to simplify any4644  // addressing operations or alignment considerations.4645  // We also want power of two interleave counts to ensure that the induction4646  // variable of the vector loop wraps to zero, when tail is folded by masking;4647  // this currently happens when OptForSize, in which case IC is set to 1 above.4648  unsigned IC = UINT_MAX;4649 4650  for (const auto &Pair : R.MaxLocalUsers) {4651    unsigned TargetNumRegisters = TTI.getNumberOfRegisters(Pair.first);4652    LLVM_DEBUG(dbgs() << "LV: The target has " << TargetNumRegisters4653                      << " registers of "4654                      << TTI.getRegisterClassName(Pair.first)4655                      << " register class\n");4656    if (VF.isScalar()) {4657      if (ForceTargetNumScalarRegs.getNumOccurrences() > 0)4658        TargetNumRegisters = ForceTargetNumScalarRegs;4659    } else {4660      if (ForceTargetNumVectorRegs.getNumOccurrences() > 0)4661        TargetNumRegisters = ForceTargetNumVectorRegs;4662    }4663    unsigned MaxLocalUsers = Pair.second;4664    unsigned LoopInvariantRegs = 0;4665    if (R.LoopInvariantRegs.contains(Pair.first))4666      LoopInvariantRegs = R.LoopInvariantRegs[Pair.first];4667 4668    unsigned TmpIC = llvm::bit_floor((TargetNumRegisters - LoopInvariantRegs) /4669                                     MaxLocalUsers);4670    // Don't count the induction variable as interleaved.4671    if (EnableIndVarRegisterHeur) {4672      TmpIC = llvm::bit_floor((TargetNumRegisters - LoopInvariantRegs - 1) /4673                              std::max(1U, (MaxLocalUsers - 1)));4674    }4675 4676    IC = std::min(IC, TmpIC);4677  }4678 4679  // Clamp the interleave ranges to reasonable counts.4680  unsigned MaxInterleaveCount = TTI.getMaxInterleaveFactor(VF);4681 4682  // Check if the user has overridden the max.4683  if (VF.isScalar()) {4684    if (ForceTargetMaxScalarInterleaveFactor.getNumOccurrences() > 0)4685      MaxInterleaveCount = ForceTargetMaxScalarInterleaveFactor;4686  } else {4687    if (ForceTargetMaxVectorInterleaveFactor.getNumOccurrences() > 0)4688      MaxInterleaveCount = ForceTargetMaxVectorInterleaveFactor;4689  }4690 4691  // Try to get the exact trip count, or an estimate based on profiling data or4692  // ConstantMax from PSE, failing that.4693  auto BestKnownTC = getSmallBestKnownTC(PSE, OrigLoop);4694 4695  // For fixed length VFs treat a scalable trip count as unknown.4696  if (BestKnownTC && (BestKnownTC->isFixed() || VF.isScalable())) {4697    // Re-evaluate trip counts and VFs to be in the same numerical space.4698    unsigned AvailableTC =4699        estimateElementCount(*BestKnownTC, CM.getVScaleForTuning());4700    unsigned EstimatedVF = estimateElementCount(VF, CM.getVScaleForTuning());4701 4702    // At least one iteration must be scalar when this constraint holds. So the4703    // maximum available iterations for interleaving is one less.4704    if (CM.requiresScalarEpilogue(VF.isVector()))4705      --AvailableTC;4706 4707    unsigned InterleaveCountLB = bit_floor(std::max(4708        1u, std::min(AvailableTC / (EstimatedVF * 2), MaxInterleaveCount)));4709 4710    if (getSmallConstantTripCount(PSE.getSE(), OrigLoop).isNonZero()) {4711      // If the best known trip count is exact, we select between two4712      // prospective ICs, where4713      //4714      // 1) the aggressive IC is capped by the trip count divided by VF4715      // 2) the conservative IC is capped by the trip count divided by (VF * 2)4716      //4717      // The final IC is selected in a way that the epilogue loop trip count is4718      // minimized while maximizing the IC itself, so that we either run the4719      // vector loop at least once if it generates a small epilogue loop, or4720      // else we run the vector loop at least twice.4721 4722      unsigned InterleaveCountUB = bit_floor(std::max(4723          1u, std::min(AvailableTC / EstimatedVF, MaxInterleaveCount)));4724      MaxInterleaveCount = InterleaveCountLB;4725 4726      if (InterleaveCountUB != InterleaveCountLB) {4727        unsigned TailTripCountUB =4728            (AvailableTC % (EstimatedVF * InterleaveCountUB));4729        unsigned TailTripCountLB =4730            (AvailableTC % (EstimatedVF * InterleaveCountLB));4731        // If both produce same scalar tail, maximize the IC to do the same work4732        // in fewer vector loop iterations4733        if (TailTripCountUB == TailTripCountLB)4734          MaxInterleaveCount = InterleaveCountUB;4735      }4736    } else {4737      // If trip count is an estimated compile time constant, limit the4738      // IC to be capped by the trip count divided by VF * 2, such that the4739      // vector loop runs at least twice to make interleaving seem profitable4740      // when there is an epilogue loop present. Since exact Trip count is not4741      // known we choose to be conservative in our IC estimate.4742      MaxInterleaveCount = InterleaveCountLB;4743    }4744  }4745 4746  assert(MaxInterleaveCount > 0 &&4747         "Maximum interleave count must be greater than 0");4748 4749  // Clamp the calculated IC to be between the 1 and the max interleave count4750  // that the target and trip count allows.4751  if (IC > MaxInterleaveCount)4752    IC = MaxInterleaveCount;4753  else4754    // Make sure IC is greater than 0.4755    IC = std::max(1u, IC);4756 4757  assert(IC > 0 && "Interleave count must be greater than 0.");4758 4759  // Interleave if we vectorized this loop and there is a reduction that could4760  // benefit from interleaving.4761  if (VF.isVector() && HasReductions) {4762    LLVM_DEBUG(dbgs() << "LV: Interleaving because of reductions.\n");4763    return IC;4764  }4765 4766  // For any scalar loop that either requires runtime checks or predication we4767  // are better off leaving this to the unroller. Note that if we've already4768  // vectorized the loop we will have done the runtime check and so interleaving4769  // won't require further checks.4770  bool ScalarInterleavingRequiresPredication =4771      (VF.isScalar() && any_of(OrigLoop->blocks(), [this](BasicBlock *BB) {4772         return Legal->blockNeedsPredication(BB);4773       }));4774  bool ScalarInterleavingRequiresRuntimePointerCheck =4775      (VF.isScalar() && Legal->getRuntimePointerChecking()->Need);4776 4777  // We want to interleave small loops in order to reduce the loop overhead and4778  // potentially expose ILP opportunities.4779  LLVM_DEBUG(dbgs() << "LV: Loop cost is " << LoopCost << '\n'4780                    << "LV: IC is " << IC << '\n'4781                    << "LV: VF is " << VF << '\n');4782  const bool AggressivelyInterleaveReductions =4783      TTI.enableAggressiveInterleaving(HasReductions);4784  if (!ScalarInterleavingRequiresRuntimePointerCheck &&4785      !ScalarInterleavingRequiresPredication && LoopCost < SmallLoopCost) {4786    // We assume that the cost overhead is 1 and we use the cost model4787    // to estimate the cost of the loop and interleave until the cost of the4788    // loop overhead is about 5% of the cost of the loop.4789    unsigned SmallIC = std::min(IC, (unsigned)llvm::bit_floor<uint64_t>(4790                                        SmallLoopCost / LoopCost.getValue()));4791 4792    // Interleave until store/load ports (estimated by max interleave count) are4793    // saturated.4794    unsigned NumStores = 0;4795    unsigned NumLoads = 0;4796    for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(4797             vp_depth_first_deep(Plan.getVectorLoopRegion()->getEntry()))) {4798      for (VPRecipeBase &R : *VPBB) {4799        if (isa<VPWidenLoadRecipe, VPWidenLoadEVLRecipe>(&R)) {4800          NumLoads++;4801          continue;4802        }4803        if (isa<VPWidenStoreRecipe, VPWidenStoreEVLRecipe>(&R)) {4804          NumStores++;4805          continue;4806        }4807 4808        if (auto *InterleaveR = dyn_cast<VPInterleaveRecipe>(&R)) {4809          if (unsigned StoreOps = InterleaveR->getNumStoreOperands())4810            NumStores += StoreOps;4811          else4812            NumLoads += InterleaveR->getNumDefinedValues();4813          continue;4814        }4815        if (auto *RepR = dyn_cast<VPReplicateRecipe>(&R)) {4816          NumLoads += isa<LoadInst>(RepR->getUnderlyingInstr());4817          NumStores += isa<StoreInst>(RepR->getUnderlyingInstr());4818          continue;4819        }4820        if (isa<VPHistogramRecipe>(&R)) {4821          NumLoads++;4822          NumStores++;4823          continue;4824        }4825      }4826    }4827    unsigned StoresIC = IC / (NumStores ? NumStores : 1);4828    unsigned LoadsIC = IC / (NumLoads ? NumLoads : 1);4829 4830    // There is little point in interleaving for reductions containing selects4831    // and compares when VF=1 since it may just create more overhead than it's4832    // worth for loops with small trip counts. This is because we still have to4833    // do the final reduction after the loop.4834    bool HasSelectCmpReductions =4835        HasReductions &&4836        any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(),4837               [](VPRecipeBase &R) {4838                 auto *RedR = dyn_cast<VPReductionPHIRecipe>(&R);4839                 return RedR && (RecurrenceDescriptor::isAnyOfRecurrenceKind(4840                                     RedR->getRecurrenceKind()) ||4841                                 RecurrenceDescriptor::isFindIVRecurrenceKind(4842                                     RedR->getRecurrenceKind()));4843               });4844    if (HasSelectCmpReductions) {4845      LLVM_DEBUG(dbgs() << "LV: Not interleaving select-cmp reductions.\n");4846      return 1;4847    }4848 4849    // If we have a scalar reduction (vector reductions are already dealt with4850    // by this point), we can increase the critical path length if the loop4851    // we're interleaving is inside another loop. For tree-wise reductions4852    // set the limit to 2, and for ordered reductions it's best to disable4853    // interleaving entirely.4854    if (HasReductions && OrigLoop->getLoopDepth() > 1) {4855      bool HasOrderedReductions =4856          any_of(Plan.getVectorLoopRegion()->getEntryBasicBlock()->phis(),4857                 [](VPRecipeBase &R) {4858                   auto *RedR = dyn_cast<VPReductionPHIRecipe>(&R);4859 4860                   return RedR && RedR->isOrdered();4861                 });4862      if (HasOrderedReductions) {4863        LLVM_DEBUG(4864            dbgs() << "LV: Not interleaving scalar ordered reductions.\n");4865        return 1;4866      }4867 4868      unsigned F = MaxNestedScalarReductionIC;4869      SmallIC = std::min(SmallIC, F);4870      StoresIC = std::min(StoresIC, F);4871      LoadsIC = std::min(LoadsIC, F);4872    }4873 4874    if (EnableLoadStoreRuntimeInterleave &&4875        std::max(StoresIC, LoadsIC) > SmallIC) {4876      LLVM_DEBUG(4877          dbgs() << "LV: Interleaving to saturate store or load ports.\n");4878      return std::max(StoresIC, LoadsIC);4879    }4880 4881    // If there are scalar reductions and TTI has enabled aggressive4882    // interleaving for reductions, we will interleave to expose ILP.4883    if (VF.isScalar() && AggressivelyInterleaveReductions) {4884      LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");4885      // Interleave no less than SmallIC but not as aggressive as the normal IC4886      // to satisfy the rare situation when resources are too limited.4887      return std::max(IC / 2, SmallIC);4888    }4889 4890    LLVM_DEBUG(dbgs() << "LV: Interleaving to reduce branch cost.\n");4891    return SmallIC;4892  }4893 4894  // Interleave if this is a large loop (small loops are already dealt with by4895  // this point) that could benefit from interleaving.4896  if (AggressivelyInterleaveReductions) {4897    LLVM_DEBUG(dbgs() << "LV: Interleaving to expose ILP.\n");4898    return IC;4899  }4900 4901  LLVM_DEBUG(dbgs() << "LV: Not Interleaving.\n");4902  return 1;4903}4904 4905bool LoopVectorizationCostModel::useEmulatedMaskMemRefHack(Instruction *I,4906                                                           ElementCount VF) {4907  // TODO: Cost model for emulated masked load/store is completely4908  // broken. This hack guides the cost model to use an artificially4909  // high enough value to practically disable vectorization with such4910  // operations, except where previously deployed legality hack allowed4911  // using very low cost values. This is to avoid regressions coming simply4912  // from moving "masked load/store" check from legality to cost model.4913  // Masked Load/Gather emulation was previously never allowed.4914  // Limited number of Masked Store/Scatter emulation was allowed.4915  assert((isPredicatedInst(I)) &&4916         "Expecting a scalar emulated instruction");4917  return isa<LoadInst>(I) ||4918         (isa<StoreInst>(I) &&4919          NumPredStores > NumberOfStoresToPredicate);4920}4921 4922void LoopVectorizationCostModel::collectInstsToScalarize(ElementCount VF) {4923  assert(VF.isVector() && "Expected VF >= 2");4924 4925  // If we've already collected the instructions to scalarize or the predicated4926  // BBs after vectorization, there's nothing to do. Collection may already have4927  // occurred if we have a user-selected VF and are now computing the expected4928  // cost for interleaving.4929  if (InstsToScalarize.contains(VF) ||4930      PredicatedBBsAfterVectorization.contains(VF))4931    return;4932 4933  // Initialize a mapping for VF in InstsToScalalarize. If we find that it's4934  // not profitable to scalarize any instructions, the presence of VF in the4935  // map will indicate that we've analyzed it already.4936  ScalarCostsTy &ScalarCostsVF = InstsToScalarize[VF];4937 4938  // Find all the instructions that are scalar with predication in the loop and4939  // determine if it would be better to not if-convert the blocks they are in.4940  // If so, we also record the instructions to scalarize.4941  for (BasicBlock *BB : TheLoop->blocks()) {4942    if (!blockNeedsPredicationForAnyReason(BB))4943      continue;4944    for (Instruction &I : *BB)4945      if (isScalarWithPredication(&I, VF)) {4946        ScalarCostsTy ScalarCosts;4947        // Do not apply discount logic for:4948        // 1. Scalars after vectorization, as there will only be a single copy4949        // of the instruction.4950        // 2. Scalable VF, as that would lead to invalid scalarization costs.4951        // 3. Emulated masked memrefs, if a hacked cost is needed.4952        if (!isScalarAfterVectorization(&I, VF) && !VF.isScalable() &&4953            !useEmulatedMaskMemRefHack(&I, VF) &&4954            computePredInstDiscount(&I, ScalarCosts, VF) >= 0) {4955          for (const auto &[I, IC] : ScalarCosts)4956            ScalarCostsVF.insert({I, IC});4957          // Check if we decided to scalarize a call. If so, update the widening4958          // decision of the call to CM_Scalarize with the computed scalar cost.4959          for (const auto &[I, Cost] : ScalarCosts) {4960            auto *CI = dyn_cast<CallInst>(I);4961            if (!CI || !CallWideningDecisions.contains({CI, VF}))4962              continue;4963            CallWideningDecisions[{CI, VF}].Kind = CM_Scalarize;4964            CallWideningDecisions[{CI, VF}].Cost = Cost;4965          }4966        }4967        // Remember that BB will remain after vectorization.4968        PredicatedBBsAfterVectorization[VF].insert(BB);4969        for (auto *Pred : predecessors(BB)) {4970          if (Pred->getSingleSuccessor() == BB)4971            PredicatedBBsAfterVectorization[VF].insert(Pred);4972        }4973      }4974  }4975}4976 4977InstructionCost LoopVectorizationCostModel::computePredInstDiscount(4978    Instruction *PredInst, ScalarCostsTy &ScalarCosts, ElementCount VF) {4979  assert(!isUniformAfterVectorization(PredInst, VF) &&4980         "Instruction marked uniform-after-vectorization will be predicated");4981 4982  // Initialize the discount to zero, meaning that the scalar version and the4983  // vector version cost the same.4984  InstructionCost Discount = 0;4985 4986  // Holds instructions to analyze. The instructions we visit are mapped in4987  // ScalarCosts. Those instructions are the ones that would be scalarized if4988  // we find that the scalar version costs less.4989  SmallVector<Instruction *, 8> Worklist;4990 4991  // Returns true if the given instruction can be scalarized.4992  auto CanBeScalarized = [&](Instruction *I) -> bool {4993    // We only attempt to scalarize instructions forming a single-use chain4994    // from the original predicated block that would otherwise be vectorized.4995    // Although not strictly necessary, we give up on instructions we know will4996    // already be scalar to avoid traversing chains that are unlikely to be4997    // beneficial.4998    if (!I->hasOneUse() || PredInst->getParent() != I->getParent() ||4999        isScalarAfterVectorization(I, VF))5000      return false;5001 5002    // If the instruction is scalar with predication, it will be analyzed5003    // separately. We ignore it within the context of PredInst.5004    if (isScalarWithPredication(I, VF))5005      return false;5006 5007    // If any of the instruction's operands are uniform after vectorization,5008    // the instruction cannot be scalarized. This prevents, for example, a5009    // masked load from being scalarized.5010    //5011    // We assume we will only emit a value for lane zero of an instruction5012    // marked uniform after vectorization, rather than VF identical values.5013    // Thus, if we scalarize an instruction that uses a uniform, we would5014    // create uses of values corresponding to the lanes we aren't emitting code5015    // for. This behavior can be changed by allowing getScalarValue to clone5016    // the lane zero values for uniforms rather than asserting.5017    for (Use &U : I->operands())5018      if (auto *J = dyn_cast<Instruction>(U.get()))5019        if (isUniformAfterVectorization(J, VF))5020          return false;5021 5022    // Otherwise, we can scalarize the instruction.5023    return true;5024  };5025 5026  // Compute the expected cost discount from scalarizing the entire expression5027  // feeding the predicated instruction. We currently only consider expressions5028  // that are single-use instruction chains.5029  Worklist.push_back(PredInst);5030  while (!Worklist.empty()) {5031    Instruction *I = Worklist.pop_back_val();5032 5033    // If we've already analyzed the instruction, there's nothing to do.5034    if (ScalarCosts.contains(I))5035      continue;5036 5037    // Cannot scalarize fixed-order recurrence phis at the moment.5038    if (isa<PHINode>(I) && Legal->isFixedOrderRecurrence(cast<PHINode>(I)))5039      continue;5040 5041    // Compute the cost of the vector instruction. Note that this cost already5042    // includes the scalarization overhead of the predicated instruction.5043    InstructionCost VectorCost = getInstructionCost(I, VF);5044 5045    // Compute the cost of the scalarized instruction. This cost is the cost of5046    // the instruction as if it wasn't if-converted and instead remained in the5047    // predicated block. We will scale this cost by block probability after5048    // computing the scalarization overhead.5049    InstructionCost ScalarCost =5050        VF.getFixedValue() * getInstructionCost(I, ElementCount::getFixed(1));5051 5052    // Compute the scalarization overhead of needed insertelement instructions5053    // and phi nodes.5054    if (isScalarWithPredication(I, VF) && !I->getType()->isVoidTy()) {5055      Type *WideTy = toVectorizedTy(I->getType(), VF);5056      for (Type *VectorTy : getContainedTypes(WideTy)) {5057        ScalarCost += TTI.getScalarizationOverhead(5058            cast<VectorType>(VectorTy), APInt::getAllOnes(VF.getFixedValue()),5059            /*Insert=*/true,5060            /*Extract=*/false, CostKind);5061      }5062      ScalarCost +=5063          VF.getFixedValue() * TTI.getCFInstrCost(Instruction::PHI, CostKind);5064    }5065 5066    // Compute the scalarization overhead of needed extractelement5067    // instructions. For each of the instruction's operands, if the operand can5068    // be scalarized, add it to the worklist; otherwise, account for the5069    // overhead.5070    for (Use &U : I->operands())5071      if (auto *J = dyn_cast<Instruction>(U.get())) {5072        assert(canVectorizeTy(J->getType()) &&5073               "Instruction has non-scalar type");5074        if (CanBeScalarized(J))5075          Worklist.push_back(J);5076        else if (needsExtract(J, VF)) {5077          Type *WideTy = toVectorizedTy(J->getType(), VF);5078          for (Type *VectorTy : getContainedTypes(WideTy)) {5079            ScalarCost += TTI.getScalarizationOverhead(5080                cast<VectorType>(VectorTy),5081                APInt::getAllOnes(VF.getFixedValue()), /*Insert*/ false,5082                /*Extract*/ true, CostKind);5083          }5084        }5085      }5086 5087    // Scale the total scalar cost by block probability.5088    ScalarCost /= getPredBlockCostDivisor(CostKind, I->getParent());5089 5090    // Compute the discount. A non-negative discount means the vector version5091    // of the instruction costs more, and scalarizing would be beneficial.5092    Discount += VectorCost - ScalarCost;5093    ScalarCosts[I] = ScalarCost;5094  }5095 5096  return Discount;5097}5098 5099InstructionCost LoopVectorizationCostModel::expectedCost(ElementCount VF) {5100  InstructionCost Cost;5101 5102  // If the vector loop gets executed exactly once with the given VF, ignore the5103  // costs of comparison and induction instructions, as they'll get simplified5104  // away.5105  SmallPtrSet<Instruction *, 2> ValuesToIgnoreForVF;5106  auto TC = getSmallConstantTripCount(PSE.getSE(), TheLoop);5107  if (TC == VF && !foldTailByMasking())5108    addFullyUnrolledInstructionsToIgnore(TheLoop, Legal->getInductionVars(),5109                                         ValuesToIgnoreForVF);5110 5111  // For each block.5112  for (BasicBlock *BB : TheLoop->blocks()) {5113    InstructionCost BlockCost;5114 5115    // For each instruction in the old loop.5116    for (Instruction &I : BB->instructionsWithoutDebug()) {5117      // Skip ignored values.5118      if (ValuesToIgnore.count(&I) || ValuesToIgnoreForVF.count(&I) ||5119          (VF.isVector() && VecValuesToIgnore.count(&I)))5120        continue;5121 5122      InstructionCost C = getInstructionCost(&I, VF);5123 5124      // Check if we should override the cost.5125      if (C.isValid() && ForceTargetInstructionCost.getNumOccurrences() > 0)5126        C = InstructionCost(ForceTargetInstructionCost);5127 5128      BlockCost += C;5129      LLVM_DEBUG(dbgs() << "LV: Found an estimated cost of " << C << " for VF "5130                        << VF << " For instruction: " << I << '\n');5131    }5132 5133    // If we are vectorizing a predicated block, it will have been5134    // if-converted. This means that the block's instructions (aside from5135    // stores and instructions that may divide by zero) will now be5136    // unconditionally executed. For the scalar case, we may not always execute5137    // the predicated block, if it is an if-else block. Thus, scale the block's5138    // cost by the probability of executing it.5139    // getPredBlockCostDivisor will return 1 for blocks that are only predicated5140    // by the header mask when folding the tail.5141    if (VF.isScalar())5142      BlockCost /= getPredBlockCostDivisor(CostKind, BB);5143 5144    Cost += BlockCost;5145  }5146 5147  return Cost;5148}5149 5150/// Gets Address Access SCEV after verifying that the access pattern5151/// is loop invariant except the induction variable dependence.5152///5153/// This SCEV can be sent to the Target in order to estimate the address5154/// calculation cost.5155static const SCEV *getAddressAccessSCEV(5156              Value *Ptr,5157              LoopVectorizationLegality *Legal,5158              PredicatedScalarEvolution &PSE,5159              const Loop *TheLoop) {5160 5161  auto *Gep = dyn_cast<GetElementPtrInst>(Ptr);5162  if (!Gep)5163    return nullptr;5164 5165  // We are looking for a gep with all loop invariant indices except for one5166  // which should be an induction variable.5167  auto *SE = PSE.getSE();5168  unsigned NumOperands = Gep->getNumOperands();5169  for (unsigned Idx = 1; Idx < NumOperands; ++Idx) {5170    Value *Opd = Gep->getOperand(Idx);5171    if (!SE->isLoopInvariant(SE->getSCEV(Opd), TheLoop) &&5172        !Legal->isInductionVariable(Opd))5173      return nullptr;5174  }5175 5176  // Now we know we have a GEP ptr, %inv, %ind, %inv. return the Ptr SCEV.5177  return PSE.getSCEV(Ptr);5178}5179 5180InstructionCost5181LoopVectorizationCostModel::getMemInstScalarizationCost(Instruction *I,5182                                                        ElementCount VF) {5183  assert(VF.isVector() &&5184         "Scalarization cost of instruction implies vectorization.");5185  if (VF.isScalable())5186    return InstructionCost::getInvalid();5187 5188  Type *ValTy = getLoadStoreType(I);5189  auto *SE = PSE.getSE();5190 5191  unsigned AS = getLoadStoreAddressSpace(I);5192  Value *Ptr = getLoadStorePointerOperand(I);5193  Type *PtrTy = toVectorTy(Ptr->getType(), VF);5194  // NOTE: PtrTy is a vector to signal `TTI::getAddressComputationCost`5195  //       that it is being called from this specific place.5196 5197  // Figure out whether the access is strided and get the stride value5198  // if it's known in compile time5199  const SCEV *PtrSCEV = getAddressAccessSCEV(Ptr, Legal, PSE, TheLoop);5200 5201  // Get the cost of the scalar memory instruction and address computation.5202  InstructionCost Cost = VF.getFixedValue() * TTI.getAddressComputationCost(5203                                                  PtrTy, SE, PtrSCEV, CostKind);5204 5205  // Don't pass *I here, since it is scalar but will actually be part of a5206  // vectorized loop where the user of it is a vectorized instruction.5207  const Align Alignment = getLoadStoreAlignment(I);5208  TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0));5209  Cost += VF.getFixedValue() *5210          TTI.getMemoryOpCost(I->getOpcode(), ValTy->getScalarType(), Alignment,5211                              AS, CostKind, OpInfo);5212 5213  // Get the overhead of the extractelement and insertelement instructions5214  // we might create due to scalarization.5215  Cost += getScalarizationOverhead(I, VF);5216 5217  // If we have a predicated load/store, it will need extra i1 extracts and5218  // conditional branches, but may not be executed for each vector lane. Scale5219  // the cost by the probability of executing the predicated block.5220  if (isPredicatedInst(I)) {5221    Cost /= getPredBlockCostDivisor(CostKind, I->getParent());5222 5223    // Add the cost of an i1 extract and a branch5224    auto *VecI1Ty =5225        VectorType::get(IntegerType::getInt1Ty(ValTy->getContext()), VF);5226    Cost += TTI.getScalarizationOverhead(5227        VecI1Ty, APInt::getAllOnes(VF.getFixedValue()),5228        /*Insert=*/false, /*Extract=*/true, CostKind);5229    Cost += TTI.getCFInstrCost(Instruction::Br, CostKind);5230 5231    if (useEmulatedMaskMemRefHack(I, VF))5232      // Artificially setting to a high enough value to practically disable5233      // vectorization with such operations.5234      Cost = 3000000;5235  }5236 5237  return Cost;5238}5239 5240InstructionCost5241LoopVectorizationCostModel::getConsecutiveMemOpCost(Instruction *I,5242                                                    ElementCount VF) {5243  Type *ValTy = getLoadStoreType(I);5244  auto *VectorTy = cast<VectorType>(toVectorTy(ValTy, VF));5245  Value *Ptr = getLoadStorePointerOperand(I);5246  unsigned AS = getLoadStoreAddressSpace(I);5247  int ConsecutiveStride = Legal->isConsecutivePtr(ValTy, Ptr);5248 5249  assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&5250         "Stride should be 1 or -1 for consecutive memory access");5251  const Align Alignment = getLoadStoreAlignment(I);5252  InstructionCost Cost = 0;5253  if (Legal->isMaskRequired(I)) {5254    unsigned IID = I->getOpcode() == Instruction::Load5255                       ? Intrinsic::masked_load5256                       : Intrinsic::masked_store;5257    Cost += TTI.getMemIntrinsicInstrCost(5258        MemIntrinsicCostAttributes(IID, VectorTy, Alignment, AS), CostKind);5259  } else {5260    TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0));5261    Cost += TTI.getMemoryOpCost(I->getOpcode(), VectorTy, Alignment, AS,5262                                CostKind, OpInfo, I);5263  }5264 5265  bool Reverse = ConsecutiveStride < 0;5266  if (Reverse)5267    Cost += TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy,5268                               VectorTy, {}, CostKind, 0);5269  return Cost;5270}5271 5272InstructionCost5273LoopVectorizationCostModel::getUniformMemOpCost(Instruction *I,5274                                                ElementCount VF) {5275  assert(Legal->isUniformMemOp(*I, VF));5276 5277  Type *ValTy = getLoadStoreType(I);5278  Type *PtrTy = getLoadStorePointerOperand(I)->getType();5279  auto *VectorTy = cast<VectorType>(toVectorTy(ValTy, VF));5280  const Align Alignment = getLoadStoreAlignment(I);5281  unsigned AS = getLoadStoreAddressSpace(I);5282  if (isa<LoadInst>(I)) {5283    return TTI.getAddressComputationCost(PtrTy, nullptr, nullptr, CostKind) +5284           TTI.getMemoryOpCost(Instruction::Load, ValTy, Alignment, AS,5285                               CostKind) +5286           TTI.getShuffleCost(TargetTransformInfo::SK_Broadcast, VectorTy,5287                              VectorTy, {}, CostKind);5288  }5289  StoreInst *SI = cast<StoreInst>(I);5290 5291  bool IsLoopInvariantStoreValue = Legal->isInvariant(SI->getValueOperand());5292  // TODO: We have existing tests that request the cost of extracting element5293  // VF.getKnownMinValue() - 1 from a scalable vector. This does not represent5294  // the actual generated code, which involves extracting the last element of5295  // a scalable vector where the lane to extract is unknown at compile time.5296  InstructionCost Cost =5297      TTI.getAddressComputationCost(PtrTy, nullptr, nullptr, CostKind) +5298      TTI.getMemoryOpCost(Instruction::Store, ValTy, Alignment, AS, CostKind);5299  if (!IsLoopInvariantStoreValue)5300    Cost += TTI.getIndexedVectorInstrCostFromEnd(Instruction::ExtractElement,5301                                                 VectorTy, CostKind, 0);5302  return Cost;5303}5304 5305InstructionCost5306LoopVectorizationCostModel::getGatherScatterCost(Instruction *I,5307                                                 ElementCount VF) {5308  Type *ValTy = getLoadStoreType(I);5309  auto *VectorTy = cast<VectorType>(toVectorTy(ValTy, VF));5310  const Align Alignment = getLoadStoreAlignment(I);5311  Value *Ptr = getLoadStorePointerOperand(I);5312  Type *PtrTy = Ptr->getType();5313 5314  if (!Legal->isUniform(Ptr, VF))5315    PtrTy = toVectorTy(PtrTy, VF);5316 5317  unsigned IID = I->getOpcode() == Instruction::Load5318                     ? Intrinsic::masked_gather5319                     : Intrinsic::masked_scatter;5320  return TTI.getAddressComputationCost(PtrTy, nullptr, nullptr, CostKind) +5321         TTI.getMemIntrinsicInstrCost(5322             MemIntrinsicCostAttributes(IID, VectorTy, Ptr,5323                                        Legal->isMaskRequired(I), Alignment, I),5324             CostKind);5325}5326 5327InstructionCost5328LoopVectorizationCostModel::getInterleaveGroupCost(Instruction *I,5329                                                   ElementCount VF) {5330  const auto *Group = getInterleavedAccessGroup(I);5331  assert(Group && "Fail to get an interleaved access group.");5332 5333  Instruction *InsertPos = Group->getInsertPos();5334  Type *ValTy = getLoadStoreType(InsertPos);5335  auto *VectorTy = cast<VectorType>(toVectorTy(ValTy, VF));5336  unsigned AS = getLoadStoreAddressSpace(InsertPos);5337 5338  unsigned InterleaveFactor = Group->getFactor();5339  auto *WideVecTy = VectorType::get(ValTy, VF * InterleaveFactor);5340 5341  // Holds the indices of existing members in the interleaved group.5342  SmallVector<unsigned, 4> Indices;5343  for (unsigned IF = 0; IF < InterleaveFactor; IF++)5344    if (Group->getMember(IF))5345      Indices.push_back(IF);5346 5347  // Calculate the cost of the whole interleaved group.5348  bool UseMaskForGaps =5349      (Group->requiresScalarEpilogue() && !isScalarEpilogueAllowed()) ||5350      (isa<StoreInst>(I) && !Group->isFull());5351  InstructionCost Cost = TTI.getInterleavedMemoryOpCost(5352      InsertPos->getOpcode(), WideVecTy, Group->getFactor(), Indices,5353      Group->getAlign(), AS, CostKind, Legal->isMaskRequired(I),5354      UseMaskForGaps);5355 5356  if (Group->isReverse()) {5357    // TODO: Add support for reversed masked interleaved access.5358    assert(!Legal->isMaskRequired(I) &&5359           "Reverse masked interleaved access not supported.");5360    Cost += Group->getNumMembers() *5361            TTI.getShuffleCost(TargetTransformInfo::SK_Reverse, VectorTy,5362                               VectorTy, {}, CostKind, 0);5363  }5364  return Cost;5365}5366 5367std::optional<InstructionCost>5368LoopVectorizationCostModel::getReductionPatternCost(Instruction *I,5369                                                    ElementCount VF,5370                                                    Type *Ty) const {5371  using namespace llvm::PatternMatch;5372  // Early exit for no inloop reductions5373  if (InLoopReductions.empty() || VF.isScalar() || !isa<VectorType>(Ty))5374    return std::nullopt;5375  auto *VectorTy = cast<VectorType>(Ty);5376 5377  // We are looking for a pattern of, and finding the minimal acceptable cost:5378  //  reduce(mul(ext(A), ext(B))) or5379  //  reduce(mul(A, B)) or5380  //  reduce(ext(A)) or5381  //  reduce(A).5382  // The basic idea is that we walk down the tree to do that, finding the root5383  // reduction instruction in InLoopReductionImmediateChains. From there we find5384  // the pattern of mul/ext and test the cost of the entire pattern vs the cost5385  // of the components. If the reduction cost is lower then we return it for the5386  // reduction instruction and 0 for the other instructions in the pattern. If5387  // it is not we return an invalid cost specifying the orignal cost method5388  // should be used.5389  Instruction *RetI = I;5390  if (match(RetI, m_ZExtOrSExt(m_Value()))) {5391    if (!RetI->hasOneUser())5392      return std::nullopt;5393    RetI = RetI->user_back();5394  }5395 5396  if (match(RetI, m_OneUse(m_Mul(m_Value(), m_Value()))) &&5397      RetI->user_back()->getOpcode() == Instruction::Add) {5398    RetI = RetI->user_back();5399  }5400 5401  // Test if the found instruction is a reduction, and if not return an invalid5402  // cost specifying the parent to use the original cost modelling.5403  Instruction *LastChain = InLoopReductionImmediateChains.lookup(RetI);5404  if (!LastChain)5405    return std::nullopt;5406 5407  // Find the reduction this chain is a part of and calculate the basic cost of5408  // the reduction on its own.5409  Instruction *ReductionPhi = LastChain;5410  while (!isa<PHINode>(ReductionPhi))5411    ReductionPhi = InLoopReductionImmediateChains.at(ReductionPhi);5412 5413  const RecurrenceDescriptor &RdxDesc =5414      Legal->getRecurrenceDescriptor(cast<PHINode>(ReductionPhi));5415 5416  InstructionCost BaseCost;5417  RecurKind RK = RdxDesc.getRecurrenceKind();5418  if (RecurrenceDescriptor::isMinMaxRecurrenceKind(RK)) {5419    Intrinsic::ID MinMaxID = getMinMaxReductionIntrinsicOp(RK);5420    BaseCost = TTI.getMinMaxReductionCost(MinMaxID, VectorTy,5421                                          RdxDesc.getFastMathFlags(), CostKind);5422  } else {5423    BaseCost = TTI.getArithmeticReductionCost(5424        RdxDesc.getOpcode(), VectorTy, RdxDesc.getFastMathFlags(), CostKind);5425  }5426 5427  // For a call to the llvm.fmuladd intrinsic we need to add the cost of a5428  // normal fmul instruction to the cost of the fadd reduction.5429  if (RK == RecurKind::FMulAdd)5430    BaseCost +=5431        TTI.getArithmeticInstrCost(Instruction::FMul, VectorTy, CostKind);5432 5433  // If we're using ordered reductions then we can just return the base cost5434  // here, since getArithmeticReductionCost calculates the full ordered5435  // reduction cost when FP reassociation is not allowed.5436  if (useOrderedReductions(RdxDesc))5437    return BaseCost;5438 5439  // Get the operand that was not the reduction chain and match it to one of the5440  // patterns, returning the better cost if it is found.5441  Instruction *RedOp = RetI->getOperand(1) == LastChain5442                           ? dyn_cast<Instruction>(RetI->getOperand(0))5443                           : dyn_cast<Instruction>(RetI->getOperand(1));5444 5445  VectorTy = VectorType::get(I->getOperand(0)->getType(), VectorTy);5446 5447  Instruction *Op0, *Op1;5448  if (RedOp && RdxDesc.getOpcode() == Instruction::Add &&5449      match(RedOp,5450            m_ZExtOrSExt(m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) &&5451      match(Op0, m_ZExtOrSExt(m_Value())) &&5452      Op0->getOpcode() == Op1->getOpcode() &&5453      Op0->getOperand(0)->getType() == Op1->getOperand(0)->getType() &&5454      !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1) &&5455      (Op0->getOpcode() == RedOp->getOpcode() || Op0 == Op1)) {5456 5457    // Matched reduce.add(ext(mul(ext(A), ext(B)))5458    // Note that the extend opcodes need to all match, or if A==B they will have5459    // been converted to zext(mul(sext(A), sext(A))) as it is known positive,5460    // which is equally fine.5461    bool IsUnsigned = isa<ZExtInst>(Op0);5462    auto *ExtType = VectorType::get(Op0->getOperand(0)->getType(), VectorTy);5463    auto *MulType = VectorType::get(Op0->getType(), VectorTy);5464 5465    InstructionCost ExtCost =5466        TTI.getCastInstrCost(Op0->getOpcode(), MulType, ExtType,5467                             TTI::CastContextHint::None, CostKind, Op0);5468    InstructionCost MulCost =5469        TTI.getArithmeticInstrCost(Instruction::Mul, MulType, CostKind);5470    InstructionCost Ext2Cost =5471        TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, MulType,5472                             TTI::CastContextHint::None, CostKind, RedOp);5473 5474    InstructionCost RedCost = TTI.getMulAccReductionCost(5475        IsUnsigned, RdxDesc.getOpcode(), RdxDesc.getRecurrenceType(), ExtType,5476        CostKind);5477 5478    if (RedCost.isValid() &&5479        RedCost < ExtCost * 2 + MulCost + Ext2Cost + BaseCost)5480      return I == RetI ? RedCost : 0;5481  } else if (RedOp && match(RedOp, m_ZExtOrSExt(m_Value())) &&5482             !TheLoop->isLoopInvariant(RedOp)) {5483    // Matched reduce(ext(A))5484    bool IsUnsigned = isa<ZExtInst>(RedOp);5485    auto *ExtType = VectorType::get(RedOp->getOperand(0)->getType(), VectorTy);5486    InstructionCost RedCost = TTI.getExtendedReductionCost(5487        RdxDesc.getOpcode(), IsUnsigned, RdxDesc.getRecurrenceType(), ExtType,5488        RdxDesc.getFastMathFlags(), CostKind);5489 5490    InstructionCost ExtCost =5491        TTI.getCastInstrCost(RedOp->getOpcode(), VectorTy, ExtType,5492                             TTI::CastContextHint::None, CostKind, RedOp);5493    if (RedCost.isValid() && RedCost < BaseCost + ExtCost)5494      return I == RetI ? RedCost : 0;5495  } else if (RedOp && RdxDesc.getOpcode() == Instruction::Add &&5496             match(RedOp, m_Mul(m_Instruction(Op0), m_Instruction(Op1)))) {5497    if (match(Op0, m_ZExtOrSExt(m_Value())) &&5498        Op0->getOpcode() == Op1->getOpcode() &&5499        !TheLoop->isLoopInvariant(Op0) && !TheLoop->isLoopInvariant(Op1)) {5500      bool IsUnsigned = isa<ZExtInst>(Op0);5501      Type *Op0Ty = Op0->getOperand(0)->getType();5502      Type *Op1Ty = Op1->getOperand(0)->getType();5503      Type *LargestOpTy =5504          Op0Ty->getIntegerBitWidth() < Op1Ty->getIntegerBitWidth() ? Op1Ty5505                                                                    : Op0Ty;5506      auto *ExtType = VectorType::get(LargestOpTy, VectorTy);5507 5508      // Matched reduce.add(mul(ext(A), ext(B))), where the two ext may be of5509      // different sizes. We take the largest type as the ext to reduce, and add5510      // the remaining cost as, for example reduce(mul(ext(ext(A)), ext(B))).5511      InstructionCost ExtCost0 = TTI.getCastInstrCost(5512          Op0->getOpcode(), VectorTy, VectorType::get(Op0Ty, VectorTy),5513          TTI::CastContextHint::None, CostKind, Op0);5514      InstructionCost ExtCost1 = TTI.getCastInstrCost(5515          Op1->getOpcode(), VectorTy, VectorType::get(Op1Ty, VectorTy),5516          TTI::CastContextHint::None, CostKind, Op1);5517      InstructionCost MulCost =5518          TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind);5519 5520      InstructionCost RedCost = TTI.getMulAccReductionCost(5521          IsUnsigned, RdxDesc.getOpcode(), RdxDesc.getRecurrenceType(), ExtType,5522          CostKind);5523      InstructionCost ExtraExtCost = 0;5524      if (Op0Ty != LargestOpTy || Op1Ty != LargestOpTy) {5525        Instruction *ExtraExtOp = (Op0Ty != LargestOpTy) ? Op0 : Op1;5526        ExtraExtCost = TTI.getCastInstrCost(5527            ExtraExtOp->getOpcode(), ExtType,5528            VectorType::get(ExtraExtOp->getOperand(0)->getType(), VectorTy),5529            TTI::CastContextHint::None, CostKind, ExtraExtOp);5530      }5531 5532      if (RedCost.isValid() &&5533          (RedCost + ExtraExtCost) < (ExtCost0 + ExtCost1 + MulCost + BaseCost))5534        return I == RetI ? RedCost : 0;5535    } else if (!match(I, m_ZExtOrSExt(m_Value()))) {5536      // Matched reduce.add(mul())5537      InstructionCost MulCost =5538          TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind);5539 5540      InstructionCost RedCost = TTI.getMulAccReductionCost(5541          true, RdxDesc.getOpcode(), RdxDesc.getRecurrenceType(), VectorTy,5542          CostKind);5543 5544      if (RedCost.isValid() && RedCost < MulCost + BaseCost)5545        return I == RetI ? RedCost : 0;5546    }5547  }5548 5549  return I == RetI ? std::optional<InstructionCost>(BaseCost) : std::nullopt;5550}5551 5552InstructionCost5553LoopVectorizationCostModel::getMemoryInstructionCost(Instruction *I,5554                                                     ElementCount VF) {5555  // Calculate scalar cost only. Vectorization cost should be ready at this5556  // moment.5557  if (VF.isScalar()) {5558    Type *ValTy = getLoadStoreType(I);5559    Type *PtrTy = getLoadStorePointerOperand(I)->getType();5560    const Align Alignment = getLoadStoreAlignment(I);5561    unsigned AS = getLoadStoreAddressSpace(I);5562 5563    TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(I->getOperand(0));5564    return TTI.getAddressComputationCost(PtrTy, nullptr, nullptr, CostKind) +5565           TTI.getMemoryOpCost(I->getOpcode(), ValTy, Alignment, AS, CostKind,5566                               OpInfo, I);5567  }5568  return getWideningCost(I, VF);5569}5570 5571InstructionCost5572LoopVectorizationCostModel::getScalarizationOverhead(Instruction *I,5573                                                     ElementCount VF) const {5574 5575  // There is no mechanism yet to create a scalable scalarization loop,5576  // so this is currently Invalid.5577  if (VF.isScalable())5578    return InstructionCost::getInvalid();5579 5580  if (VF.isScalar())5581    return 0;5582 5583  InstructionCost Cost = 0;5584  Type *RetTy = toVectorizedTy(I->getType(), VF);5585  if (!RetTy->isVoidTy() &&5586      (!isa<LoadInst>(I) || !TTI.supportsEfficientVectorElementLoadStore())) {5587 5588    for (Type *VectorTy : getContainedTypes(RetTy)) {5589      Cost += TTI.getScalarizationOverhead(5590          cast<VectorType>(VectorTy), APInt::getAllOnes(VF.getFixedValue()),5591          /*Insert=*/true,5592          /*Extract=*/false, CostKind);5593    }5594  }5595 5596  // Some targets keep addresses scalar.5597  if (isa<LoadInst>(I) && !TTI.prefersVectorizedAddressing())5598    return Cost;5599 5600  // Some targets support efficient element stores.5601  if (isa<StoreInst>(I) && TTI.supportsEfficientVectorElementLoadStore())5602    return Cost;5603 5604  // Collect operands to consider.5605  CallInst *CI = dyn_cast<CallInst>(I);5606  Instruction::op_range Ops = CI ? CI->args() : I->operands();5607 5608  // Skip operands that do not require extraction/scalarization and do not incur5609  // any overhead.5610  SmallVector<Type *> Tys;5611  for (auto *V : filterExtractingOperands(Ops, VF))5612    Tys.push_back(maybeVectorizeType(V->getType(), VF));5613  return Cost + TTI.getOperandsScalarizationOverhead(Tys, CostKind);5614}5615 5616void LoopVectorizationCostModel::setCostBasedWideningDecision(ElementCount VF) {5617  if (VF.isScalar())5618    return;5619  NumPredStores = 0;5620  for (BasicBlock *BB : TheLoop->blocks()) {5621    // For each instruction in the old loop.5622    for (Instruction &I : *BB) {5623      Value *Ptr =  getLoadStorePointerOperand(&I);5624      if (!Ptr)5625        continue;5626 5627      // TODO: We should generate better code and update the cost model for5628      // predicated uniform stores. Today they are treated as any other5629      // predicated store (see added test cases in5630      // invariant-store-vectorization.ll).5631      if (isa<StoreInst>(&I) && isScalarWithPredication(&I, VF))5632        NumPredStores++;5633 5634      if (Legal->isUniformMemOp(I, VF)) {5635        auto IsLegalToScalarize = [&]() {5636          if (!VF.isScalable())5637            // Scalarization of fixed length vectors "just works".5638            return true;5639 5640          // We have dedicated lowering for unpredicated uniform loads and5641          // stores.  Note that even with tail folding we know that at least5642          // one lane is active (i.e. generalized predication is not possible5643          // here), and the logic below depends on this fact.5644          if (!foldTailByMasking())5645            return true;5646 5647          // For scalable vectors, a uniform memop load is always5648          // uniform-by-parts  and we know how to scalarize that.5649          if (isa<LoadInst>(I))5650            return true;5651 5652          // A uniform store isn't neccessarily uniform-by-part5653          // and we can't assume scalarization.5654          auto &SI = cast<StoreInst>(I);5655          return TheLoop->isLoopInvariant(SI.getValueOperand());5656        };5657 5658        const InstructionCost GatherScatterCost =5659          isLegalGatherOrScatter(&I, VF) ?5660          getGatherScatterCost(&I, VF) : InstructionCost::getInvalid();5661 5662        // Load: Scalar load + broadcast5663        // Store: Scalar store + isLoopInvariantStoreValue ? 0 : extract5664        // FIXME: This cost is a significant under-estimate for tail folded5665        // memory ops.5666        const InstructionCost ScalarizationCost =5667            IsLegalToScalarize() ? getUniformMemOpCost(&I, VF)5668                                 : InstructionCost::getInvalid();5669 5670        // Choose better solution for the current VF,  Note that Invalid5671        // costs compare as maximumal large.  If both are invalid, we get5672        // scalable invalid which signals a failure and a vectorization abort.5673        if (GatherScatterCost < ScalarizationCost)5674          setWideningDecision(&I, VF, CM_GatherScatter, GatherScatterCost);5675        else5676          setWideningDecision(&I, VF, CM_Scalarize, ScalarizationCost);5677        continue;5678      }5679 5680      // We assume that widening is the best solution when possible.5681      if (memoryInstructionCanBeWidened(&I, VF)) {5682        InstructionCost Cost = getConsecutiveMemOpCost(&I, VF);5683        int ConsecutiveStride = Legal->isConsecutivePtr(5684            getLoadStoreType(&I), getLoadStorePointerOperand(&I));5685        assert((ConsecutiveStride == 1 || ConsecutiveStride == -1) &&5686               "Expected consecutive stride.");5687        InstWidening Decision =5688            ConsecutiveStride == 1 ? CM_Widen : CM_Widen_Reverse;5689        setWideningDecision(&I, VF, Decision, Cost);5690        continue;5691      }5692 5693      // Choose between Interleaving, Gather/Scatter or Scalarization.5694      InstructionCost InterleaveCost = InstructionCost::getInvalid();5695      unsigned NumAccesses = 1;5696      if (isAccessInterleaved(&I)) {5697        const auto *Group = getInterleavedAccessGroup(&I);5698        assert(Group && "Fail to get an interleaved access group.");5699 5700        // Make one decision for the whole group.5701        if (getWideningDecision(&I, VF) != CM_Unknown)5702          continue;5703 5704        NumAccesses = Group->getNumMembers();5705        if (interleavedAccessCanBeWidened(&I, VF))5706          InterleaveCost = getInterleaveGroupCost(&I, VF);5707      }5708 5709      InstructionCost GatherScatterCost =5710          isLegalGatherOrScatter(&I, VF)5711              ? getGatherScatterCost(&I, VF) * NumAccesses5712              : InstructionCost::getInvalid();5713 5714      InstructionCost ScalarizationCost =5715          getMemInstScalarizationCost(&I, VF) * NumAccesses;5716 5717      // Choose better solution for the current VF,5718      // write down this decision and use it during vectorization.5719      InstructionCost Cost;5720      InstWidening Decision;5721      if (InterleaveCost <= GatherScatterCost &&5722          InterleaveCost < ScalarizationCost) {5723        Decision = CM_Interleave;5724        Cost = InterleaveCost;5725      } else if (GatherScatterCost < ScalarizationCost) {5726        Decision = CM_GatherScatter;5727        Cost = GatherScatterCost;5728      } else {5729        Decision = CM_Scalarize;5730        Cost = ScalarizationCost;5731      }5732      // If the instructions belongs to an interleave group, the whole group5733      // receives the same decision. The whole group receives the cost, but5734      // the cost will actually be assigned to one instruction.5735      if (const auto *Group = getInterleavedAccessGroup(&I)) {5736        if (Decision == CM_Scalarize) {5737          for (unsigned Idx = 0; Idx < Group->getFactor(); ++Idx) {5738            if (auto *I = Group->getMember(Idx)) {5739              setWideningDecision(I, VF, Decision,5740                                  getMemInstScalarizationCost(I, VF));5741            }5742          }5743        } else {5744          setWideningDecision(Group, VF, Decision, Cost);5745        }5746      } else5747        setWideningDecision(&I, VF, Decision, Cost);5748    }5749  }5750 5751  // Make sure that any load of address and any other address computation5752  // remains scalar unless there is gather/scatter support. This avoids5753  // inevitable extracts into address registers, and also has the benefit of5754  // activating LSR more, since that pass can't optimize vectorized5755  // addresses.5756  if (TTI.prefersVectorizedAddressing())5757    return;5758 5759  // Start with all scalar pointer uses.5760  SmallPtrSet<Instruction *, 8> AddrDefs;5761  for (BasicBlock *BB : TheLoop->blocks())5762    for (Instruction &I : *BB) {5763      Instruction *PtrDef =5764        dyn_cast_or_null<Instruction>(getLoadStorePointerOperand(&I));5765      if (PtrDef && TheLoop->contains(PtrDef) &&5766          getWideningDecision(&I, VF) != CM_GatherScatter)5767        AddrDefs.insert(PtrDef);5768    }5769 5770  // Add all instructions used to generate the addresses.5771  SmallVector<Instruction *, 4> Worklist;5772  append_range(Worklist, AddrDefs);5773  while (!Worklist.empty()) {5774    Instruction *I = Worklist.pop_back_val();5775    for (auto &Op : I->operands())5776      if (auto *InstOp = dyn_cast<Instruction>(Op))5777        if (TheLoop->contains(InstOp) && !isa<PHINode>(InstOp) &&5778            AddrDefs.insert(InstOp).second)5779          Worklist.push_back(InstOp);5780  }5781 5782  auto UpdateMemOpUserCost = [this, VF](LoadInst *LI) {5783    // If there are direct memory op users of the newly scalarized load,5784    // their cost may have changed because there's no scalarization5785    // overhead for the operand. Update it.5786    for (User *U : LI->users()) {5787      if (!isa<LoadInst, StoreInst>(U))5788        continue;5789      if (getWideningDecision(cast<Instruction>(U), VF) != CM_Scalarize)5790        continue;5791      setWideningDecision(5792          cast<Instruction>(U), VF, CM_Scalarize,5793          getMemInstScalarizationCost(cast<Instruction>(U), VF));5794    }5795  };5796  for (auto *I : AddrDefs) {5797    if (isa<LoadInst>(I)) {5798      // Setting the desired widening decision should ideally be handled in5799      // by cost functions, but since this involves the task of finding out5800      // if the loaded register is involved in an address computation, it is5801      // instead changed here when we know this is the case.5802      InstWidening Decision = getWideningDecision(I, VF);5803      if (Decision == CM_Widen || Decision == CM_Widen_Reverse ||5804          (!isPredicatedInst(I) && !Legal->isUniformMemOp(*I, VF) &&5805           Decision == CM_Scalarize)) {5806        // Scalarize a widened load of address or update the cost of a scalar5807        // load of an address.5808        setWideningDecision(5809            I, VF, CM_Scalarize,5810            (VF.getKnownMinValue() *5811             getMemoryInstructionCost(I, ElementCount::getFixed(1))));5812        UpdateMemOpUserCost(cast<LoadInst>(I));5813      } else if (const auto *Group = getInterleavedAccessGroup(I)) {5814        // Scalarize all members of this interleaved group when any member5815        // is used as an address. The address-used load skips scalarization5816        // overhead, other members include it.5817        for (unsigned Idx = 0; Idx < Group->getFactor(); ++Idx) {5818          if (Instruction *Member = Group->getMember(Idx)) {5819            InstructionCost Cost =5820                AddrDefs.contains(Member)5821                    ? (VF.getKnownMinValue() *5822                       getMemoryInstructionCost(Member,5823                                                ElementCount::getFixed(1)))5824                    : getMemInstScalarizationCost(Member, VF);5825            setWideningDecision(Member, VF, CM_Scalarize, Cost);5826            UpdateMemOpUserCost(cast<LoadInst>(Member));5827          }5828        }5829      }5830    } else {5831      // Cannot scalarize fixed-order recurrence phis at the moment.5832      if (isa<PHINode>(I) && Legal->isFixedOrderRecurrence(cast<PHINode>(I)))5833        continue;5834 5835      // Make sure I gets scalarized and a cost estimate without5836      // scalarization overhead.5837      ForcedScalars[VF].insert(I);5838    }5839  }5840}5841 5842void LoopVectorizationCostModel::setVectorizedCallDecision(ElementCount VF) {5843  assert(!VF.isScalar() &&5844         "Trying to set a vectorization decision for a scalar VF");5845 5846  auto ForcedScalar = ForcedScalars.find(VF);5847  for (BasicBlock *BB : TheLoop->blocks()) {5848    // For each instruction in the old loop.5849    for (Instruction &I : *BB) {5850      CallInst *CI = dyn_cast<CallInst>(&I);5851 5852      if (!CI)5853        continue;5854 5855      InstructionCost ScalarCost = InstructionCost::getInvalid();5856      InstructionCost VectorCost = InstructionCost::getInvalid();5857      InstructionCost IntrinsicCost = InstructionCost::getInvalid();5858      Function *ScalarFunc = CI->getCalledFunction();5859      Type *ScalarRetTy = CI->getType();5860      SmallVector<Type *, 4> Tys, ScalarTys;5861      for (auto &ArgOp : CI->args())5862        ScalarTys.push_back(ArgOp->getType());5863 5864      // Estimate cost of scalarized vector call. The source operands are5865      // assumed to be vectors, so we need to extract individual elements from5866      // there, execute VF scalar calls, and then gather the result into the5867      // vector return value.5868      if (VF.isFixed()) {5869        InstructionCost ScalarCallCost =5870            TTI.getCallInstrCost(ScalarFunc, ScalarRetTy, ScalarTys, CostKind);5871 5872        // Compute costs of unpacking argument values for the scalar calls and5873        // packing the return values to a vector.5874        InstructionCost ScalarizationCost = getScalarizationOverhead(CI, VF);5875        ScalarCost = ScalarCallCost * VF.getKnownMinValue() + ScalarizationCost;5876      } else {5877        // There is no point attempting to calculate the scalar cost for a5878        // scalable VF as we know it will be Invalid.5879        assert(!getScalarizationOverhead(CI, VF).isValid() &&5880               "Unexpected valid cost for scalarizing scalable vectors");5881        ScalarCost = InstructionCost::getInvalid();5882      }5883 5884      // Honor ForcedScalars and UniformAfterVectorization decisions.5885      // TODO: For calls, it might still be more profitable to widen. Use5886      // VPlan-based cost model to compare different options.5887      if (VF.isVector() && ((ForcedScalar != ForcedScalars.end() &&5888                             ForcedScalar->second.contains(CI)) ||5889                            isUniformAfterVectorization(CI, VF))) {5890        setCallWideningDecision(CI, VF, CM_Scalarize, nullptr,5891                                Intrinsic::not_intrinsic, std::nullopt,5892                                ScalarCost);5893        continue;5894      }5895 5896      bool MaskRequired = Legal->isMaskRequired(CI);5897      // Compute corresponding vector type for return value and arguments.5898      Type *RetTy = toVectorizedTy(ScalarRetTy, VF);5899      for (Type *ScalarTy : ScalarTys)5900        Tys.push_back(toVectorizedTy(ScalarTy, VF));5901 5902      // An in-loop reduction using an fmuladd intrinsic is a special case;5903      // we don't want the normal cost for that intrinsic.5904      if (RecurrenceDescriptor::isFMulAddIntrinsic(CI))5905        if (auto RedCost = getReductionPatternCost(CI, VF, RetTy)) {5906          setCallWideningDecision(CI, VF, CM_IntrinsicCall, nullptr,5907                                  getVectorIntrinsicIDForCall(CI, TLI),5908                                  std::nullopt, *RedCost);5909          continue;5910        }5911 5912      // Find the cost of vectorizing the call, if we can find a suitable5913      // vector variant of the function.5914      VFInfo FuncInfo;5915      Function *VecFunc = nullptr;5916      // Search through any available variants for one we can use at this VF.5917      for (VFInfo &Info : VFDatabase::getMappings(*CI)) {5918        // Must match requested VF.5919        if (Info.Shape.VF != VF)5920          continue;5921 5922        // Must take a mask argument if one is required5923        if (MaskRequired && !Info.isMasked())5924          continue;5925 5926        // Check that all parameter kinds are supported5927        bool ParamsOk = true;5928        for (VFParameter Param : Info.Shape.Parameters) {5929          switch (Param.ParamKind) {5930          case VFParamKind::Vector:5931            break;5932          case VFParamKind::OMP_Uniform: {5933            Value *ScalarParam = CI->getArgOperand(Param.ParamPos);5934            // Make sure the scalar parameter in the loop is invariant.5935            if (!PSE.getSE()->isLoopInvariant(PSE.getSCEV(ScalarParam),5936                                              TheLoop))5937              ParamsOk = false;5938            break;5939          }5940          case VFParamKind::OMP_Linear: {5941            Value *ScalarParam = CI->getArgOperand(Param.ParamPos);5942            // Find the stride for the scalar parameter in this loop and see if5943            // it matches the stride for the variant.5944            // TODO: do we need to figure out the cost of an extract to get the5945            // first lane? Or do we hope that it will be folded away?5946            ScalarEvolution *SE = PSE.getSE();5947            if (!match(SE->getSCEV(ScalarParam),5948                       m_scev_AffineAddRec(5949                           m_SCEV(), m_scev_SpecificSInt(Param.LinearStepOrPos),5950                           m_SpecificLoop(TheLoop))))5951              ParamsOk = false;5952            break;5953          }5954          case VFParamKind::GlobalPredicate:5955            break;5956          default:5957            ParamsOk = false;5958            break;5959          }5960        }5961 5962        if (!ParamsOk)5963          continue;5964 5965        // Found a suitable candidate, stop here.5966        VecFunc = CI->getModule()->getFunction(Info.VectorName);5967        FuncInfo = Info;5968        break;5969      }5970 5971      if (TLI && VecFunc && !CI->isNoBuiltin())5972        VectorCost = TTI.getCallInstrCost(nullptr, RetTy, Tys, CostKind);5973 5974      // Find the cost of an intrinsic; some targets may have instructions that5975      // perform the operation without needing an actual call.5976      Intrinsic::ID IID = getVectorIntrinsicIDForCall(CI, TLI);5977      if (IID != Intrinsic::not_intrinsic)5978        IntrinsicCost = getVectorIntrinsicCost(CI, VF);5979 5980      InstructionCost Cost = ScalarCost;5981      InstWidening Decision = CM_Scalarize;5982 5983      if (VectorCost <= Cost) {5984        Cost = VectorCost;5985        Decision = CM_VectorCall;5986      }5987 5988      if (IntrinsicCost <= Cost) {5989        Cost = IntrinsicCost;5990        Decision = CM_IntrinsicCall;5991      }5992 5993      setCallWideningDecision(CI, VF, Decision, VecFunc, IID,5994                              FuncInfo.getParamIndexForOptionalMask(), Cost);5995    }5996  }5997}5998 5999bool LoopVectorizationCostModel::shouldConsiderInvariant(Value *Op) {6000  if (!Legal->isInvariant(Op))6001    return false;6002  // Consider Op invariant, if it or its operands aren't predicated6003  // instruction in the loop. In that case, it is not trivially hoistable.6004  auto *OpI = dyn_cast<Instruction>(Op);6005  return !OpI || !TheLoop->contains(OpI) ||6006         (!isPredicatedInst(OpI) &&6007          (!isa<PHINode>(OpI) || OpI->getParent() != TheLoop->getHeader()) &&6008          all_of(OpI->operands(),6009                 [this](Value *Op) { return shouldConsiderInvariant(Op); }));6010}6011 6012InstructionCost6013LoopVectorizationCostModel::getInstructionCost(Instruction *I,6014                                               ElementCount VF) {6015  // If we know that this instruction will remain uniform, check the cost of6016  // the scalar version.6017  if (isUniformAfterVectorization(I, VF))6018    VF = ElementCount::getFixed(1);6019 6020  if (VF.isVector() && isProfitableToScalarize(I, VF))6021    return InstsToScalarize[VF][I];6022 6023  // Forced scalars do not have any scalarization overhead.6024  auto ForcedScalar = ForcedScalars.find(VF);6025  if (VF.isVector() && ForcedScalar != ForcedScalars.end()) {6026    auto InstSet = ForcedScalar->second;6027    if (InstSet.count(I))6028      return getInstructionCost(I, ElementCount::getFixed(1)) *6029             VF.getKnownMinValue();6030  }6031 6032  Type *RetTy = I->getType();6033  if (canTruncateToMinimalBitwidth(I, VF))6034    RetTy = IntegerType::get(RetTy->getContext(), MinBWs[I]);6035  auto *SE = PSE.getSE();6036 6037  Type *VectorTy;6038  if (isScalarAfterVectorization(I, VF)) {6039    [[maybe_unused]] auto HasSingleCopyAfterVectorization =6040        [this](Instruction *I, ElementCount VF) -> bool {6041      if (VF.isScalar())6042        return true;6043 6044      auto Scalarized = InstsToScalarize.find(VF);6045      assert(Scalarized != InstsToScalarize.end() &&6046             "VF not yet analyzed for scalarization profitability");6047      return !Scalarized->second.count(I) &&6048             llvm::all_of(I->users(), [&](User *U) {6049               auto *UI = cast<Instruction>(U);6050               return !Scalarized->second.count(UI);6051             });6052    };6053 6054    // With the exception of GEPs and PHIs, after scalarization there should6055    // only be one copy of the instruction generated in the loop. This is6056    // because the VF is either 1, or any instructions that need scalarizing6057    // have already been dealt with by the time we get here. As a result,6058    // it means we don't have to multiply the instruction cost by VF.6059    assert(I->getOpcode() == Instruction::GetElementPtr ||6060           I->getOpcode() == Instruction::PHI ||6061           (I->getOpcode() == Instruction::BitCast &&6062            I->getType()->isPointerTy()) ||6063           HasSingleCopyAfterVectorization(I, VF));6064    VectorTy = RetTy;6065  } else6066    VectorTy = toVectorizedTy(RetTy, VF);6067 6068  if (VF.isVector() && VectorTy->isVectorTy() &&6069      !TTI.getNumberOfParts(VectorTy))6070    return InstructionCost::getInvalid();6071 6072  // TODO: We need to estimate the cost of intrinsic calls.6073  switch (I->getOpcode()) {6074  case Instruction::GetElementPtr:6075    // We mark this instruction as zero-cost because the cost of GEPs in6076    // vectorized code depends on whether the corresponding memory instruction6077    // is scalarized or not. Therefore, we handle GEPs with the memory6078    // instruction cost.6079    return 0;6080  case Instruction::Br: {6081    // In cases of scalarized and predicated instructions, there will be VF6082    // predicated blocks in the vectorized loop. Each branch around these6083    // blocks requires also an extract of its vector compare i1 element.6084    // Note that the conditional branch from the loop latch will be replaced by6085    // a single branch controlling the loop, so there is no extra overhead from6086    // scalarization.6087    bool ScalarPredicatedBB = false;6088    BranchInst *BI = cast<BranchInst>(I);6089    if (VF.isVector() && BI->isConditional() &&6090        (PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(0)) ||6091         PredicatedBBsAfterVectorization[VF].count(BI->getSuccessor(1))) &&6092        BI->getParent() != TheLoop->getLoopLatch())6093      ScalarPredicatedBB = true;6094 6095    if (ScalarPredicatedBB) {6096      // Not possible to scalarize scalable vector with predicated instructions.6097      if (VF.isScalable())6098        return InstructionCost::getInvalid();6099      // Return cost for branches around scalarized and predicated blocks.6100      auto *VecI1Ty =6101          VectorType::get(IntegerType::getInt1Ty(RetTy->getContext()), VF);6102      return (6103          TTI.getScalarizationOverhead(6104              VecI1Ty, APInt::getAllOnes(VF.getFixedValue()),6105              /*Insert*/ false, /*Extract*/ true, CostKind) +6106          (TTI.getCFInstrCost(Instruction::Br, CostKind) * VF.getFixedValue()));6107    }6108 6109    if (I->getParent() == TheLoop->getLoopLatch() || VF.isScalar())6110      // The back-edge branch will remain, as will all scalar branches.6111      return TTI.getCFInstrCost(Instruction::Br, CostKind);6112 6113    // This branch will be eliminated by if-conversion.6114    return 0;6115    // Note: We currently assume zero cost for an unconditional branch inside6116    // a predicated block since it will become a fall-through, although we6117    // may decide in the future to call TTI for all branches.6118  }6119  case Instruction::Switch: {6120    if (VF.isScalar())6121      return TTI.getCFInstrCost(Instruction::Switch, CostKind);6122    auto *Switch = cast<SwitchInst>(I);6123    return Switch->getNumCases() *6124           TTI.getCmpSelInstrCost(6125               Instruction::ICmp,6126               toVectorTy(Switch->getCondition()->getType(), VF),6127               toVectorTy(Type::getInt1Ty(I->getContext()), VF),6128               CmpInst::ICMP_EQ, CostKind);6129  }6130  case Instruction::PHI: {6131    auto *Phi = cast<PHINode>(I);6132 6133    // First-order recurrences are replaced by vector shuffles inside the loop.6134    if (VF.isVector() && Legal->isFixedOrderRecurrence(Phi)) {6135      SmallVector<int> Mask(VF.getKnownMinValue());6136      std::iota(Mask.begin(), Mask.end(), VF.getKnownMinValue() - 1);6137      return TTI.getShuffleCost(TargetTransformInfo::SK_Splice,6138                                cast<VectorType>(VectorTy),6139                                cast<VectorType>(VectorTy), Mask, CostKind,6140                                VF.getKnownMinValue() - 1);6141    }6142 6143    // Phi nodes in non-header blocks (not inductions, reductions, etc.) are6144    // converted into select instructions. We require N - 1 selects per phi6145    // node, where N is the number of incoming values.6146    if (VF.isVector() && Phi->getParent() != TheLoop->getHeader()) {6147      Type *ResultTy = Phi->getType();6148 6149      // All instructions in an Any-of reduction chain are narrowed to bool.6150      // Check if that is the case for this phi node.6151      auto *HeaderUser = cast_if_present<PHINode>(6152          find_singleton<User>(Phi->users(), [this](User *U, bool) -> User * {6153            auto *Phi = dyn_cast<PHINode>(U);6154            if (Phi && Phi->getParent() == TheLoop->getHeader())6155              return Phi;6156            return nullptr;6157          }));6158      if (HeaderUser) {6159        auto &ReductionVars = Legal->getReductionVars();6160        auto Iter = ReductionVars.find(HeaderUser);6161        if (Iter != ReductionVars.end() &&6162            RecurrenceDescriptor::isAnyOfRecurrenceKind(6163                Iter->second.getRecurrenceKind()))6164          ResultTy = Type::getInt1Ty(Phi->getContext());6165      }6166      return (Phi->getNumIncomingValues() - 1) *6167             TTI.getCmpSelInstrCost(6168                 Instruction::Select, toVectorTy(ResultTy, VF),6169                 toVectorTy(Type::getInt1Ty(Phi->getContext()), VF),6170                 CmpInst::BAD_ICMP_PREDICATE, CostKind);6171    }6172 6173    // When tail folding with EVL, if the phi is part of an out of loop6174    // reduction then it will be transformed into a wide vp_merge.6175    if (VF.isVector() && foldTailWithEVL() &&6176        Legal->getReductionVars().contains(Phi) && !isInLoopReduction(Phi)) {6177      IntrinsicCostAttributes ICA(6178          Intrinsic::vp_merge, toVectorTy(Phi->getType(), VF),6179          {toVectorTy(Type::getInt1Ty(Phi->getContext()), VF)});6180      return TTI.getIntrinsicInstrCost(ICA, CostKind);6181    }6182 6183    return TTI.getCFInstrCost(Instruction::PHI, CostKind);6184  }6185  case Instruction::UDiv:6186  case Instruction::SDiv:6187  case Instruction::URem:6188  case Instruction::SRem:6189    if (VF.isVector() && isPredicatedInst(I)) {6190      const auto [ScalarCost, SafeDivisorCost] = getDivRemSpeculationCost(I, VF);6191      return isDivRemScalarWithPredication(ScalarCost, SafeDivisorCost) ?6192        ScalarCost : SafeDivisorCost;6193    }6194    // We've proven all lanes safe to speculate, fall through.6195    [[fallthrough]];6196  case Instruction::Add:6197  case Instruction::Sub: {6198    auto Info = Legal->getHistogramInfo(I);6199    if (Info && VF.isVector()) {6200      const HistogramInfo *HGram = Info.value();6201      // Assume that a non-constant update value (or a constant != 1) requires6202      // a multiply, and add that into the cost.6203      InstructionCost MulCost = TTI::TCC_Free;6204      ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1));6205      if (!RHS || RHS->getZExtValue() != 1)6206        MulCost =6207            TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind);6208 6209      // Find the cost of the histogram operation itself.6210      Type *PtrTy = VectorType::get(HGram->Load->getPointerOperandType(), VF);6211      Type *ScalarTy = I->getType();6212      Type *MaskTy = VectorType::get(Type::getInt1Ty(I->getContext()), VF);6213      IntrinsicCostAttributes ICA(Intrinsic::experimental_vector_histogram_add,6214                                  Type::getVoidTy(I->getContext()),6215                                  {PtrTy, ScalarTy, MaskTy});6216 6217      // Add the costs together with the add/sub operation.6218      return TTI.getIntrinsicInstrCost(ICA, CostKind) + MulCost +6219             TTI.getArithmeticInstrCost(I->getOpcode(), VectorTy, CostKind);6220    }6221    [[fallthrough]];6222  }6223  case Instruction::FAdd:6224  case Instruction::FSub:6225  case Instruction::Mul:6226  case Instruction::FMul:6227  case Instruction::FDiv:6228  case Instruction::FRem:6229  case Instruction::Shl:6230  case Instruction::LShr:6231  case Instruction::AShr:6232  case Instruction::And:6233  case Instruction::Or:6234  case Instruction::Xor: {6235    // If we're speculating on the stride being 1, the multiplication may6236    // fold away.  We can generalize this for all operations using the notion6237    // of neutral elements.  (TODO)6238    if (I->getOpcode() == Instruction::Mul &&6239        ((TheLoop->isLoopInvariant(I->getOperand(0)) &&6240          PSE.getSCEV(I->getOperand(0))->isOne()) ||6241         (TheLoop->isLoopInvariant(I->getOperand(1)) &&6242          PSE.getSCEV(I->getOperand(1))->isOne())))6243      return 0;6244 6245    // Detect reduction patterns6246    if (auto RedCost = getReductionPatternCost(I, VF, VectorTy))6247      return *RedCost;6248 6249    // Certain instructions can be cheaper to vectorize if they have a constant6250    // second vector operand. One example of this are shifts on x86.6251    Value *Op2 = I->getOperand(1);6252    if (!isa<Constant>(Op2) && TheLoop->isLoopInvariant(Op2) &&6253        PSE.getSE()->isSCEVable(Op2->getType()) &&6254        isa<SCEVConstant>(PSE.getSCEV(Op2))) {6255      Op2 = cast<SCEVConstant>(PSE.getSCEV(Op2))->getValue();6256    }6257    auto Op2Info = TTI.getOperandInfo(Op2);6258    if (Op2Info.Kind == TargetTransformInfo::OK_AnyValue &&6259        shouldConsiderInvariant(Op2))6260      Op2Info.Kind = TargetTransformInfo::OK_UniformValue;6261 6262    SmallVector<const Value *, 4> Operands(I->operand_values());6263    return TTI.getArithmeticInstrCost(6264        I->getOpcode(), VectorTy, CostKind,6265        {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None},6266        Op2Info, Operands, I, TLI);6267  }6268  case Instruction::FNeg: {6269    return TTI.getArithmeticInstrCost(6270        I->getOpcode(), VectorTy, CostKind,6271        {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None},6272        {TargetTransformInfo::OK_AnyValue, TargetTransformInfo::OP_None},6273        I->getOperand(0), I);6274  }6275  case Instruction::Select: {6276    SelectInst *SI = cast<SelectInst>(I);6277    const SCEV *CondSCEV = SE->getSCEV(SI->getCondition());6278    bool ScalarCond = (SE->isLoopInvariant(CondSCEV, TheLoop));6279 6280    const Value *Op0, *Op1;6281    using namespace llvm::PatternMatch;6282    if (!ScalarCond && (match(I, m_LogicalAnd(m_Value(Op0), m_Value(Op1))) ||6283                        match(I, m_LogicalOr(m_Value(Op0), m_Value(Op1))))) {6284      // select x, y, false --> x & y6285      // select x, true, y --> x | y6286      const auto [Op1VK, Op1VP] = TTI::getOperandInfo(Op0);6287      const auto [Op2VK, Op2VP] = TTI::getOperandInfo(Op1);6288      assert(Op0->getType()->getScalarSizeInBits() == 1 &&6289              Op1->getType()->getScalarSizeInBits() == 1);6290 6291      return TTI.getArithmeticInstrCost(6292          match(I, m_LogicalOr()) ? Instruction::Or : Instruction::And,6293          VectorTy, CostKind, {Op1VK, Op1VP}, {Op2VK, Op2VP}, {Op0, Op1}, I);6294    }6295 6296    Type *CondTy = SI->getCondition()->getType();6297    if (!ScalarCond)6298      CondTy = VectorType::get(CondTy, VF);6299 6300    CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;6301    if (auto *Cmp = dyn_cast<CmpInst>(SI->getCondition()))6302      Pred = Cmp->getPredicate();6303    return TTI.getCmpSelInstrCost(I->getOpcode(), VectorTy, CondTy, Pred,6304                                  CostKind, {TTI::OK_AnyValue, TTI::OP_None},6305                                  {TTI::OK_AnyValue, TTI::OP_None}, I);6306  }6307  case Instruction::ICmp:6308  case Instruction::FCmp: {6309    Type *ValTy = I->getOperand(0)->getType();6310 6311    if (canTruncateToMinimalBitwidth(I, VF)) {6312      [[maybe_unused]] Instruction *Op0AsInstruction =6313          dyn_cast<Instruction>(I->getOperand(0));6314      assert((!canTruncateToMinimalBitwidth(Op0AsInstruction, VF) ||6315              MinBWs[I] == MinBWs[Op0AsInstruction]) &&6316             "if both the operand and the compare are marked for "6317             "truncation, they must have the same bitwidth");6318      ValTy = IntegerType::get(ValTy->getContext(), MinBWs[I]);6319    }6320 6321    VectorTy = toVectorTy(ValTy, VF);6322    return TTI.getCmpSelInstrCost(6323        I->getOpcode(), VectorTy, CmpInst::makeCmpResultType(VectorTy),6324        cast<CmpInst>(I)->getPredicate(), CostKind,6325        {TTI::OK_AnyValue, TTI::OP_None}, {TTI::OK_AnyValue, TTI::OP_None}, I);6326  }6327  case Instruction::Store:6328  case Instruction::Load: {6329    ElementCount Width = VF;6330    if (Width.isVector()) {6331      InstWidening Decision = getWideningDecision(I, Width);6332      assert(Decision != CM_Unknown &&6333             "CM decision should be taken at this point");6334      if (getWideningCost(I, VF) == InstructionCost::getInvalid())6335        return InstructionCost::getInvalid();6336      if (Decision == CM_Scalarize)6337        Width = ElementCount::getFixed(1);6338    }6339    VectorTy = toVectorTy(getLoadStoreType(I), Width);6340    return getMemoryInstructionCost(I, VF);6341  }6342  case Instruction::BitCast:6343    if (I->getType()->isPointerTy())6344      return 0;6345    [[fallthrough]];6346  case Instruction::ZExt:6347  case Instruction::SExt:6348  case Instruction::FPToUI:6349  case Instruction::FPToSI:6350  case Instruction::FPExt:6351  case Instruction::PtrToInt:6352  case Instruction::IntToPtr:6353  case Instruction::SIToFP:6354  case Instruction::UIToFP:6355  case Instruction::Trunc:6356  case Instruction::FPTrunc: {6357    // Computes the CastContextHint from a Load/Store instruction.6358    auto ComputeCCH = [&](Instruction *I) -> TTI::CastContextHint {6359      assert((isa<LoadInst>(I) || isa<StoreInst>(I)) &&6360             "Expected a load or a store!");6361 6362      if (VF.isScalar() || !TheLoop->contains(I))6363        return TTI::CastContextHint::Normal;6364 6365      switch (getWideningDecision(I, VF)) {6366      case LoopVectorizationCostModel::CM_GatherScatter:6367        return TTI::CastContextHint::GatherScatter;6368      case LoopVectorizationCostModel::CM_Interleave:6369        return TTI::CastContextHint::Interleave;6370      case LoopVectorizationCostModel::CM_Scalarize:6371      case LoopVectorizationCostModel::CM_Widen:6372        return isPredicatedInst(I) ? TTI::CastContextHint::Masked6373                                   : TTI::CastContextHint::Normal;6374      case LoopVectorizationCostModel::CM_Widen_Reverse:6375        return TTI::CastContextHint::Reversed;6376      case LoopVectorizationCostModel::CM_Unknown:6377        llvm_unreachable("Instr did not go through cost modelling?");6378      case LoopVectorizationCostModel::CM_VectorCall:6379      case LoopVectorizationCostModel::CM_IntrinsicCall:6380        llvm_unreachable_internal("Instr has invalid widening decision");6381      }6382 6383      llvm_unreachable("Unhandled case!");6384    };6385 6386    unsigned Opcode = I->getOpcode();6387    TTI::CastContextHint CCH = TTI::CastContextHint::None;6388    // For Trunc, the context is the only user, which must be a StoreInst.6389    if (Opcode == Instruction::Trunc || Opcode == Instruction::FPTrunc) {6390      if (I->hasOneUse())6391        if (StoreInst *Store = dyn_cast<StoreInst>(*I->user_begin()))6392          CCH = ComputeCCH(Store);6393    }6394    // For Z/Sext, the context is the operand, which must be a LoadInst.6395    else if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt ||6396             Opcode == Instruction::FPExt) {6397      if (LoadInst *Load = dyn_cast<LoadInst>(I->getOperand(0)))6398        CCH = ComputeCCH(Load);6399    }6400 6401    // We optimize the truncation of induction variables having constant6402    // integer steps. The cost of these truncations is the same as the scalar6403    // operation.6404    if (isOptimizableIVTruncate(I, VF)) {6405      auto *Trunc = cast<TruncInst>(I);6406      return TTI.getCastInstrCost(Instruction::Trunc, Trunc->getDestTy(),6407                                  Trunc->getSrcTy(), CCH, CostKind, Trunc);6408    }6409 6410    // Detect reduction patterns6411    if (auto RedCost = getReductionPatternCost(I, VF, VectorTy))6412      return *RedCost;6413 6414    Type *SrcScalarTy = I->getOperand(0)->getType();6415    Instruction *Op0AsInstruction = dyn_cast<Instruction>(I->getOperand(0));6416    if (canTruncateToMinimalBitwidth(Op0AsInstruction, VF))6417      SrcScalarTy =6418          IntegerType::get(SrcScalarTy->getContext(), MinBWs[Op0AsInstruction]);6419    Type *SrcVecTy =6420        VectorTy->isVectorTy() ? toVectorTy(SrcScalarTy, VF) : SrcScalarTy;6421 6422    if (canTruncateToMinimalBitwidth(I, VF)) {6423      // If the result type is <= the source type, there will be no extend6424      // after truncating the users to the minimal required bitwidth.6425      if (VectorTy->getScalarSizeInBits() <= SrcVecTy->getScalarSizeInBits() &&6426          (I->getOpcode() == Instruction::ZExt ||6427           I->getOpcode() == Instruction::SExt))6428        return 0;6429    }6430 6431    return TTI.getCastInstrCost(Opcode, VectorTy, SrcVecTy, CCH, CostKind, I);6432  }6433  case Instruction::Call:6434    return getVectorCallCost(cast<CallInst>(I), VF);6435  case Instruction::ExtractValue:6436    return TTI.getInstructionCost(I, CostKind);6437  case Instruction::Alloca:6438    // We cannot easily widen alloca to a scalable alloca, as6439    // the result would need to be a vector of pointers.6440    if (VF.isScalable())6441      return InstructionCost::getInvalid();6442    [[fallthrough]];6443  default:6444    // This opcode is unknown. Assume that it is the same as 'mul'.6445    return TTI.getArithmeticInstrCost(Instruction::Mul, VectorTy, CostKind);6446  } // end of switch.6447}6448 6449void LoopVectorizationCostModel::collectValuesToIgnore() {6450  // Ignore ephemeral values.6451  CodeMetrics::collectEphemeralValues(TheLoop, AC, ValuesToIgnore);6452 6453  SmallVector<Value *, 4> DeadInterleavePointerOps;6454  SmallVector<Value *, 4> DeadOps;6455 6456  // If a scalar epilogue is required, users outside the loop won't use6457  // live-outs from the vector loop but from the scalar epilogue. Ignore them if6458  // that is the case.6459  bool RequiresScalarEpilogue = requiresScalarEpilogue(true);6460  auto IsLiveOutDead = [this, RequiresScalarEpilogue](User *U) {6461    return RequiresScalarEpilogue &&6462           !TheLoop->contains(cast<Instruction>(U)->getParent());6463  };6464 6465  LoopBlocksDFS DFS(TheLoop);6466  DFS.perform(LI);6467  for (BasicBlock *BB : reverse(make_range(DFS.beginRPO(), DFS.endRPO())))6468    for (Instruction &I : reverse(*BB)) {6469      if (VecValuesToIgnore.contains(&I) || ValuesToIgnore.contains(&I))6470        continue;6471 6472      // Add instructions that would be trivially dead and are only used by6473      // values already ignored to DeadOps to seed worklist.6474      if (wouldInstructionBeTriviallyDead(&I, TLI) &&6475          all_of(I.users(), [this, IsLiveOutDead](User *U) {6476            return VecValuesToIgnore.contains(U) ||6477                   ValuesToIgnore.contains(U) || IsLiveOutDead(U);6478          }))6479        DeadOps.push_back(&I);6480 6481      // For interleave groups, we only create a pointer for the start of the6482      // interleave group. Queue up addresses of group members except the insert6483      // position for further processing.6484      if (isAccessInterleaved(&I)) {6485        auto *Group = getInterleavedAccessGroup(&I);6486        if (Group->getInsertPos() == &I)6487          continue;6488        Value *PointerOp = getLoadStorePointerOperand(&I);6489        DeadInterleavePointerOps.push_back(PointerOp);6490      }6491 6492      // Queue branches for analysis. They are dead, if their successors only6493      // contain dead instructions.6494      if (auto *Br = dyn_cast<BranchInst>(&I)) {6495        if (Br->isConditional())6496          DeadOps.push_back(&I);6497      }6498    }6499 6500  // Mark ops feeding interleave group members as free, if they are only used6501  // by other dead computations.6502  for (unsigned I = 0; I != DeadInterleavePointerOps.size(); ++I) {6503    auto *Op = dyn_cast<Instruction>(DeadInterleavePointerOps[I]);6504    if (!Op || !TheLoop->contains(Op) || any_of(Op->users(), [this](User *U) {6505          Instruction *UI = cast<Instruction>(U);6506          return !VecValuesToIgnore.contains(U) &&6507                 (!isAccessInterleaved(UI) ||6508                  getInterleavedAccessGroup(UI)->getInsertPos() == UI);6509        }))6510      continue;6511    VecValuesToIgnore.insert(Op);6512    append_range(DeadInterleavePointerOps, Op->operands());6513  }6514 6515  // Mark ops that would be trivially dead and are only used by ignored6516  // instructions as free.6517  BasicBlock *Header = TheLoop->getHeader();6518 6519  // Returns true if the block contains only dead instructions. Such blocks will6520  // be removed by VPlan-to-VPlan transforms and won't be considered by the6521  // VPlan-based cost model, so skip them in the legacy cost-model as well.6522  auto IsEmptyBlock = [this](BasicBlock *BB) {6523    return all_of(*BB, [this](Instruction &I) {6524      return ValuesToIgnore.contains(&I) || VecValuesToIgnore.contains(&I) ||6525             (isa<BranchInst>(&I) && !cast<BranchInst>(&I)->isConditional());6526    });6527  };6528  for (unsigned I = 0; I != DeadOps.size(); ++I) {6529    auto *Op = dyn_cast<Instruction>(DeadOps[I]);6530 6531    // Check if the branch should be considered dead.6532    if (auto *Br = dyn_cast_or_null<BranchInst>(Op)) {6533      BasicBlock *ThenBB = Br->getSuccessor(0);6534      BasicBlock *ElseBB = Br->getSuccessor(1);6535      // Don't considers branches leaving the loop for simplification.6536      if (!TheLoop->contains(ThenBB) || !TheLoop->contains(ElseBB))6537        continue;6538      bool ThenEmpty = IsEmptyBlock(ThenBB);6539      bool ElseEmpty = IsEmptyBlock(ElseBB);6540      if ((ThenEmpty && ElseEmpty) ||6541          (ThenEmpty && ThenBB->getSingleSuccessor() == ElseBB &&6542           ElseBB->phis().empty()) ||6543          (ElseEmpty && ElseBB->getSingleSuccessor() == ThenBB &&6544           ThenBB->phis().empty())) {6545        VecValuesToIgnore.insert(Br);6546        DeadOps.push_back(Br->getCondition());6547      }6548      continue;6549    }6550 6551    // Skip any op that shouldn't be considered dead.6552    if (!Op || !TheLoop->contains(Op) ||6553        (isa<PHINode>(Op) && Op->getParent() == Header) ||6554        !wouldInstructionBeTriviallyDead(Op, TLI) ||6555        any_of(Op->users(), [this, IsLiveOutDead](User *U) {6556          return !VecValuesToIgnore.contains(U) &&6557                 !ValuesToIgnore.contains(U) && !IsLiveOutDead(U);6558        }))6559      continue;6560 6561    // If all of Op's users are in ValuesToIgnore, add it to ValuesToIgnore6562    // which applies for both scalar and vector versions. Otherwise it is only6563    // dead in vector versions, so only add it to VecValuesToIgnore.6564    if (all_of(Op->users(),6565               [this](User *U) { return ValuesToIgnore.contains(U); }))6566      ValuesToIgnore.insert(Op);6567 6568    VecValuesToIgnore.insert(Op);6569    append_range(DeadOps, Op->operands());6570  }6571 6572  // Ignore type-promoting instructions we identified during reduction6573  // detection.6574  for (const auto &Reduction : Legal->getReductionVars()) {6575    const RecurrenceDescriptor &RedDes = Reduction.second;6576    const SmallPtrSetImpl<Instruction *> &Casts = RedDes.getCastInsts();6577    VecValuesToIgnore.insert_range(Casts);6578  }6579  // Ignore type-casting instructions we identified during induction6580  // detection.6581  for (const auto &Induction : Legal->getInductionVars()) {6582    const InductionDescriptor &IndDes = Induction.second;6583    VecValuesToIgnore.insert_range(IndDes.getCastInsts());6584  }6585}6586 6587void LoopVectorizationCostModel::collectInLoopReductions() {6588  // Avoid duplicating work finding in-loop reductions.6589  if (!InLoopReductions.empty())6590    return;6591 6592  for (const auto &Reduction : Legal->getReductionVars()) {6593    PHINode *Phi = Reduction.first;6594    const RecurrenceDescriptor &RdxDesc = Reduction.second;6595 6596    // Multi-use reductions (e.g., used in FindLastIV patterns) are handled6597    // separately and should not be considered for in-loop reductions.6598    if (RdxDesc.hasUsesOutsideReductionChain())6599      continue;6600 6601    // We don't collect reductions that are type promoted (yet).6602    if (RdxDesc.getRecurrenceType() != Phi->getType())6603      continue;6604 6605    // In-loop AnyOf and FindIV reductions are not yet supported.6606    RecurKind Kind = RdxDesc.getRecurrenceKind();6607    if (RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) ||6608        RecurrenceDescriptor::isFindIVRecurrenceKind(Kind))6609      continue;6610 6611    // If the target would prefer this reduction to happen "in-loop", then we6612    // want to record it as such.6613    if (!PreferInLoopReductions && !useOrderedReductions(RdxDesc) &&6614        !TTI.preferInLoopReduction(Kind, Phi->getType()))6615      continue;6616 6617    // Check that we can correctly put the reductions into the loop, by6618    // finding the chain of operations that leads from the phi to the loop6619    // exit value.6620    SmallVector<Instruction *, 4> ReductionOperations =6621        RdxDesc.getReductionOpChain(Phi, TheLoop);6622    bool InLoop = !ReductionOperations.empty();6623 6624    if (InLoop) {6625      InLoopReductions.insert(Phi);6626      // Add the elements to InLoopReductionImmediateChains for cost modelling.6627      Instruction *LastChain = Phi;6628      for (auto *I : ReductionOperations) {6629        InLoopReductionImmediateChains[I] = LastChain;6630        LastChain = I;6631      }6632    }6633    LLVM_DEBUG(dbgs() << "LV: Using " << (InLoop ? "inloop" : "out of loop")6634                      << " reduction for phi: " << *Phi << "\n");6635  }6636}6637 6638// This function will select a scalable VF if the target supports scalable6639// vectors and a fixed one otherwise.6640// TODO: we could return a pair of values that specify the max VF and6641// min VF, to be used in `buildVPlans(MinVF, MaxVF)` instead of6642// `buildVPlans(VF, VF)`. We cannot do it because VPLAN at the moment6643// doesn't have a cost model that can choose which plan to execute if6644// more than one is generated.6645static ElementCount determineVPlanVF(const TargetTransformInfo &TTI,6646                                     LoopVectorizationCostModel &CM) {6647  unsigned WidestType;6648  std::tie(std::ignore, WidestType) = CM.getSmallestAndWidestTypes();6649 6650  TargetTransformInfo::RegisterKind RegKind =6651      TTI.enableScalableVectorization()6652          ? TargetTransformInfo::RGK_ScalableVector6653          : TargetTransformInfo::RGK_FixedWidthVector;6654 6655  TypeSize RegSize = TTI.getRegisterBitWidth(RegKind);6656  unsigned N = RegSize.getKnownMinValue() / WidestType;6657  return ElementCount::get(N, RegSize.isScalable());6658}6659 6660VectorizationFactor6661LoopVectorizationPlanner::planInVPlanNativePath(ElementCount UserVF) {6662  ElementCount VF = UserVF;6663  // Outer loop handling: They may require CFG and instruction level6664  // transformations before even evaluating whether vectorization is profitable.6665  // Since we cannot modify the incoming IR, we need to build VPlan upfront in6666  // the vectorization pipeline.6667  if (!OrigLoop->isInnermost()) {6668    // If the user doesn't provide a vectorization factor, determine a6669    // reasonable one.6670    if (UserVF.isZero()) {6671      VF = determineVPlanVF(TTI, CM);6672      LLVM_DEBUG(dbgs() << "LV: VPlan computed VF " << VF << ".\n");6673 6674      // Make sure we have a VF > 1 for stress testing.6675      if (VPlanBuildStressTest && (VF.isScalar() || VF.isZero())) {6676        LLVM_DEBUG(dbgs() << "LV: VPlan stress testing: "6677                          << "overriding computed VF.\n");6678        VF = ElementCount::getFixed(4);6679      }6680    } else if (UserVF.isScalable() && !TTI.supportsScalableVectors() &&6681               !ForceTargetSupportsScalableVectors) {6682      LLVM_DEBUG(dbgs() << "LV: Not vectorizing. Scalable VF requested, but "6683                        << "not supported by the target.\n");6684      reportVectorizationFailure(6685          "Scalable vectorization requested but not supported by the target",6686          "the scalable user-specified vectorization width for outer-loop "6687          "vectorization cannot be used because the target does not support "6688          "scalable vectors.",6689          "ScalableVFUnfeasible", ORE, OrigLoop);6690      return VectorizationFactor::Disabled();6691    }6692    assert(EnableVPlanNativePath && "VPlan-native path is not enabled.");6693    assert(isPowerOf2_32(VF.getKnownMinValue()) &&6694           "VF needs to be a power of two");6695    LLVM_DEBUG(dbgs() << "LV: Using " << (!UserVF.isZero() ? "user " : "")6696                      << "VF " << VF << " to build VPlans.\n");6697    buildVPlans(VF, VF);6698 6699    if (VPlans.empty())6700      return VectorizationFactor::Disabled();6701 6702    // For VPlan build stress testing, we bail out after VPlan construction.6703    if (VPlanBuildStressTest)6704      return VectorizationFactor::Disabled();6705 6706    return {VF, 0 /*Cost*/, 0 /* ScalarCost */};6707  }6708 6709  LLVM_DEBUG(6710      dbgs() << "LV: Not vectorizing. Inner loops aren't supported in the "6711                "VPlan-native path.\n");6712  return VectorizationFactor::Disabled();6713}6714 6715void LoopVectorizationPlanner::plan(ElementCount UserVF, unsigned UserIC) {6716  assert(OrigLoop->isInnermost() && "Inner loop expected.");6717  CM.collectValuesToIgnore();6718  CM.collectElementTypesForWidening();6719 6720  FixedScalableVFPair MaxFactors = CM.computeMaxVF(UserVF, UserIC);6721  if (!MaxFactors) // Cases that should not to be vectorized nor interleaved.6722    return;6723 6724  // Invalidate interleave groups if all blocks of loop will be predicated.6725  if (CM.blockNeedsPredicationForAnyReason(OrigLoop->getHeader()) &&6726      !useMaskedInterleavedAccesses(TTI)) {6727    LLVM_DEBUG(6728        dbgs()6729        << "LV: Invalidate all interleaved groups due to fold-tail by masking "6730           "which requires masked-interleaved support.\n");6731    if (CM.InterleaveInfo.invalidateGroups())6732      // Invalidating interleave groups also requires invalidating all decisions6733      // based on them, which includes widening decisions and uniform and scalar6734      // values.6735      CM.invalidateCostModelingDecisions();6736  }6737 6738  if (CM.foldTailByMasking())6739    Legal->prepareToFoldTailByMasking();6740 6741  ElementCount MaxUserVF =6742      UserVF.isScalable() ? MaxFactors.ScalableVF : MaxFactors.FixedVF;6743  if (UserVF) {6744    if (!ElementCount::isKnownLE(UserVF, MaxUserVF)) {6745      reportVectorizationInfo(6746          "UserVF ignored because it may be larger than the maximal safe VF",6747          "InvalidUserVF", ORE, OrigLoop);6748    } else {6749      assert(isPowerOf2_32(UserVF.getKnownMinValue()) &&6750             "VF needs to be a power of two");6751      // Collect the instructions (and their associated costs) that will be more6752      // profitable to scalarize.6753      CM.collectInLoopReductions();6754      if (CM.selectUserVectorizationFactor(UserVF)) {6755        LLVM_DEBUG(dbgs() << "LV: Using user VF " << UserVF << ".\n");6756        buildVPlansWithVPRecipes(UserVF, UserVF);6757        LLVM_DEBUG(printPlans(dbgs()));6758        return;6759      }6760      reportVectorizationInfo("UserVF ignored because of invalid costs.",6761                              "InvalidCost", ORE, OrigLoop);6762    }6763  }6764 6765  // Collect the Vectorization Factor Candidates.6766  SmallVector<ElementCount> VFCandidates;6767  for (auto VF = ElementCount::getFixed(1);6768       ElementCount::isKnownLE(VF, MaxFactors.FixedVF); VF *= 2)6769    VFCandidates.push_back(VF);6770  for (auto VF = ElementCount::getScalable(1);6771       ElementCount::isKnownLE(VF, MaxFactors.ScalableVF); VF *= 2)6772    VFCandidates.push_back(VF);6773 6774  CM.collectInLoopReductions();6775  for (const auto &VF : VFCandidates) {6776    // Collect Uniform and Scalar instructions after vectorization with VF.6777    CM.collectNonVectorizedAndSetWideningDecisions(VF);6778  }6779 6780  buildVPlansWithVPRecipes(ElementCount::getFixed(1), MaxFactors.FixedVF);6781  buildVPlansWithVPRecipes(ElementCount::getScalable(1), MaxFactors.ScalableVF);6782 6783  LLVM_DEBUG(printPlans(dbgs()));6784}6785 6786InstructionCost VPCostContext::getLegacyCost(Instruction *UI,6787                                             ElementCount VF) const {6788  InstructionCost Cost = CM.getInstructionCost(UI, VF);6789  if (Cost.isValid() && ForceTargetInstructionCost.getNumOccurrences())6790    return InstructionCost(ForceTargetInstructionCost);6791  return Cost;6792}6793 6794bool VPCostContext::isLegacyUniformAfterVectorization(Instruction *I,6795                                                      ElementCount VF) const {6796  return CM.isUniformAfterVectorization(I, VF);6797}6798 6799bool VPCostContext::skipCostComputation(Instruction *UI, bool IsVector) const {6800  return CM.ValuesToIgnore.contains(UI) ||6801         (IsVector && CM.VecValuesToIgnore.contains(UI)) ||6802         SkipCostComputation.contains(UI);6803}6804 6805unsigned VPCostContext::getPredBlockCostDivisor(BasicBlock *BB) const {6806  return CM.getPredBlockCostDivisor(CostKind, BB);6807}6808 6809InstructionCost6810LoopVectorizationPlanner::precomputeCosts(VPlan &Plan, ElementCount VF,6811                                          VPCostContext &CostCtx) const {6812  InstructionCost Cost;6813  // Cost modeling for inductions is inaccurate in the legacy cost model6814  // compared to the recipes that are generated. To match here initially during6815  // VPlan cost model bring up directly use the induction costs from the legacy6816  // cost model. Note that we do this as pre-processing; the VPlan may not have6817  // any recipes associated with the original induction increment instruction6818  // and may replace truncates with VPWidenIntOrFpInductionRecipe. We precompute6819  // the cost of induction phis and increments (both that are represented by6820  // recipes and those that are not), to avoid distinguishing between them here,6821  // and skip all recipes that represent induction phis and increments (the6822  // former case) later on, if they exist, to avoid counting them twice.6823  // Similarly we pre-compute the cost of any optimized truncates.6824  // TODO: Switch to more accurate costing based on VPlan.6825  for (const auto &[IV, IndDesc] : Legal->getInductionVars()) {6826    Instruction *IVInc = cast<Instruction>(6827        IV->getIncomingValueForBlock(OrigLoop->getLoopLatch()));6828    SmallVector<Instruction *> IVInsts = {IVInc};6829    for (unsigned I = 0; I != IVInsts.size(); I++) {6830      for (Value *Op : IVInsts[I]->operands()) {6831        auto *OpI = dyn_cast<Instruction>(Op);6832        if (Op == IV || !OpI || !OrigLoop->contains(OpI) || !Op->hasOneUse())6833          continue;6834        IVInsts.push_back(OpI);6835      }6836    }6837    IVInsts.push_back(IV);6838    for (User *U : IV->users()) {6839      auto *CI = cast<Instruction>(U);6840      if (!CostCtx.CM.isOptimizableIVTruncate(CI, VF))6841        continue;6842      IVInsts.push_back(CI);6843    }6844 6845    // If the vector loop gets executed exactly once with the given VF, ignore6846    // the costs of comparison and induction instructions, as they'll get6847    // simplified away.6848    // TODO: Remove this code after stepping away from the legacy cost model and6849    // adding code to simplify VPlans before calculating their costs.6850    auto TC = getSmallConstantTripCount(PSE.getSE(), OrigLoop);6851    if (TC == VF && !CM.foldTailByMasking())6852      addFullyUnrolledInstructionsToIgnore(OrigLoop, Legal->getInductionVars(),6853                                           CostCtx.SkipCostComputation);6854 6855    for (Instruction *IVInst : IVInsts) {6856      if (CostCtx.skipCostComputation(IVInst, VF.isVector()))6857        continue;6858      InstructionCost InductionCost = CostCtx.getLegacyCost(IVInst, VF);6859      LLVM_DEBUG({6860        dbgs() << "Cost of " << InductionCost << " for VF " << VF6861               << ": induction instruction " << *IVInst << "\n";6862      });6863      Cost += InductionCost;6864      CostCtx.SkipCostComputation.insert(IVInst);6865    }6866  }6867 6868  /// Compute the cost of all exiting conditions of the loop using the legacy6869  /// cost model. This is to match the legacy behavior, which adds the cost of6870  /// all exit conditions. Note that this over-estimates the cost, as there will6871  /// be a single condition to control the vector loop.6872  SmallVector<BasicBlock *> Exiting;6873  CM.TheLoop->getExitingBlocks(Exiting);6874  SetVector<Instruction *> ExitInstrs;6875  // Collect all exit conditions.6876  for (BasicBlock *EB : Exiting) {6877    auto *Term = dyn_cast<BranchInst>(EB->getTerminator());6878    if (!Term || CostCtx.skipCostComputation(Term, VF.isVector()))6879      continue;6880    if (auto *CondI = dyn_cast<Instruction>(Term->getOperand(0))) {6881      ExitInstrs.insert(CondI);6882    }6883  }6884  // Compute the cost of all instructions only feeding the exit conditions.6885  for (unsigned I = 0; I != ExitInstrs.size(); ++I) {6886    Instruction *CondI = ExitInstrs[I];6887    if (!OrigLoop->contains(CondI) ||6888        !CostCtx.SkipCostComputation.insert(CondI).second)6889      continue;6890    InstructionCost CondICost = CostCtx.getLegacyCost(CondI, VF);6891    LLVM_DEBUG({6892      dbgs() << "Cost of " << CondICost << " for VF " << VF6893             << ": exit condition instruction " << *CondI << "\n";6894    });6895    Cost += CondICost;6896    for (Value *Op : CondI->operands()) {6897      auto *OpI = dyn_cast<Instruction>(Op);6898      if (!OpI || CostCtx.skipCostComputation(OpI, VF.isVector()) ||6899          any_of(OpI->users(), [&ExitInstrs, this](User *U) {6900            return OrigLoop->contains(cast<Instruction>(U)->getParent()) &&6901                   !ExitInstrs.contains(cast<Instruction>(U));6902          }))6903        continue;6904      ExitInstrs.insert(OpI);6905    }6906  }6907 6908  // Pre-compute the costs for branches except for the backedge, as the number6909  // of replicate regions in a VPlan may not directly match the number of6910  // branches, which would lead to different decisions.6911  // TODO: Compute cost of branches for each replicate region in the VPlan,6912  // which is more accurate than the legacy cost model.6913  for (BasicBlock *BB : OrigLoop->blocks()) {6914    if (CostCtx.skipCostComputation(BB->getTerminator(), VF.isVector()))6915      continue;6916    CostCtx.SkipCostComputation.insert(BB->getTerminator());6917    if (BB == OrigLoop->getLoopLatch())6918      continue;6919    auto BranchCost = CostCtx.getLegacyCost(BB->getTerminator(), VF);6920    Cost += BranchCost;6921  }6922 6923  // Pre-compute costs for instructions that are forced-scalar or profitable to6924  // scalarize. Their costs will be computed separately in the legacy cost6925  // model.6926  for (Instruction *ForcedScalar : CM.ForcedScalars[VF]) {6927    if (CostCtx.skipCostComputation(ForcedScalar, VF.isVector()))6928      continue;6929    CostCtx.SkipCostComputation.insert(ForcedScalar);6930    InstructionCost ForcedCost = CostCtx.getLegacyCost(ForcedScalar, VF);6931    LLVM_DEBUG({6932      dbgs() << "Cost of " << ForcedCost << " for VF " << VF6933             << ": forced scalar " << *ForcedScalar << "\n";6934    });6935    Cost += ForcedCost;6936  }6937  for (const auto &[Scalarized, ScalarCost] : CM.InstsToScalarize[VF]) {6938    if (CostCtx.skipCostComputation(Scalarized, VF.isVector()))6939      continue;6940    CostCtx.SkipCostComputation.insert(Scalarized);6941    LLVM_DEBUG({6942      dbgs() << "Cost of " << ScalarCost << " for VF " << VF6943             << ": profitable to scalarize " << *Scalarized << "\n";6944    });6945    Cost += ScalarCost;6946  }6947 6948  return Cost;6949}6950 6951InstructionCost LoopVectorizationPlanner::cost(VPlan &Plan,6952                                               ElementCount VF) const {6953  VPCostContext CostCtx(CM.TTI, *CM.TLI, Plan, CM, CM.CostKind, *PSE.getSE(),6954                        OrigLoop);6955  InstructionCost Cost = precomputeCosts(Plan, VF, CostCtx);6956 6957  // Now compute and add the VPlan-based cost.6958  Cost += Plan.cost(VF, CostCtx);6959#ifndef NDEBUG6960  unsigned EstimatedWidth = estimateElementCount(VF, CM.getVScaleForTuning());6961  LLVM_DEBUG(dbgs() << "Cost for VF " << VF << ": " << Cost6962                    << " (Estimated cost per lane: ");6963  if (Cost.isValid()) {6964    double CostPerLane = double(Cost.getValue()) / EstimatedWidth;6965    LLVM_DEBUG(dbgs() << format("%.1f", CostPerLane));6966  } else /* No point dividing an invalid cost - it will still be invalid */6967    LLVM_DEBUG(dbgs() << "Invalid");6968  LLVM_DEBUG(dbgs() << ")\n");6969#endif6970  return Cost;6971}6972 6973#ifndef NDEBUG6974/// Return true if the original loop \ TheLoop contains any instructions that do6975/// not have corresponding recipes in \p Plan and are not marked to be ignored6976/// in \p CostCtx. This means the VPlan contains simplification that the legacy6977/// cost-model did not account for.6978static bool planContainsAdditionalSimplifications(VPlan &Plan,6979                                                  VPCostContext &CostCtx,6980                                                  Loop *TheLoop,6981                                                  ElementCount VF) {6982  // First collect all instructions for the recipes in Plan.6983  auto GetInstructionForCost = [](const VPRecipeBase *R) -> Instruction * {6984    if (auto *S = dyn_cast<VPSingleDefRecipe>(R))6985      return dyn_cast_or_null<Instruction>(S->getUnderlyingValue());6986    if (auto *WidenMem = dyn_cast<VPWidenMemoryRecipe>(R))6987      return &WidenMem->getIngredient();6988    return nullptr;6989  };6990 6991  // Check if a select for a safe divisor was hoisted to the pre-header. If so,6992  // the select doesn't need to be considered for the vector loop cost; go with6993  // the more accurate VPlan-based cost model.6994  for (VPRecipeBase &R : *Plan.getVectorPreheader()) {6995    auto *VPI = dyn_cast<VPInstruction>(&R);6996    if (!VPI || VPI->getOpcode() != Instruction::Select)6997      continue;6998 6999    if (auto *WR = dyn_cast_or_null<VPWidenRecipe>(VPI->getSingleUser())) {7000      switch (WR->getOpcode()) {7001      case Instruction::UDiv:7002      case Instruction::SDiv:7003      case Instruction::URem:7004      case Instruction::SRem:7005        return true;7006      default:7007        break;7008      }7009    }7010  }7011 7012  DenseSet<Instruction *> SeenInstrs;7013  auto Iter = vp_depth_first_deep(Plan.getVectorLoopRegion()->getEntry());7014  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(Iter)) {7015    for (VPRecipeBase &R : *VPBB) {7016      if (auto *IR = dyn_cast<VPInterleaveRecipe>(&R)) {7017        auto *IG = IR->getInterleaveGroup();7018        unsigned NumMembers = IG->getNumMembers();7019        for (unsigned I = 0; I != NumMembers; ++I) {7020          if (Instruction *M = IG->getMember(I))7021            SeenInstrs.insert(M);7022        }7023        continue;7024      }7025      // Unused FOR splices are removed by VPlan transforms, so the VPlan-based7026      // cost model won't cost it whilst the legacy will.7027      if (auto *FOR = dyn_cast<VPFirstOrderRecurrencePHIRecipe>(&R)) {7028        using namespace VPlanPatternMatch;7029        if (none_of(FOR->users(),7030                    match_fn(m_VPInstruction<7031                             VPInstruction::FirstOrderRecurrenceSplice>())))7032          return true;7033      }7034      // The VPlan-based cost model is more accurate for partial reductions and7035      // comparing against the legacy cost isn't desirable.7036      if (auto *VPR = dyn_cast<VPReductionRecipe>(&R))7037        if (VPR->isPartialReduction())7038          return true;7039 7040      // The VPlan-based cost model can analyze if recipes are scalar7041      // recursively, but the legacy cost model cannot.7042      if (auto *WidenMemR = dyn_cast<VPWidenMemoryRecipe>(&R)) {7043        auto *AddrI = dyn_cast<Instruction>(7044            getLoadStorePointerOperand(&WidenMemR->getIngredient()));7045        if (AddrI && vputils::isSingleScalar(WidenMemR->getAddr()) !=7046                         CostCtx.isLegacyUniformAfterVectorization(AddrI, VF))7047          return true;7048      }7049 7050      /// If a VPlan transform folded a recipe to one producing a single-scalar,7051      /// but the original instruction wasn't uniform-after-vectorization in the7052      /// legacy cost model, the legacy cost overestimates the actual cost.7053      if (auto *RepR = dyn_cast<VPReplicateRecipe>(&R)) {7054        if (RepR->isSingleScalar() &&7055            !CostCtx.isLegacyUniformAfterVectorization(7056                RepR->getUnderlyingInstr(), VF))7057          return true;7058      }7059      if (Instruction *UI = GetInstructionForCost(&R)) {7060        // If we adjusted the predicate of the recipe, the cost in the legacy7061        // cost model may be different.7062        using namespace VPlanPatternMatch;7063        CmpPredicate Pred;7064        if (match(&R, m_Cmp(Pred, m_VPValue(), m_VPValue())) &&7065            cast<VPRecipeWithIRFlags>(R).getPredicate() !=7066                cast<CmpInst>(UI)->getPredicate())7067          return true;7068        SeenInstrs.insert(UI);7069      }7070    }7071  }7072 7073  // Return true if the loop contains any instructions that are not also part of7074  // the VPlan or are skipped for VPlan-based cost computations. This indicates7075  // that the VPlan contains extra simplifications.7076  return any_of(TheLoop->blocks(), [&SeenInstrs, &CostCtx,7077                                    TheLoop](BasicBlock *BB) {7078    return any_of(*BB, [&SeenInstrs, &CostCtx, TheLoop, BB](Instruction &I) {7079      // Skip induction phis when checking for simplifications, as they may not7080      // be lowered directly be lowered to a corresponding PHI recipe.7081      if (isa<PHINode>(&I) && BB == TheLoop->getHeader() &&7082          CostCtx.CM.Legal->isInductionPhi(cast<PHINode>(&I)))7083        return false;7084      return !SeenInstrs.contains(&I) && !CostCtx.skipCostComputation(&I, true);7085    });7086  });7087}7088#endif7089 7090VectorizationFactor LoopVectorizationPlanner::computeBestVF() {7091  if (VPlans.empty())7092    return VectorizationFactor::Disabled();7093  // If there is a single VPlan with a single VF, return it directly.7094  VPlan &FirstPlan = *VPlans[0];7095  if (VPlans.size() == 1 && size(FirstPlan.vectorFactors()) == 1)7096    return {*FirstPlan.vectorFactors().begin(), 0, 0};7097 7098  LLVM_DEBUG(dbgs() << "LV: Computing best VF using cost kind: "7099                    << (CM.CostKind == TTI::TCK_RecipThroughput7100                            ? "Reciprocal Throughput\n"7101                        : CM.CostKind == TTI::TCK_Latency7102                            ? "Instruction Latency\n"7103                        : CM.CostKind == TTI::TCK_CodeSize ? "Code Size\n"7104                        : CM.CostKind == TTI::TCK_SizeAndLatency7105                            ? "Code Size and Latency\n"7106                            : "Unknown\n"));7107 7108  ElementCount ScalarVF = ElementCount::getFixed(1);7109  assert(hasPlanWithVF(ScalarVF) &&7110         "More than a single plan/VF w/o any plan having scalar VF");7111 7112  // TODO: Compute scalar cost using VPlan-based cost model.7113  InstructionCost ScalarCost = CM.expectedCost(ScalarVF);7114  LLVM_DEBUG(dbgs() << "LV: Scalar loop costs: " << ScalarCost << ".\n");7115  VectorizationFactor ScalarFactor(ScalarVF, ScalarCost, ScalarCost);7116  VectorizationFactor BestFactor = ScalarFactor;7117 7118  bool ForceVectorization = Hints.getForce() == LoopVectorizeHints::FK_Enabled;7119  if (ForceVectorization) {7120    // Ignore scalar width, because the user explicitly wants vectorization.7121    // Initialize cost to max so that VF = 2 is, at least, chosen during cost7122    // evaluation.7123    BestFactor.Cost = InstructionCost::getMax();7124  }7125 7126  for (auto &P : VPlans) {7127    ArrayRef<ElementCount> VFs(P->vectorFactors().begin(),7128                               P->vectorFactors().end());7129 7130    SmallVector<VPRegisterUsage, 8> RUs;7131    if (any_of(VFs, [this](ElementCount VF) {7132          return CM.shouldConsiderRegPressureForVF(VF);7133        }))7134      RUs = calculateRegisterUsageForPlan(*P, VFs, TTI, CM.ValuesToIgnore);7135 7136    for (unsigned I = 0; I < VFs.size(); I++) {7137      ElementCount VF = VFs[I];7138      if (VF.isScalar())7139        continue;7140      if (!ForceVectorization && !willGenerateVectors(*P, VF, TTI)) {7141        LLVM_DEBUG(7142            dbgs()7143            << "LV: Not considering vector loop of width " << VF7144            << " because it will not generate any vector instructions.\n");7145        continue;7146      }7147      if (CM.OptForSize && !ForceVectorization && hasReplicatorRegion(*P)) {7148        LLVM_DEBUG(7149            dbgs()7150            << "LV: Not considering vector loop of width " << VF7151            << " because it would cause replicated blocks to be generated,"7152            << " which isn't allowed when optimizing for size.\n");7153        continue;7154      }7155 7156      InstructionCost Cost = cost(*P, VF);7157      VectorizationFactor CurrentFactor(VF, Cost, ScalarCost);7158 7159      if (CM.shouldConsiderRegPressureForVF(VF) &&7160          RUs[I].exceedsMaxNumRegs(TTI, ForceTargetNumVectorRegs)) {7161        LLVM_DEBUG(dbgs() << "LV(REG): Not considering vector loop of width "7162                          << VF << " because it uses too many registers\n");7163        continue;7164      }7165 7166      if (isMoreProfitable(CurrentFactor, BestFactor, P->hasScalarTail()))7167        BestFactor = CurrentFactor;7168 7169      // If profitable add it to ProfitableVF list.7170      if (isMoreProfitable(CurrentFactor, ScalarFactor, P->hasScalarTail()))7171        ProfitableVFs.push_back(CurrentFactor);7172    }7173  }7174 7175#ifndef NDEBUG7176  // Select the optimal vectorization factor according to the legacy cost-model.7177  // This is now only used to verify the decisions by the new VPlan-based7178  // cost-model and will be retired once the VPlan-based cost-model is7179  // stabilized.7180  VectorizationFactor LegacyVF = selectVectorizationFactor();7181  VPlan &BestPlan = getPlanFor(BestFactor.Width);7182 7183  // Pre-compute the cost and use it to check if BestPlan contains any7184  // simplifications not accounted for in the legacy cost model. If that's the7185  // case, don't trigger the assertion, as the extra simplifications may cause a7186  // different VF to be picked by the VPlan-based cost model.7187  VPCostContext CostCtx(CM.TTI, *CM.TLI, BestPlan, CM, CM.CostKind,7188                        *CM.PSE.getSE(), OrigLoop);7189  precomputeCosts(BestPlan, BestFactor.Width, CostCtx);7190  // Verify that the VPlan-based and legacy cost models agree, except for7191  // * VPlans with early exits,7192  // * VPlans with additional VPlan simplifications,7193  // * EVL-based VPlans with gather/scatters (the VPlan-based cost model uses7194  //   vp_scatter/vp_gather).7195  // The legacy cost model doesn't properly model costs for such loops.7196  bool UsesEVLGatherScatter =7197      any_of(VPBlockUtils::blocksOnly<VPBasicBlock>(vp_depth_first_shallow(7198                 BestPlan.getVectorLoopRegion()->getEntry())),7199             [](VPBasicBlock *VPBB) {7200               return any_of(*VPBB, [](VPRecipeBase &R) {7201                 return isa<VPWidenLoadEVLRecipe, VPWidenStoreEVLRecipe>(&R) &&7202                        !cast<VPWidenMemoryRecipe>(&R)->isConsecutive();7203               });7204             });7205  assert(7206      (BestFactor.Width == LegacyVF.Width || BestPlan.hasEarlyExit() ||7207       !Legal->getLAI()->getSymbolicStrides().empty() || UsesEVLGatherScatter ||7208       planContainsAdditionalSimplifications(7209           getPlanFor(BestFactor.Width), CostCtx, OrigLoop, BestFactor.Width) ||7210       planContainsAdditionalSimplifications(7211           getPlanFor(LegacyVF.Width), CostCtx, OrigLoop, LegacyVF.Width)) &&7212      " VPlan cost model and legacy cost model disagreed");7213  assert((BestFactor.Width.isScalar() || BestFactor.ScalarCost > 0) &&7214         "when vectorizing, the scalar cost must be computed.");7215#endif7216 7217  LLVM_DEBUG(dbgs() << "LV: Selecting VF: " << BestFactor.Width << ".\n");7218  return BestFactor;7219}7220 7221static Value *getStartValueFromReductionResult(VPInstruction *RdxResult) {7222  using namespace VPlanPatternMatch;7223  assert(RdxResult->getOpcode() == VPInstruction::ComputeFindIVResult &&7224         "RdxResult must be ComputeFindIVResult");7225  VPValue *StartVPV = RdxResult->getOperand(1);7226  match(StartVPV, m_Freeze(m_VPValue(StartVPV)));7227  return StartVPV->getLiveInIRValue();7228}7229 7230// If \p EpiResumePhiR is resume VPPhi for a reduction when vectorizing the7231// epilog loop, fix the reduction's scalar PHI node by adding the incoming value7232// from the main vector loop.7233static void fixReductionScalarResumeWhenVectorizingEpilog(7234    VPPhi *EpiResumePhiR, PHINode &EpiResumePhi, BasicBlock *BypassBlock) {7235  // Get the VPInstruction computing the reduction result in the middle block.7236  // The first operand may not be from the middle block if it is not connected7237  // to the scalar preheader. In that case, there's nothing to fix.7238  VPValue *Incoming = EpiResumePhiR->getOperand(0);7239  match(Incoming, VPlanPatternMatch::m_ZExtOrSExt(7240                      VPlanPatternMatch::m_VPValue(Incoming)));7241  auto *EpiRedResult = dyn_cast<VPInstruction>(Incoming);7242  if (!EpiRedResult ||7243      (EpiRedResult->getOpcode() != VPInstruction::ComputeAnyOfResult &&7244       EpiRedResult->getOpcode() != VPInstruction::ComputeReductionResult &&7245       EpiRedResult->getOpcode() != VPInstruction::ComputeFindIVResult))7246    return;7247 7248  auto *EpiRedHeaderPhi =7249      cast<VPReductionPHIRecipe>(EpiRedResult->getOperand(0));7250  RecurKind Kind = EpiRedHeaderPhi->getRecurrenceKind();7251  Value *MainResumeValue;7252  if (auto *VPI = dyn_cast<VPInstruction>(EpiRedHeaderPhi->getStartValue())) {7253    assert((VPI->getOpcode() == VPInstruction::Broadcast ||7254            VPI->getOpcode() == VPInstruction::ReductionStartVector) &&7255           "unexpected start recipe");7256    MainResumeValue = VPI->getOperand(0)->getUnderlyingValue();7257  } else7258    MainResumeValue = EpiRedHeaderPhi->getStartValue()->getUnderlyingValue();7259  if (RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind)) {7260    [[maybe_unused]] Value *StartV =7261        EpiRedResult->getOperand(1)->getLiveInIRValue();7262    auto *Cmp = cast<ICmpInst>(MainResumeValue);7263    assert(Cmp->getPredicate() == CmpInst::ICMP_NE &&7264           "AnyOf expected to start with ICMP_NE");7265    assert(Cmp->getOperand(1) == StartV &&7266           "AnyOf expected to start by comparing main resume value to original "7267           "start value");7268    MainResumeValue = Cmp->getOperand(0);7269  } else if (RecurrenceDescriptor::isFindIVRecurrenceKind(Kind)) {7270    Value *StartV = getStartValueFromReductionResult(EpiRedResult);7271    Value *SentinelV = EpiRedResult->getOperand(2)->getLiveInIRValue();7272    using namespace llvm::PatternMatch;7273    Value *Cmp, *OrigResumeV, *CmpOp;7274    [[maybe_unused]] bool IsExpectedPattern =7275        match(MainResumeValue,7276              m_Select(m_OneUse(m_Value(Cmp)), m_Specific(SentinelV),7277                       m_Value(OrigResumeV))) &&7278        (match(Cmp, m_SpecificICmp(ICmpInst::ICMP_EQ, m_Specific(OrigResumeV),7279                                   m_Value(CmpOp))) &&7280         ((CmpOp == StartV && isGuaranteedNotToBeUndefOrPoison(CmpOp))));7281    assert(IsExpectedPattern && "Unexpected reduction resume pattern");7282    MainResumeValue = OrigResumeV;7283  }7284  PHINode *MainResumePhi = cast<PHINode>(MainResumeValue);7285 7286  // When fixing reductions in the epilogue loop we should already have7287  // created a bc.merge.rdx Phi after the main vector body. Ensure that we carry7288  // over the incoming values correctly.7289  EpiResumePhi.setIncomingValueForBlock(7290      BypassBlock, MainResumePhi->getIncomingValueForBlock(BypassBlock));7291}7292 7293DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(7294    ElementCount BestVF, unsigned BestUF, VPlan &BestVPlan,7295    InnerLoopVectorizer &ILV, DominatorTree *DT, bool VectorizingEpilogue) {7296  assert(BestVPlan.hasVF(BestVF) &&7297         "Trying to execute plan with unsupported VF");7298  assert(BestVPlan.hasUF(BestUF) &&7299         "Trying to execute plan with unsupported UF");7300  if (BestVPlan.hasEarlyExit())7301    ++LoopsEarlyExitVectorized;7302  // TODO: Move to VPlan transform stage once the transition to the VPlan-based7303  // cost model is complete for better cost estimates.7304  VPlanTransforms::runPass(VPlanTransforms::unrollByUF, BestVPlan, BestUF);7305  VPlanTransforms::runPass(VPlanTransforms::materializePacksAndUnpacks,7306                           BestVPlan);7307  VPlanTransforms::runPass(VPlanTransforms::materializeBroadcasts, BestVPlan);7308  VPlanTransforms::runPass(VPlanTransforms::replicateByVF, BestVPlan, BestVF);7309  bool HasBranchWeights =7310      hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator());7311  if (HasBranchWeights) {7312    std::optional<unsigned> VScale = CM.getVScaleForTuning();7313    VPlanTransforms::runPass(VPlanTransforms::addBranchWeightToMiddleTerminator,7314                             BestVPlan, BestVF, VScale);7315  }7316 7317  // Checks are the same for all VPlans, added to BestVPlan only for7318  // compactness.7319  attachRuntimeChecks(BestVPlan, ILV.RTChecks, HasBranchWeights);7320 7321  // Retrieving VectorPH now when it's easier while VPlan still has Regions.7322  VPBasicBlock *VectorPH = cast<VPBasicBlock>(BestVPlan.getVectorPreheader());7323 7324  VPlanTransforms::optimizeForVFAndUF(BestVPlan, BestVF, BestUF, PSE);7325  VPlanTransforms::simplifyRecipes(BestVPlan);7326  VPlanTransforms::removeBranchOnConst(BestVPlan);7327  if (BestVPlan.getEntry()->getSingleSuccessor() ==7328      BestVPlan.getScalarPreheader()) {7329    // TODO: The vector loop would be dead, should not even try to vectorize.7330    ORE->emit([&]() {7331      return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationDead",7332                                        OrigLoop->getStartLoc(),7333                                        OrigLoop->getHeader())7334             << "Created vector loop never executes due to insufficient trip "7335                "count.";7336    });7337    return DenseMap<const SCEV *, Value *>();7338  }7339 7340  VPlanTransforms::narrowInterleaveGroups(7341      BestVPlan, BestVF,7342      TTI.getRegisterBitWidth(BestVF.isScalable()7343                                  ? TargetTransformInfo::RGK_ScalableVector7344                                  : TargetTransformInfo::RGK_FixedWidthVector));7345  VPlanTransforms::removeDeadRecipes(BestVPlan);7346 7347  VPlanTransforms::convertToConcreteRecipes(BestVPlan);7348  // Regions are dissolved after optimizing for VF and UF, which completely7349  // removes unneeded loop regions first.7350  VPlanTransforms::dissolveLoopRegions(BestVPlan);7351  // Canonicalize EVL loops after regions are dissolved.7352  VPlanTransforms::canonicalizeEVLLoops(BestVPlan);7353  VPlanTransforms::materializeBackedgeTakenCount(BestVPlan, VectorPH);7354  VPlanTransforms::materializeVectorTripCount(7355      BestVPlan, VectorPH, CM.foldTailByMasking(),7356      CM.requiresScalarEpilogue(BestVF.isVector()));7357  VPlanTransforms::materializeVFAndVFxUF(BestVPlan, VectorPH, BestVF);7358  VPlanTransforms::cse(BestVPlan);7359  VPlanTransforms::simplifyRecipes(BestVPlan);7360 7361  // 0. Generate SCEV-dependent code in the entry, including TripCount, before7362  // making any changes to the CFG.7363  DenseMap<const SCEV *, Value *> ExpandedSCEVs =7364      VPlanTransforms::expandSCEVs(BestVPlan, *PSE.getSE());7365  if (!ILV.getTripCount())7366    ILV.setTripCount(BestVPlan.getTripCount()->getLiveInIRValue());7367  else7368    assert(VectorizingEpilogue && "should only re-use the existing trip "7369                                  "count during epilogue vectorization");7370 7371  // Perform the actual loop transformation.7372  VPTransformState State(&TTI, BestVF, LI, DT, ILV.AC, ILV.Builder, &BestVPlan,7373                         OrigLoop->getParentLoop(),7374                         Legal->getWidestInductionType());7375 7376#ifdef EXPENSIVE_CHECKS7377  assert(DT->verify(DominatorTree::VerificationLevel::Fast));7378#endif7379 7380  // 1. Set up the skeleton for vectorization, including vector pre-header and7381  // middle block. The vector loop is created during VPlan execution.7382  State.CFG.PrevBB = ILV.createVectorizedLoopSkeleton();7383  replaceVPBBWithIRVPBB(BestVPlan.getScalarPreheader(),7384                        State.CFG.PrevBB->getSingleSuccessor(), &BestVPlan);7385  VPlanTransforms::removeDeadRecipes(BestVPlan);7386 7387  assert(verifyVPlanIsValid(BestVPlan, true /*VerifyLate*/) &&7388         "final VPlan is invalid");7389 7390  // After vectorization, the exit blocks of the original loop will have7391  // additional predecessors. Invalidate SCEVs for the exit phis in case SE7392  // looked through single-entry phis.7393  ScalarEvolution &SE = *PSE.getSE();7394  for (VPIRBasicBlock *Exit : BestVPlan.getExitBlocks()) {7395    if (!Exit->hasPredecessors())7396      continue;7397    for (VPRecipeBase &PhiR : Exit->phis())7398      SE.forgetLcssaPhiWithNewPredecessor(OrigLoop,7399                                          &cast<VPIRPhi>(PhiR).getIRPhi());7400  }7401  // Forget the original loop and block dispositions.7402  SE.forgetLoop(OrigLoop);7403  SE.forgetBlockAndLoopDispositions();7404 7405  ILV.printDebugTracesAtStart();7406 7407  //===------------------------------------------------===//7408  //7409  // Notice: any optimization or new instruction that go7410  // into the code below should also be implemented in7411  // the cost-model.7412  //7413  //===------------------------------------------------===//7414 7415  // Retrieve loop information before executing the plan, which may remove the7416  // original loop, if it becomes unreachable.7417  MDNode *LID = OrigLoop->getLoopID();7418  unsigned OrigLoopInvocationWeight = 0;7419  std::optional<unsigned> OrigAverageTripCount =7420      getLoopEstimatedTripCount(OrigLoop, &OrigLoopInvocationWeight);7421 7422  BestVPlan.execute(&State);7423 7424  // 2.6. Maintain Loop Hints7425  // Keep all loop hints from the original loop on the vector loop (we'll7426  // replace the vectorizer-specific hints below).7427  VPBasicBlock *HeaderVPBB = vputils::getFirstLoopHeader(BestVPlan, State.VPDT);7428  // Add metadata to disable runtime unrolling a scalar loop when there7429  // are no runtime checks about strides and memory. A scalar loop that is7430  // rarely used is not worth unrolling.7431  bool DisableRuntimeUnroll = !ILV.RTChecks.hasChecks() && !BestVF.isScalar();7432  updateLoopMetadataAndProfileInfo(7433      HeaderVPBB ? LI->getLoopFor(State.CFG.VPBB2IRBB.lookup(HeaderVPBB))7434                 : nullptr,7435      HeaderVPBB, BestVPlan, VectorizingEpilogue, LID, OrigAverageTripCount,7436      OrigLoopInvocationWeight,7437      estimateElementCount(BestVF * BestUF, CM.getVScaleForTuning()),7438      DisableRuntimeUnroll);7439 7440  // 3. Fix the vectorized code: take care of header phi's, live-outs,7441  //    predication, updating analyses.7442  ILV.fixVectorizedLoop(State);7443 7444  ILV.printDebugTracesAtEnd();7445 7446  return ExpandedSCEVs;7447}7448 7449//===--------------------------------------------------------------------===//7450// EpilogueVectorizerMainLoop7451//===--------------------------------------------------------------------===//7452 7453/// This function is partially responsible for generating the control flow7454/// depicted in https://llvm.org/docs/Vectorizers.html#epilogue-vectorization.7455BasicBlock *EpilogueVectorizerMainLoop::createVectorizedLoopSkeleton() {7456  BasicBlock *ScalarPH = createScalarPreheader("");7457  BasicBlock *VectorPH = ScalarPH->getSinglePredecessor();7458 7459  // Generate the code to check the minimum iteration count of the vector7460  // epilogue (see below).7461  EPI.EpilogueIterationCountCheck =7462      emitIterationCountCheck(VectorPH, ScalarPH, true);7463  EPI.EpilogueIterationCountCheck->setName("iter.check");7464 7465  VectorPH = cast<BranchInst>(EPI.EpilogueIterationCountCheck->getTerminator())7466                 ->getSuccessor(1);7467  // Generate the iteration count check for the main loop, *after* the check7468  // for the epilogue loop, so that the path-length is shorter for the case7469  // that goes directly through the vector epilogue. The longer-path length for7470  // the main loop is compensated for, by the gain from vectorizing the larger7471  // trip count. Note: the branch will get updated later on when we vectorize7472  // the epilogue.7473  EPI.MainLoopIterationCountCheck =7474      emitIterationCountCheck(VectorPH, ScalarPH, false);7475 7476  return cast<BranchInst>(EPI.MainLoopIterationCountCheck->getTerminator())7477      ->getSuccessor(1);7478}7479 7480void EpilogueVectorizerMainLoop::printDebugTracesAtStart() {7481  LLVM_DEBUG({7482    dbgs() << "Create Skeleton for epilogue vectorized loop (first pass)\n"7483           << "Main Loop VF:" << EPI.MainLoopVF7484           << ", Main Loop UF:" << EPI.MainLoopUF7485           << ", Epilogue Loop VF:" << EPI.EpilogueVF7486           << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n";7487  });7488}7489 7490void EpilogueVectorizerMainLoop::printDebugTracesAtEnd() {7491  DEBUG_WITH_TYPE(VerboseDebug, {7492    dbgs() << "intermediate fn:\n"7493           << *OrigLoop->getHeader()->getParent() << "\n";7494  });7495}7496 7497BasicBlock *EpilogueVectorizerMainLoop::emitIterationCountCheck(7498    BasicBlock *VectorPH, BasicBlock *Bypass, bool ForEpilogue) {7499  assert(Bypass && "Expected valid bypass basic block.");7500  Value *Count = getTripCount();7501  MinProfitableTripCount = ElementCount::getFixed(0);7502  Value *CheckMinIters = createIterationCountCheck(7503      VectorPH, ForEpilogue ? EPI.EpilogueVF : EPI.MainLoopVF,7504      ForEpilogue ? EPI.EpilogueUF : EPI.MainLoopUF);7505 7506  BasicBlock *const TCCheckBlock = VectorPH;7507  if (!ForEpilogue)7508    TCCheckBlock->setName("vector.main.loop.iter.check");7509 7510  // Create new preheader for vector loop.7511  VectorPH = SplitBlock(TCCheckBlock, TCCheckBlock->getTerminator(),7512                        static_cast<DominatorTree *>(nullptr), LI, nullptr,7513                        "vector.ph");7514  if (ForEpilogue) {7515    // Save the trip count so we don't have to regenerate it in the7516    // vec.epilog.iter.check. This is safe to do because the trip count7517    // generated here dominates the vector epilog iter check.7518    EPI.TripCount = Count;7519  } else {7520    VectorPHVPBB = replaceVPBBWithIRVPBB(VectorPHVPBB, VectorPH);7521  }7522 7523  BranchInst &BI = *BranchInst::Create(Bypass, VectorPH, CheckMinIters);7524  if (hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator()))7525    setBranchWeights(BI, MinItersBypassWeights, /*IsExpected=*/false);7526  ReplaceInstWithInst(TCCheckBlock->getTerminator(), &BI);7527 7528  // When vectorizing the main loop, its trip-count check is placed in a new7529  // block, whereas the overall trip-count check is placed in the VPlan entry7530  // block. When vectorizing the epilogue loop, its trip-count check is placed7531  // in the VPlan entry block.7532  if (!ForEpilogue)7533    introduceCheckBlockInVPlan(TCCheckBlock);7534  return TCCheckBlock;7535}7536 7537//===--------------------------------------------------------------------===//7538// EpilogueVectorizerEpilogueLoop7539//===--------------------------------------------------------------------===//7540 7541/// This function creates a new scalar preheader, using the previous one as7542/// entry block to the epilogue VPlan. The minimum iteration check is being7543/// represented in VPlan.7544BasicBlock *EpilogueVectorizerEpilogueLoop::createVectorizedLoopSkeleton() {7545  BasicBlock *NewScalarPH = createScalarPreheader("vec.epilog.");7546  BasicBlock *OriginalScalarPH = NewScalarPH->getSinglePredecessor();7547  OriginalScalarPH->setName("vec.epilog.iter.check");7548  VPIRBasicBlock *NewEntry = Plan.createVPIRBasicBlock(OriginalScalarPH);7549  VPBasicBlock *OldEntry = Plan.getEntry();7550  for (auto &R : make_early_inc_range(*OldEntry)) {7551    // Skip moving VPIRInstructions (including VPIRPhis), which are unmovable by7552    // defining.7553    if (isa<VPIRInstruction>(&R))7554      continue;7555    R.moveBefore(*NewEntry, NewEntry->end());7556  }7557 7558  VPBlockUtils::reassociateBlocks(OldEntry, NewEntry);7559  Plan.setEntry(NewEntry);7560  // OldEntry is now dead and will be cleaned up when the plan gets destroyed.7561 7562  return OriginalScalarPH;7563}7564 7565void EpilogueVectorizerEpilogueLoop::printDebugTracesAtStart() {7566  LLVM_DEBUG({7567    dbgs() << "Create Skeleton for epilogue vectorized loop (second pass)\n"7568           << "Epilogue Loop VF:" << EPI.EpilogueVF7569           << ", Epilogue Loop UF:" << EPI.EpilogueUF << "\n";7570  });7571}7572 7573void EpilogueVectorizerEpilogueLoop::printDebugTracesAtEnd() {7574  DEBUG_WITH_TYPE(VerboseDebug, {7575    dbgs() << "final fn:\n" << *OrigLoop->getHeader()->getParent() << "\n";7576  });7577}7578 7579VPWidenMemoryRecipe *VPRecipeBuilder::tryToWidenMemory(VPInstruction *VPI,7580                                                       VFRange &Range) {7581  assert((VPI->getOpcode() == Instruction::Load ||7582          VPI->getOpcode() == Instruction::Store) &&7583         "Must be called with either a load or store");7584  Instruction *I = VPI->getUnderlyingInstr();7585 7586  auto WillWiden = [&](ElementCount VF) -> bool {7587    LoopVectorizationCostModel::InstWidening Decision =7588        CM.getWideningDecision(I, VF);7589    assert(Decision != LoopVectorizationCostModel::CM_Unknown &&7590           "CM decision should be taken at this point.");7591    if (Decision == LoopVectorizationCostModel::CM_Interleave)7592      return true;7593    if (CM.isScalarAfterVectorization(I, VF) ||7594        CM.isProfitableToScalarize(I, VF))7595      return false;7596    return Decision != LoopVectorizationCostModel::CM_Scalarize;7597  };7598 7599  if (!LoopVectorizationPlanner::getDecisionAndClampRange(WillWiden, Range))7600    return nullptr;7601 7602  VPValue *Mask = nullptr;7603  if (Legal->isMaskRequired(I))7604    Mask = getBlockInMask(Builder.getInsertBlock());7605 7606  // Determine if the pointer operand of the access is either consecutive or7607  // reverse consecutive.7608  LoopVectorizationCostModel::InstWidening Decision =7609      CM.getWideningDecision(I, Range.Start);7610  bool Reverse = Decision == LoopVectorizationCostModel::CM_Widen_Reverse;7611  bool Consecutive =7612      Reverse || Decision == LoopVectorizationCostModel::CM_Widen;7613 7614  VPValue *Ptr = VPI->getOpcode() == Instruction::Load ? VPI->getOperand(0)7615                                                       : VPI->getOperand(1);7616  if (Consecutive) {7617    auto *GEP = dyn_cast<GetElementPtrInst>(7618        Ptr->getUnderlyingValue()->stripPointerCasts());7619    VPSingleDefRecipe *VectorPtr;7620    if (Reverse) {7621      // When folding the tail, we may compute an address that we don't in the7622      // original scalar loop: drop the GEP no-wrap flags in this case.7623      // Otherwise preserve existing flags without no-unsigned-wrap, as we will7624      // emit negative indices.7625      GEPNoWrapFlags Flags =7626          CM.foldTailByMasking() || !GEP7627              ? GEPNoWrapFlags::none()7628              : GEP->getNoWrapFlags().withoutNoUnsignedWrap();7629      VectorPtr = new VPVectorEndPointerRecipe(7630          Ptr, &Plan.getVF(), getLoadStoreType(I),7631          /*Stride*/ -1, Flags, VPI->getDebugLoc());7632    } else {7633      VectorPtr = new VPVectorPointerRecipe(Ptr, getLoadStoreType(I),7634                                            GEP ? GEP->getNoWrapFlags()7635                                                : GEPNoWrapFlags::none(),7636                                            VPI->getDebugLoc());7637    }7638    Builder.insert(VectorPtr);7639    Ptr = VectorPtr;7640  }7641  if (VPI->getOpcode() == Instruction::Load) {7642    auto *Load = cast<LoadInst>(I);7643    return new VPWidenLoadRecipe(*Load, Ptr, Mask, Consecutive, Reverse, *VPI,7644                                 VPI->getDebugLoc());7645  }7646 7647  StoreInst *Store = cast<StoreInst>(I);7648  return new VPWidenStoreRecipe(*Store, Ptr, VPI->getOperand(0), Mask,7649                                Consecutive, Reverse, *VPI, VPI->getDebugLoc());7650}7651 7652/// Creates a VPWidenIntOrFpInductionRecipe for \p PhiR. If needed, it will7653/// also insert a recipe to expand the step for the induction recipe.7654static VPWidenIntOrFpInductionRecipe *7655createWidenInductionRecipes(VPInstruction *PhiR,7656                            const InductionDescriptor &IndDesc, VPlan &Plan,7657                            ScalarEvolution &SE, Loop &OrigLoop) {7658  assert(SE.isLoopInvariant(IndDesc.getStep(), &OrigLoop) &&7659         "step must be loop invariant");7660 7661  VPValue *Start = PhiR->getOperand(0);7662  assert(Plan.getLiveIn(IndDesc.getStartValue()) == Start &&7663         "Start VPValue must match IndDesc's start value");7664 7665  // It is always safe to copy over the NoWrap and FastMath flags. In7666  // particular, when folding tail by masking, the masked-off lanes are never7667  // used, so it is safe.7668  VPIRFlags Flags = vputils::getFlagsFromIndDesc(IndDesc);7669  VPValue *Step =7670      vputils::getOrCreateVPValueForSCEVExpr(Plan, IndDesc.getStep());7671 7672  // Update wide induction increments to use the same step as the corresponding7673  // wide induction. This enables detecting induction increments directly in7674  // VPlan and removes redundant splats.7675  using namespace llvm::VPlanPatternMatch;7676  if (match(PhiR->getOperand(1), m_Add(m_Specific(PhiR), m_VPValue())))7677    PhiR->getOperand(1)->getDefiningRecipe()->setOperand(1, Step);7678 7679  PHINode *Phi = cast<PHINode>(PhiR->getUnderlyingInstr());7680  return new VPWidenIntOrFpInductionRecipe(Phi, Start, Step, &Plan.getVF(),7681                                           IndDesc, Flags, PhiR->getDebugLoc());7682}7683 7684VPHeaderPHIRecipe *7685VPRecipeBuilder::tryToOptimizeInductionPHI(VPInstruction *VPI) {7686  auto *Phi = cast<PHINode>(VPI->getUnderlyingInstr());7687 7688  // Check if this is an integer or fp induction. If so, build the recipe that7689  // produces its scalar and vector values.7690  if (auto *II = Legal->getIntOrFpInductionDescriptor(Phi))7691    return createWidenInductionRecipes(VPI, *II, Plan, *PSE.getSE(), *OrigLoop);7692 7693  // Check if this is pointer induction. If so, build the recipe for it.7694  if (auto *II = Legal->getPointerInductionDescriptor(Phi)) {7695    VPValue *Step = vputils::getOrCreateVPValueForSCEVExpr(Plan, II->getStep());7696    return new VPWidenPointerInductionRecipe(Phi, VPI->getOperand(0), Step,7697                                             &Plan.getVFxUF(), *II,7698                                             VPI->getDebugLoc());7699  }7700  return nullptr;7701}7702 7703VPWidenIntOrFpInductionRecipe *7704VPRecipeBuilder::tryToOptimizeInductionTruncate(VPInstruction *VPI,7705                                                VFRange &Range) {7706  auto *I = cast<TruncInst>(VPI->getUnderlyingInstr());7707  // Optimize the special case where the source is a constant integer7708  // induction variable. Notice that we can only optimize the 'trunc' case7709  // because (a) FP conversions lose precision, (b) sext/zext may wrap, and7710  // (c) other casts depend on pointer size.7711 7712  // Determine whether \p K is a truncation based on an induction variable that7713  // can be optimized.7714  auto IsOptimizableIVTruncate =7715      [&](Instruction *K) -> std::function<bool(ElementCount)> {7716    return [=](ElementCount VF) -> bool {7717      return CM.isOptimizableIVTruncate(K, VF);7718    };7719  };7720 7721  if (!LoopVectorizationPlanner::getDecisionAndClampRange(7722          IsOptimizableIVTruncate(I), Range))7723    return nullptr;7724 7725  auto *WidenIV = cast<VPWidenIntOrFpInductionRecipe>(7726      VPI->getOperand(0)->getDefiningRecipe());7727  PHINode *Phi = WidenIV->getPHINode();7728  VPValue *Start = WidenIV->getStartValue();7729  const InductionDescriptor &IndDesc = WidenIV->getInductionDescriptor();7730 7731  // It is always safe to copy over the NoWrap and FastMath flags. In7732  // particular, when folding tail by masking, the masked-off lanes are never7733  // used, so it is safe.7734  VPIRFlags Flags = vputils::getFlagsFromIndDesc(IndDesc);7735  VPValue *Step =7736      vputils::getOrCreateVPValueForSCEVExpr(Plan, IndDesc.getStep());7737  return new VPWidenIntOrFpInductionRecipe(7738      Phi, Start, Step, &Plan.getVF(), IndDesc, I, Flags, VPI->getDebugLoc());7739}7740 7741VPSingleDefRecipe *VPRecipeBuilder::tryToWidenCall(VPInstruction *VPI,7742                                                   VFRange &Range) {7743  CallInst *CI = cast<CallInst>(VPI->getUnderlyingInstr());7744  bool IsPredicated = LoopVectorizationPlanner::getDecisionAndClampRange(7745      [this, CI](ElementCount VF) {7746        return CM.isScalarWithPredication(CI, VF);7747      },7748      Range);7749 7750  if (IsPredicated)7751    return nullptr;7752 7753  Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);7754  if (ID && (ID == Intrinsic::assume || ID == Intrinsic::lifetime_end ||7755             ID == Intrinsic::lifetime_start || ID == Intrinsic::sideeffect ||7756             ID == Intrinsic::pseudoprobe ||7757             ID == Intrinsic::experimental_noalias_scope_decl))7758    return nullptr;7759 7760  SmallVector<VPValue *, 4> Ops(VPI->op_begin(),7761                                VPI->op_begin() + CI->arg_size());7762 7763  // Is it beneficial to perform intrinsic call compared to lib call?7764  bool ShouldUseVectorIntrinsic =7765      ID && LoopVectorizationPlanner::getDecisionAndClampRange(7766                [&](ElementCount VF) -> bool {7767                  return CM.getCallWideningDecision(CI, VF).Kind ==7768                         LoopVectorizationCostModel::CM_IntrinsicCall;7769                },7770                Range);7771  if (ShouldUseVectorIntrinsic)7772    return new VPWidenIntrinsicRecipe(*CI, ID, Ops, CI->getType(), *VPI, *VPI,7773                                      VPI->getDebugLoc());7774 7775  Function *Variant = nullptr;7776  std::optional<unsigned> MaskPos;7777  // Is better to call a vectorized version of the function than to to scalarize7778  // the call?7779  auto ShouldUseVectorCall = LoopVectorizationPlanner::getDecisionAndClampRange(7780      [&](ElementCount VF) -> bool {7781        // The following case may be scalarized depending on the VF.7782        // The flag shows whether we can use a usual Call for vectorized7783        // version of the instruction.7784 7785        // If we've found a variant at a previous VF, then stop looking. A7786        // vectorized variant of a function expects input in a certain shape7787        // -- basically the number of input registers, the number of lanes7788        // per register, and whether there's a mask required.7789        // We store a pointer to the variant in the VPWidenCallRecipe, so7790        // once we have an appropriate variant it's only valid for that VF.7791        // This will force a different vplan to be generated for each VF that7792        // finds a valid variant.7793        if (Variant)7794          return false;7795        LoopVectorizationCostModel::CallWideningDecision Decision =7796            CM.getCallWideningDecision(CI, VF);7797        if (Decision.Kind == LoopVectorizationCostModel::CM_VectorCall) {7798          Variant = Decision.Variant;7799          MaskPos = Decision.MaskPos;7800          return true;7801        }7802 7803        return false;7804      },7805      Range);7806  if (ShouldUseVectorCall) {7807    if (MaskPos.has_value()) {7808      // We have 2 cases that would require a mask:7809      //   1) The block needs to be predicated, either due to a conditional7810      //      in the scalar loop or use of an active lane mask with7811      //      tail-folding, and we use the appropriate mask for the block.7812      //   2) No mask is required for the block, but the only available7813      //      vector variant at this VF requires a mask, so we synthesize an7814      //      all-true mask.7815      VPValue *Mask = nullptr;7816      if (Legal->isMaskRequired(CI))7817        Mask = getBlockInMask(Builder.getInsertBlock());7818      else7819        Mask = Plan.getOrAddLiveIn(7820            ConstantInt::getTrue(IntegerType::getInt1Ty(Plan.getContext())));7821 7822      Ops.insert(Ops.begin() + *MaskPos, Mask);7823    }7824 7825    Ops.push_back(VPI->getOperand(VPI->getNumOperands() - 1));7826    return new VPWidenCallRecipe(CI, Variant, Ops, *VPI, *VPI,7827                                 VPI->getDebugLoc());7828  }7829 7830  return nullptr;7831}7832 7833bool VPRecipeBuilder::shouldWiden(Instruction *I, VFRange &Range) const {7834  assert(!isa<BranchInst>(I) && !isa<PHINode>(I) && !isa<LoadInst>(I) &&7835         !isa<StoreInst>(I) && "Instruction should have been handled earlier");7836  // Instruction should be widened, unless it is scalar after vectorization,7837  // scalarization is profitable or it is predicated.7838  auto WillScalarize = [this, I](ElementCount VF) -> bool {7839    return CM.isScalarAfterVectorization(I, VF) ||7840           CM.isProfitableToScalarize(I, VF) ||7841           CM.isScalarWithPredication(I, VF);7842  };7843  return !LoopVectorizationPlanner::getDecisionAndClampRange(WillScalarize,7844                                                             Range);7845}7846 7847VPWidenRecipe *VPRecipeBuilder::tryToWiden(VPInstruction *VPI) {7848  auto *I = VPI->getUnderlyingInstr();7849  switch (VPI->getOpcode()) {7850  default:7851    return nullptr;7852  case Instruction::SDiv:7853  case Instruction::UDiv:7854  case Instruction::SRem:7855  case Instruction::URem: {7856    // If not provably safe, use a select to form a safe divisor before widening the7857    // div/rem operation itself.  Otherwise fall through to general handling below.7858    if (CM.isPredicatedInst(I)) {7859      SmallVector<VPValue *> Ops(VPI->operands());7860      VPValue *Mask = getBlockInMask(Builder.getInsertBlock());7861      VPValue *One = Plan.getConstantInt(I->getType(), 1u);7862      auto *SafeRHS =7863          Builder.createSelect(Mask, Ops[1], One, VPI->getDebugLoc());7864      Ops[1] = SafeRHS;7865      return new VPWidenRecipe(*I, Ops, *VPI, *VPI, VPI->getDebugLoc());7866    }7867    [[fallthrough]];7868  }7869  case Instruction::Add:7870  case Instruction::And:7871  case Instruction::AShr:7872  case Instruction::FAdd:7873  case Instruction::FCmp:7874  case Instruction::FDiv:7875  case Instruction::FMul:7876  case Instruction::FNeg:7877  case Instruction::FRem:7878  case Instruction::FSub:7879  case Instruction::ICmp:7880  case Instruction::LShr:7881  case Instruction::Mul:7882  case Instruction::Or:7883  case Instruction::Select:7884  case Instruction::Shl:7885  case Instruction::Sub:7886  case Instruction::Xor:7887  case Instruction::Freeze: {7888    SmallVector<VPValue *> NewOps(VPI->operands());7889    if (Instruction::isBinaryOp(VPI->getOpcode())) {7890      // The legacy cost model uses SCEV to check if some of the operands are7891      // constants. To match the legacy cost model's behavior, use SCEV to try7892      // to replace operands with constants.7893      ScalarEvolution &SE = *PSE.getSE();7894      auto GetConstantViaSCEV = [this, &SE](VPValue *Op) {7895        if (!Op->isLiveIn())7896          return Op;7897        Value *V = Op->getUnderlyingValue();7898        if (isa<Constant>(V) || !SE.isSCEVable(V->getType()))7899          return Op;7900        auto *C = dyn_cast<SCEVConstant>(SE.getSCEV(V));7901        if (!C)7902          return Op;7903        return Plan.getOrAddLiveIn(C->getValue());7904      };7905      // For Mul, the legacy cost model checks both operands.7906      if (VPI->getOpcode() == Instruction::Mul)7907        NewOps[0] = GetConstantViaSCEV(NewOps[0]);7908      // For other binops, the legacy cost model only checks the second operand.7909      NewOps[1] = GetConstantViaSCEV(NewOps[1]);7910    }7911    return new VPWidenRecipe(*I, NewOps, *VPI, *VPI, VPI->getDebugLoc());7912  }7913  case Instruction::ExtractValue: {7914    SmallVector<VPValue *> NewOps(VPI->operands());7915    auto *EVI = cast<ExtractValueInst>(I);7916    assert(EVI->getNumIndices() == 1 && "Expected one extractvalue index");7917    unsigned Idx = EVI->getIndices()[0];7918    NewOps.push_back(Plan.getConstantInt(32, Idx));7919    return new VPWidenRecipe(*I, NewOps, *VPI, *VPI, VPI->getDebugLoc());7920  }7921  };7922}7923 7924VPHistogramRecipe *VPRecipeBuilder::tryToWidenHistogram(const HistogramInfo *HI,7925                                                        VPInstruction *VPI) {7926  // FIXME: Support other operations.7927  unsigned Opcode = HI->Update->getOpcode();7928  assert((Opcode == Instruction::Add || Opcode == Instruction::Sub) &&7929         "Histogram update operation must be an Add or Sub");7930 7931  SmallVector<VPValue *, 3> HGramOps;7932  // Bucket address.7933  HGramOps.push_back(VPI->getOperand(1));7934  // Increment value.7935  HGramOps.push_back(getVPValueOrAddLiveIn(HI->Update->getOperand(1)));7936 7937  // In case of predicated execution (due to tail-folding, or conditional7938  // execution, or both), pass the relevant mask.7939  if (Legal->isMaskRequired(HI->Store))7940    HGramOps.push_back(getBlockInMask(Builder.getInsertBlock()));7941 7942  return new VPHistogramRecipe(Opcode, HGramOps, VPI->getDebugLoc());7943}7944 7945VPReplicateRecipe *VPRecipeBuilder::handleReplication(VPInstruction *VPI,7946                                                      VFRange &Range) {7947  auto *I = VPI->getUnderlyingInstr();7948  bool IsUniform = LoopVectorizationPlanner::getDecisionAndClampRange(7949      [&](ElementCount VF) { return CM.isUniformAfterVectorization(I, VF); },7950      Range);7951 7952  bool IsPredicated = CM.isPredicatedInst(I);7953 7954  // Even if the instruction is not marked as uniform, there are certain7955  // intrinsic calls that can be effectively treated as such, so we check for7956  // them here. Conservatively, we only do this for scalable vectors, since7957  // for fixed-width VFs we can always fall back on full scalarization.7958  if (!IsUniform && Range.Start.isScalable() && isa<IntrinsicInst>(I)) {7959    switch (cast<IntrinsicInst>(I)->getIntrinsicID()) {7960    case Intrinsic::assume:7961    case Intrinsic::lifetime_start:7962    case Intrinsic::lifetime_end:7963      // For scalable vectors if one of the operands is variant then we still7964      // want to mark as uniform, which will generate one instruction for just7965      // the first lane of the vector. We can't scalarize the call in the same7966      // way as for fixed-width vectors because we don't know how many lanes7967      // there are.7968      //7969      // The reasons for doing it this way for scalable vectors are:7970      //   1. For the assume intrinsic generating the instruction for the first7971      //      lane is still be better than not generating any at all. For7972      //      example, the input may be a splat across all lanes.7973      //   2. For the lifetime start/end intrinsics the pointer operand only7974      //      does anything useful when the input comes from a stack object,7975      //      which suggests it should always be uniform. For non-stack objects7976      //      the effect is to poison the object, which still allows us to7977      //      remove the call.7978      IsUniform = true;7979      break;7980    default:7981      break;7982    }7983  }7984  VPValue *BlockInMask = nullptr;7985  if (!IsPredicated) {7986    // Finalize the recipe for Instr, first if it is not predicated.7987    LLVM_DEBUG(dbgs() << "LV: Scalarizing:" << *I << "\n");7988  } else {7989    LLVM_DEBUG(dbgs() << "LV: Scalarizing and predicating:" << *I << "\n");7990    // Instructions marked for predication are replicated and a mask operand is7991    // added initially. Masked replicate recipes will later be placed under an7992    // if-then construct to prevent side-effects. Generate recipes to compute7993    // the block mask for this region.7994    BlockInMask = getBlockInMask(Builder.getInsertBlock());7995  }7996 7997  // Note that there is some custom logic to mark some intrinsics as uniform7998  // manually above for scalable vectors, which this assert needs to account for7999  // as well.8000  assert((Range.Start.isScalar() || !IsUniform || !IsPredicated ||8001          (Range.Start.isScalable() && isa<IntrinsicInst>(I))) &&8002         "Should not predicate a uniform recipe");8003  auto *Recipe =8004      new VPReplicateRecipe(I, VPI->operands(), IsUniform, BlockInMask, *VPI,8005                            *VPI, VPI->getDebugLoc());8006  return Recipe;8007}8008 8009/// Find all possible partial reductions in the loop and track all of those that8010/// are valid so recipes can be formed later.8011void VPRecipeBuilder::collectScaledReductions(VFRange &Range) {8012  // Find all possible partial reductions, grouping chains by their PHI. This8013  // grouping allows invalidating the whole chain, if any link is not a valid8014  // partial reduction.8015  MapVector<Instruction *,8016            SmallVector<std::pair<PartialReductionChain, unsigned>>>8017      ChainsByPhi;8018  for (const auto &[Phi, RdxDesc] : Legal->getReductionVars()) {8019    if (Instruction *RdxExitInstr = RdxDesc.getLoopExitInstr())8020      getScaledReductions(Phi, RdxExitInstr, Range, ChainsByPhi[Phi]);8021  }8022 8023  // A partial reduction is invalid if any of its extends are used by8024  // something that isn't another partial reduction. This is because the8025  // extends are intended to be lowered along with the reduction itself.8026 8027  // Build up a set of partial reduction ops for efficient use checking.8028  SmallPtrSet<User *, 4> PartialReductionOps;8029  for (const auto &[_, Chains] : ChainsByPhi)8030    for (const auto &[PartialRdx, _] : Chains)8031      PartialReductionOps.insert(PartialRdx.ExtendUser);8032 8033  auto ExtendIsOnlyUsedByPartialReductions =8034      [&PartialReductionOps](Instruction *Extend) {8035        return all_of(Extend->users(), [&](const User *U) {8036          return PartialReductionOps.contains(U);8037        });8038      };8039 8040  // Check if each use of a chain's two extends is a partial reduction8041  // and only add those that don't have non-partial reduction users.8042  for (const auto &[_, Chains] : ChainsByPhi) {8043    for (const auto &[Chain, Scale] : Chains) {8044      if (ExtendIsOnlyUsedByPartialReductions(Chain.ExtendA) &&8045          (!Chain.ExtendB ||8046           ExtendIsOnlyUsedByPartialReductions(Chain.ExtendB)))8047        ScaledReductionMap.try_emplace(Chain.Reduction, Scale);8048    }8049  }8050 8051  // Check that all partial reductions in a chain are only used by other8052  // partial reductions with the same scale factor. Otherwise we end up creating8053  // users of scaled reductions where the types of the other operands don't8054  // match.8055  for (const auto &[Phi, Chains] : ChainsByPhi) {8056    for (const auto &[Chain, Scale] : Chains) {8057      auto AllUsersPartialRdx = [ScaleVal = Scale, RdxPhi = Phi,8058                                 this](const User *U) {8059        auto *UI = cast<Instruction>(U);8060        if (isa<PHINode>(UI) && UI->getParent() == OrigLoop->getHeader())8061          return UI == RdxPhi;8062        return ScaledReductionMap.lookup_or(UI, 0) == ScaleVal ||8063               !OrigLoop->contains(UI->getParent());8064      };8065 8066      // If any partial reduction entry for the phi is invalid, invalidate the8067      // whole chain.8068      if (!all_of(Chain.Reduction->users(), AllUsersPartialRdx)) {8069        for (const auto &[Chain, _] : Chains)8070          ScaledReductionMap.erase(Chain.Reduction);8071        break;8072      }8073    }8074  }8075}8076 8077bool VPRecipeBuilder::getScaledReductions(8078    Instruction *PHI, Instruction *RdxExitInstr, VFRange &Range,8079    SmallVectorImpl<std::pair<PartialReductionChain, unsigned>> &Chains) {8080  if (!CM.TheLoop->contains(RdxExitInstr))8081    return false;8082 8083  auto *Update = dyn_cast<BinaryOperator>(RdxExitInstr);8084  if (!Update)8085    return false;8086 8087  Value *Op = Update->getOperand(0);8088  Value *PhiOp = Update->getOperand(1);8089  if (Op == PHI)8090    std::swap(Op, PhiOp);8091 8092  using namespace llvm::PatternMatch;8093  // If Op is an extend, then it's still a valid partial reduction if the8094  // extended mul fulfills the other requirements.8095  // For example, reduce.add(ext(mul(ext(A), ext(B)))) is still a valid partial8096  // reduction since the inner extends will be widened. We already have oneUse8097  // checks on the inner extends so widening them is safe.8098  std::optional<TTI::PartialReductionExtendKind> OuterExtKind = std::nullopt;8099  if (match(Op, m_ZExtOrSExt(m_Mul(m_Value(), m_Value())))) {8100    auto *Cast = cast<CastInst>(Op);8101    OuterExtKind = TTI::getPartialReductionExtendKind(Cast->getOpcode());8102    Op = Cast->getOperand(0);8103  }8104 8105  // Try and get a scaled reduction from the first non-phi operand.8106  // If one is found, we use the discovered reduction instruction in8107  // place of the accumulator for costing.8108  if (auto *OpInst = dyn_cast<Instruction>(Op)) {8109    if (getScaledReductions(PHI, OpInst, Range, Chains)) {8110      PHI = Chains.rbegin()->first.Reduction;8111 8112      Op = Update->getOperand(0);8113      PhiOp = Update->getOperand(1);8114      if (Op == PHI)8115        std::swap(Op, PhiOp);8116    }8117  }8118  if (PhiOp != PHI)8119    return false;8120 8121  // If the update is a binary operator, check both of its operands to see if8122  // they are extends. Otherwise, see if the update comes directly from an8123  // extend.8124  Instruction *Exts[2] = {nullptr};8125  BinaryOperator *ExtendUser = dyn_cast<BinaryOperator>(Op);8126  std::optional<unsigned> BinOpc;8127  Type *ExtOpTypes[2] = {nullptr};8128  TTI::PartialReductionExtendKind ExtKinds[2] = {TTI::PR_None};8129 8130  auto CollectExtInfo = [this, OuterExtKind, &Exts, &ExtOpTypes,8131                         &ExtKinds](SmallVectorImpl<Value *> &Ops) -> bool {8132    for (const auto &[I, OpI] : enumerate(Ops)) {8133      const APInt *C;8134      if (I > 0 && match(OpI, m_APInt(C)) &&8135          canConstantBeExtended(C, ExtOpTypes[0], ExtKinds[0])) {8136        ExtOpTypes[I] = ExtOpTypes[0];8137        ExtKinds[I] = ExtKinds[0];8138        continue;8139      }8140      Value *ExtOp;8141      if (!match(OpI, m_ZExtOrSExt(m_Value(ExtOp))))8142        return false;8143      Exts[I] = cast<Instruction>(OpI);8144 8145      // TODO: We should be able to support live-ins.8146      if (!CM.TheLoop->contains(Exts[I]))8147        return false;8148 8149      ExtOpTypes[I] = ExtOp->getType();8150      ExtKinds[I] = TTI::getPartialReductionExtendKind(Exts[I]);8151      // The outer extend kind must be the same as the inner extends, so that8152      // they can be folded together.8153      if (OuterExtKind.has_value() && OuterExtKind.value() != ExtKinds[I])8154        return false;8155    }8156    return true;8157  };8158 8159  if (ExtendUser) {8160    if (!ExtendUser->hasOneUse())8161      return false;8162 8163    // Use the side-effect of match to replace BinOp only if the pattern is8164    // matched, we don't care at this point whether it actually matched.8165    match(ExtendUser, m_Neg(m_BinOp(ExtendUser)));8166 8167    SmallVector<Value *> Ops(ExtendUser->operands());8168    if (!CollectExtInfo(Ops))8169      return false;8170 8171    BinOpc = std::make_optional(ExtendUser->getOpcode());8172  } else if (match(Update, m_Add(m_Value(), m_Value()))) {8173    // We already know the operands for Update are Op and PhiOp.8174    SmallVector<Value *> Ops({Op});8175    if (!CollectExtInfo(Ops))8176      return false;8177 8178    ExtendUser = Update;8179    BinOpc = std::nullopt;8180  } else8181    return false;8182 8183  PartialReductionChain Chain(RdxExitInstr, Exts[0], Exts[1], ExtendUser);8184 8185  TypeSize PHISize = PHI->getType()->getPrimitiveSizeInBits();8186  TypeSize ASize = ExtOpTypes[0]->getPrimitiveSizeInBits();8187  if (!PHISize.hasKnownScalarFactor(ASize))8188    return false;8189  unsigned TargetScaleFactor = PHISize.getKnownScalarFactor(ASize);8190 8191  if (LoopVectorizationPlanner::getDecisionAndClampRange(8192          [&](ElementCount VF) {8193            InstructionCost Cost = TTI->getPartialReductionCost(8194                Update->getOpcode(), ExtOpTypes[0], ExtOpTypes[1],8195                PHI->getType(), VF, ExtKinds[0], ExtKinds[1], BinOpc,8196                CM.CostKind);8197            return Cost.isValid();8198          },8199          Range)) {8200    Chains.emplace_back(Chain, TargetScaleFactor);8201    return true;8202  }8203 8204  return false;8205}8206 8207VPRecipeBase *VPRecipeBuilder::tryToCreateWidenRecipe(VPSingleDefRecipe *R,8208                                                      VFRange &Range) {8209  // First, check for specific widening recipes that deal with inductions, Phi8210  // nodes, calls and memory operations.8211  VPRecipeBase *Recipe;8212  if (auto *PhiR = dyn_cast<VPPhi>(R)) {8213    VPBasicBlock *Parent = PhiR->getParent();8214    [[maybe_unused]] VPRegionBlock *LoopRegionOf =8215        Parent->getEnclosingLoopRegion();8216    assert(LoopRegionOf && LoopRegionOf->getEntry() == Parent &&8217           "Non-header phis should have been handled during predication");8218    auto *Phi = cast<PHINode>(R->getUnderlyingInstr());8219    assert(R->getNumOperands() == 2 && "Must have 2 operands for header phis");8220    if ((Recipe = tryToOptimizeInductionPHI(PhiR)))8221      return Recipe;8222 8223    VPHeaderPHIRecipe *PhiRecipe = nullptr;8224    assert((Legal->isReductionVariable(Phi) ||8225            Legal->isFixedOrderRecurrence(Phi)) &&8226           "can only widen reductions and fixed-order recurrences here");8227    VPValue *StartV = R->getOperand(0);8228    if (Legal->isReductionVariable(Phi)) {8229      const RecurrenceDescriptor &RdxDesc = Legal->getRecurrenceDescriptor(Phi);8230      assert(RdxDesc.getRecurrenceStartValue() ==8231             Phi->getIncomingValueForBlock(OrigLoop->getLoopPreheader()));8232 8233      // If the PHI is used by a partial reduction, set the scale factor.8234      bool UseInLoopReduction = CM.isInLoopReduction(Phi);8235      bool UseOrderedReductions = CM.useOrderedReductions(RdxDesc);8236      unsigned ScaleFactor =8237          getScalingForReduction(RdxDesc.getLoopExitInstr()).value_or(1);8238 8239      PhiRecipe = new VPReductionPHIRecipe(8240          Phi, RdxDesc.getRecurrenceKind(), *StartV,8241          getReductionStyle(UseInLoopReduction, UseOrderedReductions,8242                            ScaleFactor),8243          RdxDesc.hasUsesOutsideReductionChain());8244    } else {8245      // TODO: Currently fixed-order recurrences are modeled as chains of8246      // first-order recurrences. If there are no users of the intermediate8247      // recurrences in the chain, the fixed order recurrence should be modeled8248      // directly, enabling more efficient codegen.8249      PhiRecipe = new VPFirstOrderRecurrencePHIRecipe(Phi, *StartV);8250    }8251    // Add backedge value.8252    PhiRecipe->addOperand(R->getOperand(1));8253    return PhiRecipe;8254  }8255  assert(!R->isPhi() && "only VPPhi nodes expected at this point");8256 8257  auto *VPI = cast<VPInstruction>(R);8258  Instruction *Instr = R->getUnderlyingInstr();8259  if (VPI->getOpcode() == Instruction::Trunc &&8260      (Recipe = tryToOptimizeInductionTruncate(VPI, Range)))8261    return Recipe;8262 8263  // All widen recipes below deal only with VF > 1.8264  if (LoopVectorizationPlanner::getDecisionAndClampRange(8265          [&](ElementCount VF) { return VF.isScalar(); }, Range))8266    return nullptr;8267 8268  if (VPI->getOpcode() == Instruction::Call)8269    return tryToWidenCall(VPI, Range);8270 8271  if (VPI->getOpcode() == Instruction::Store)8272    if (auto HistInfo = Legal->getHistogramInfo(cast<StoreInst>(Instr)))8273      return tryToWidenHistogram(*HistInfo, VPI);8274 8275  if (VPI->getOpcode() == Instruction::Load ||8276      VPI->getOpcode() == Instruction::Store)8277    return tryToWidenMemory(VPI, Range);8278 8279  if (std::optional<unsigned> ScaleFactor = getScalingForReduction(Instr))8280    return tryToCreatePartialReduction(VPI, ScaleFactor.value());8281 8282  if (!shouldWiden(Instr, Range))8283    return nullptr;8284 8285  if (VPI->getOpcode() == Instruction::GetElementPtr)8286    return new VPWidenGEPRecipe(cast<GetElementPtrInst>(Instr), R->operands(),8287                                *VPI, VPI->getDebugLoc());8288 8289  if (VPI->getOpcode() == Instruction::Select)8290    return new VPWidenSelectRecipe(cast<SelectInst>(Instr), R->operands(), *VPI,8291                                   *VPI, VPI->getDebugLoc());8292 8293  if (Instruction::isCast(VPI->getOpcode())) {8294    auto *CI = cast<CastInst>(Instr);8295    auto *CastR = cast<VPInstructionWithType>(VPI);8296    return new VPWidenCastRecipe(CI->getOpcode(), VPI->getOperand(0),8297                                 CastR->getResultType(), CI, *VPI, *VPI,8298                                 VPI->getDebugLoc());8299  }8300 8301  return tryToWiden(VPI);8302}8303 8304VPRecipeBase *8305VPRecipeBuilder::tryToCreatePartialReduction(VPInstruction *Reduction,8306                                             unsigned ScaleFactor) {8307  assert(Reduction->getNumOperands() == 2 &&8308         "Unexpected number of operands for partial reduction");8309 8310  VPValue *BinOp = Reduction->getOperand(0);8311  VPValue *Accumulator = Reduction->getOperand(1);8312  VPRecipeBase *BinOpRecipe = BinOp->getDefiningRecipe();8313  if (isa<VPReductionPHIRecipe>(BinOpRecipe) ||8314      (isa<VPReductionRecipe>(BinOpRecipe) &&8315       cast<VPReductionRecipe>(BinOpRecipe)->isPartialReduction()))8316    std::swap(BinOp, Accumulator);8317 8318  assert(ScaleFactor ==8319             vputils::getVFScaleFactor(Accumulator->getDefiningRecipe()) &&8320         "all accumulators in chain must have same scale factor");8321 8322  auto *ReductionI = Reduction->getUnderlyingInstr();8323  if (Reduction->getOpcode() == Instruction::Sub) {8324    auto *const Zero = ConstantInt::get(ReductionI->getType(), 0);8325    SmallVector<VPValue *, 2> Ops;8326    Ops.push_back(Plan.getOrAddLiveIn(Zero));8327    Ops.push_back(BinOp);8328    BinOp = new VPWidenRecipe(*ReductionI, Ops, VPIRFlags(*ReductionI),8329                              VPIRMetadata(), ReductionI->getDebugLoc());8330    Builder.insert(BinOp->getDefiningRecipe());8331  }8332 8333  VPValue *Cond = nullptr;8334  if (CM.blockNeedsPredicationForAnyReason(ReductionI->getParent()))8335    Cond = getBlockInMask(Builder.getInsertBlock());8336 8337  return new VPReductionRecipe(8338      RecurKind::Add, FastMathFlags(), ReductionI, Accumulator, BinOp, Cond,8339      RdxUnordered{/*VFScaleFactor=*/ScaleFactor}, ReductionI->getDebugLoc());8340}8341 8342void LoopVectorizationPlanner::buildVPlansWithVPRecipes(ElementCount MinVF,8343                                                        ElementCount MaxVF) {8344  if (ElementCount::isKnownGT(MinVF, MaxVF))8345    return;8346 8347  assert(OrigLoop->isInnermost() && "Inner loop expected.");8348 8349  const LoopAccessInfo *LAI = Legal->getLAI();8350  LoopVersioning LVer(*LAI, LAI->getRuntimePointerChecking()->getChecks(),8351                      OrigLoop, LI, DT, PSE.getSE());8352  if (!LAI->getRuntimePointerChecking()->getChecks().empty() &&8353      !LAI->getRuntimePointerChecking()->getDiffChecks()) {8354    // Only use noalias metadata when using memory checks guaranteeing no8355    // overlap across all iterations.8356    LVer.prepareNoAliasMetadata();8357  }8358 8359  // Create initial base VPlan0, to serve as common starting point for all8360  // candidates built later for specific VF ranges.8361  auto VPlan0 = VPlanTransforms::buildVPlan0(8362      OrigLoop, *LI, Legal->getWidestInductionType(),8363      getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()), PSE, &LVer);8364 8365  auto MaxVFTimes2 = MaxVF * 2;8366  for (ElementCount VF = MinVF; ElementCount::isKnownLT(VF, MaxVFTimes2);) {8367    VFRange SubRange = {VF, MaxVFTimes2};8368    if (auto Plan = tryToBuildVPlanWithVPRecipes(8369            std::unique_ptr<VPlan>(VPlan0->duplicate()), SubRange, &LVer)) {8370      // Now optimize the initial VPlan.8371      VPlanTransforms::hoistPredicatedLoads(*Plan, *PSE.getSE(), OrigLoop);8372      VPlanTransforms::runPass(VPlanTransforms::truncateToMinimalBitwidths,8373                               *Plan, CM.getMinimalBitwidths());8374      VPlanTransforms::runPass(VPlanTransforms::optimize, *Plan);8375      // TODO: try to put it close to addActiveLaneMask().8376      if (CM.foldTailWithEVL())8377        VPlanTransforms::runPass(VPlanTransforms::addExplicitVectorLength,8378                                 *Plan, CM.getMaxSafeElements());8379      assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid");8380      VPlans.push_back(std::move(Plan));8381    }8382    VF = SubRange.End;8383  }8384}8385 8386VPlanPtr LoopVectorizationPlanner::tryToBuildVPlanWithVPRecipes(8387    VPlanPtr Plan, VFRange &Range, LoopVersioning *LVer) {8388 8389  using namespace llvm::VPlanPatternMatch;8390  SmallPtrSet<const InterleaveGroup<Instruction> *, 1> InterleaveGroups;8391 8392  // ---------------------------------------------------------------------------8393  // Build initial VPlan: Scan the body of the loop in a topological order to8394  // visit each basic block after having visited its predecessor basic blocks.8395  // ---------------------------------------------------------------------------8396 8397  bool RequiresScalarEpilogueCheck =8398      LoopVectorizationPlanner::getDecisionAndClampRange(8399          [this](ElementCount VF) {8400            return !CM.requiresScalarEpilogue(VF.isVector());8401          },8402          Range);8403  VPlanTransforms::handleEarlyExits(*Plan, Legal->hasUncountableEarlyExit());8404  VPlanTransforms::addMiddleCheck(*Plan, RequiresScalarEpilogueCheck,8405                                  CM.foldTailByMasking());8406 8407  VPlanTransforms::createLoopRegions(*Plan);8408 8409  // Don't use getDecisionAndClampRange here, because we don't know the UF8410  // so this function is better to be conservative, rather than to split8411  // it up into different VPlans.8412  // TODO: Consider using getDecisionAndClampRange here to split up VPlans.8413  bool IVUpdateMayOverflow = false;8414  for (ElementCount VF : Range)8415    IVUpdateMayOverflow |= !isIndvarOverflowCheckKnownFalse(&CM, VF);8416 8417  TailFoldingStyle Style = CM.getTailFoldingStyle(IVUpdateMayOverflow);8418  // Use NUW for the induction increment if we proved that it won't overflow in8419  // the vector loop or when not folding the tail. In the later case, we know8420  // that the canonical induction increment will not overflow as the vector trip8421  // count is >= increment and a multiple of the increment.8422  VPRegionBlock *LoopRegion = Plan->getVectorLoopRegion();8423  bool HasNUW = !IVUpdateMayOverflow || Style == TailFoldingStyle::None;8424  if (!HasNUW) {8425    auto *IVInc =8426        LoopRegion->getExitingBasicBlock()->getTerminator()->getOperand(0);8427    assert(match(IVInc,8428                 m_VPInstruction<Instruction::Add>(8429                     m_Specific(LoopRegion->getCanonicalIV()), m_VPValue())) &&8430           "Did not find the canonical IV increment");8431    cast<VPRecipeWithIRFlags>(IVInc)->dropPoisonGeneratingFlags();8432  }8433 8434  // ---------------------------------------------------------------------------8435  // Pre-construction: record ingredients whose recipes we'll need to further8436  // process after constructing the initial VPlan.8437  // ---------------------------------------------------------------------------8438 8439  // For each interleave group which is relevant for this (possibly trimmed)8440  // Range, add it to the set of groups to be later applied to the VPlan and add8441  // placeholders for its members' Recipes which we'll be replacing with a8442  // single VPInterleaveRecipe.8443  for (InterleaveGroup<Instruction> *IG : IAI.getInterleaveGroups()) {8444    auto ApplyIG = [IG, this](ElementCount VF) -> bool {8445      bool Result = (VF.isVector() && // Query is illegal for VF == 18446                     CM.getWideningDecision(IG->getInsertPos(), VF) ==8447                         LoopVectorizationCostModel::CM_Interleave);8448      // For scalable vectors, the interleave factors must be <= 8 since we8449      // require the (de)interleaveN intrinsics instead of shufflevectors.8450      assert((!Result || !VF.isScalable() || IG->getFactor() <= 8) &&8451             "Unsupported interleave factor for scalable vectors");8452      return Result;8453    };8454    if (!getDecisionAndClampRange(ApplyIG, Range))8455      continue;8456    InterleaveGroups.insert(IG);8457  }8458 8459  // ---------------------------------------------------------------------------8460  // Predicate and linearize the top-level loop region.8461  // ---------------------------------------------------------------------------8462  auto BlockMaskCache = VPlanTransforms::introduceMasksAndLinearize(8463      *Plan, CM.foldTailByMasking());8464 8465  // ---------------------------------------------------------------------------8466  // Construct wide recipes and apply predication for original scalar8467  // VPInstructions in the loop.8468  // ---------------------------------------------------------------------------8469  VPRecipeBuilder RecipeBuilder(*Plan, OrigLoop, TLI, &TTI, Legal, CM, PSE,8470                                Builder, BlockMaskCache);8471  // TODO: Handle partial reductions with EVL tail folding.8472  if (!CM.foldTailWithEVL())8473    RecipeBuilder.collectScaledReductions(Range);8474 8475  // Scan the body of the loop in a topological order to visit each basic block8476  // after having visited its predecessor basic blocks.8477  VPBasicBlock *HeaderVPBB = LoopRegion->getEntryBasicBlock();8478  ReversePostOrderTraversal<VPBlockShallowTraversalWrapper<VPBlockBase *>> RPOT(8479      HeaderVPBB);8480 8481  auto *MiddleVPBB = Plan->getMiddleBlock();8482  VPBasicBlock::iterator MBIP = MiddleVPBB->getFirstNonPhi();8483  // Mapping from VPValues in the initial plan to their widened VPValues. Needed8484  // temporarily to update created block masks.8485  DenseMap<VPValue *, VPValue *> Old2New;8486  for (VPBasicBlock *VPBB : VPBlockUtils::blocksOnly<VPBasicBlock>(RPOT)) {8487    // Convert input VPInstructions to widened recipes.8488    for (VPRecipeBase &R : make_early_inc_range(*VPBB)) {8489      auto *SingleDef = cast<VPSingleDefRecipe>(&R);8490      auto *UnderlyingValue = SingleDef->getUnderlyingValue();8491      // Skip recipes that do not need transforming, including canonical IV,8492      // wide canonical IV and VPInstructions without underlying values. The8493      // latter are added above for masking.8494      // FIXME: Migrate code relying on the underlying instruction from VPlan08495      // to construct recipes below to not use the underlying instruction.8496      if (isa<VPCanonicalIVPHIRecipe, VPWidenCanonicalIVRecipe, VPBlendRecipe>(8497              &R) ||8498          (isa<VPInstruction>(&R) && !UnderlyingValue))8499        continue;8500      assert(isa<VPInstruction>(&R) && UnderlyingValue && "unsupported recipe");8501 8502      // TODO: Gradually replace uses of underlying instruction by analyses on8503      // VPlan.8504      Instruction *Instr = cast<Instruction>(UnderlyingValue);8505      Builder.setInsertPoint(SingleDef);8506 8507      // The stores with invariant address inside the loop will be deleted, and8508      // in the exit block, a uniform store recipe will be created for the final8509      // invariant store of the reduction.8510      StoreInst *SI;8511      if ((SI = dyn_cast<StoreInst>(Instr)) &&8512          Legal->isInvariantAddressOfReduction(SI->getPointerOperand())) {8513        // Only create recipe for the final invariant store of the reduction.8514        if (Legal->isInvariantStoreOfReduction(SI)) {8515          auto *VPI = cast<VPInstruction>(SingleDef);8516          auto *Recipe = new VPReplicateRecipe(8517              SI, R.operands(), true /* IsUniform */, nullptr /*Mask*/, *VPI,8518              *VPI, VPI->getDebugLoc());8519          Recipe->insertBefore(*MiddleVPBB, MBIP);8520        }8521        R.eraseFromParent();8522        continue;8523      }8524 8525      VPRecipeBase *Recipe =8526          RecipeBuilder.tryToCreateWidenRecipe(SingleDef, Range);8527      if (!Recipe)8528        Recipe = RecipeBuilder.handleReplication(cast<VPInstruction>(SingleDef),8529                                                 Range);8530 8531      RecipeBuilder.setRecipe(Instr, Recipe);8532      if (isa<VPWidenIntOrFpInductionRecipe>(Recipe) && isa<TruncInst>(Instr)) {8533        // Optimized a truncate to VPWidenIntOrFpInductionRecipe. It needs to be8534        // moved to the phi section in the header.8535        Recipe->insertBefore(*HeaderVPBB, HeaderVPBB->getFirstNonPhi());8536      } else {8537        Builder.insert(Recipe);8538      }8539      if (Recipe->getNumDefinedValues() == 1) {8540        SingleDef->replaceAllUsesWith(Recipe->getVPSingleValue());8541        Old2New[SingleDef] = Recipe->getVPSingleValue();8542      } else {8543        assert(Recipe->getNumDefinedValues() == 0 &&8544               "Unexpected multidef recipe");8545        R.eraseFromParent();8546      }8547    }8548  }8549 8550  // replaceAllUsesWith above may invalidate the block masks. Update them here.8551  // TODO: Include the masks as operands in the predicated VPlan directly8552  // to remove the need to keep a map of masks beyond the predication8553  // transform.8554  RecipeBuilder.updateBlockMaskCache(Old2New);8555  for (VPValue *Old : Old2New.keys())8556    Old->getDefiningRecipe()->eraseFromParent();8557 8558  assert(isa<VPRegionBlock>(LoopRegion) &&8559         !LoopRegion->getEntryBasicBlock()->empty() &&8560         "entry block must be set to a VPRegionBlock having a non-empty entry "8561         "VPBasicBlock");8562 8563  // TODO: We can't call runPass on these transforms yet, due to verifier8564  // failures.8565  VPlanTransforms::addExitUsersForFirstOrderRecurrences(*Plan, Range);8566  DenseMap<VPValue *, VPValue *> IVEndValues;8567  VPlanTransforms::updateScalarResumePhis(*Plan, IVEndValues);8568 8569  // ---------------------------------------------------------------------------8570  // Transform initial VPlan: Apply previously taken decisions, in order, to8571  // bring the VPlan to its final state.8572  // ---------------------------------------------------------------------------8573 8574  // Adjust the recipes for any inloop reductions.8575  adjustRecipesForReductions(Plan, RecipeBuilder, Range.Start);8576 8577  // Apply mandatory transformation to handle reductions with multiple in-loop8578  // uses if possible, bail out otherwise.8579  if (!VPlanTransforms::runPass(VPlanTransforms::handleMultiUseReductions,8580                                *Plan))8581    return nullptr;8582  // Apply mandatory transformation to handle FP maxnum/minnum reduction with8583  // NaNs if possible, bail out otherwise.8584  if (!VPlanTransforms::runPass(VPlanTransforms::handleMaxMinNumReductions,8585                                *Plan))8586    return nullptr;8587 8588  // Transform recipes to abstract recipes if it is legal and beneficial and8589  // clamp the range for better cost estimation.8590  // TODO: Enable following transform when the EVL-version of extended-reduction8591  // and mulacc-reduction are implemented.8592  if (!CM.foldTailWithEVL()) {8593    VPCostContext CostCtx(CM.TTI, *CM.TLI, *Plan, CM, CM.CostKind,8594                          *CM.PSE.getSE(), OrigLoop);8595    VPlanTransforms::runPass(VPlanTransforms::convertToAbstractRecipes, *Plan,8596                             CostCtx, Range);8597  }8598 8599  for (ElementCount VF : Range)8600    Plan->addVF(VF);8601  Plan->setName("Initial VPlan");8602 8603  // Interleave memory: for each Interleave Group we marked earlier as relevant8604  // for this VPlan, replace the Recipes widening its memory instructions with a8605  // single VPInterleaveRecipe at its insertion point.8606  VPlanTransforms::runPass(VPlanTransforms::createInterleaveGroups, *Plan,8607                           InterleaveGroups, RecipeBuilder,8608                           CM.isScalarEpilogueAllowed());8609 8610  // Replace VPValues for known constant strides.8611  VPlanTransforms::runPass(VPlanTransforms::replaceSymbolicStrides, *Plan, PSE,8612                           Legal->getLAI()->getSymbolicStrides());8613 8614  auto BlockNeedsPredication = [this](BasicBlock *BB) {8615    return Legal->blockNeedsPredication(BB);8616  };8617  VPlanTransforms::runPass(VPlanTransforms::dropPoisonGeneratingRecipes, *Plan,8618                           BlockNeedsPredication);8619 8620  // Sink users of fixed-order recurrence past the recipe defining the previous8621  // value and introduce FirstOrderRecurrenceSplice VPInstructions.8622  if (!VPlanTransforms::runPass(VPlanTransforms::adjustFixedOrderRecurrences,8623                                *Plan, Builder))8624    return nullptr;8625 8626  if (useActiveLaneMask(Style)) {8627    // TODO: Move checks to VPlanTransforms::addActiveLaneMask once8628    // TailFoldingStyle is visible there.8629    bool ForControlFlow = useActiveLaneMaskForControlFlow(Style);8630    bool WithoutRuntimeCheck =8631        Style == TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck;8632    VPlanTransforms::addActiveLaneMask(*Plan, ForControlFlow,8633                                       WithoutRuntimeCheck);8634  }8635  VPlanTransforms::optimizeInductionExitUsers(*Plan, IVEndValues, *PSE.getSE());8636 8637  assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid");8638  return Plan;8639}8640 8641VPlanPtr LoopVectorizationPlanner::tryToBuildVPlan(VFRange &Range) {8642  // Outer loop handling: They may require CFG and instruction level8643  // transformations before even evaluating whether vectorization is profitable.8644  // Since we cannot modify the incoming IR, we need to build VPlan upfront in8645  // the vectorization pipeline.8646  assert(!OrigLoop->isInnermost());8647  assert(EnableVPlanNativePath && "VPlan-native path is not enabled.");8648 8649  auto Plan = VPlanTransforms::buildVPlan0(8650      OrigLoop, *LI, Legal->getWidestInductionType(),8651      getDebugLocFromInstOrOperands(Legal->getPrimaryInduction()), PSE);8652  VPlanTransforms::handleEarlyExits(*Plan,8653                                    /*HasUncountableExit*/ false);8654  VPlanTransforms::addMiddleCheck(*Plan, /*RequiresScalarEpilogue*/ true,8655                                  /*TailFolded*/ false);8656 8657  VPlanTransforms::createLoopRegions(*Plan);8658 8659  for (ElementCount VF : Range)8660    Plan->addVF(VF);8661 8662  if (!VPlanTransforms::tryToConvertVPInstructionsToVPRecipes(8663          *Plan,8664          [this](PHINode *P) {8665            return Legal->getIntOrFpInductionDescriptor(P);8666          },8667          *TLI))8668    return nullptr;8669 8670  // TODO: IVEndValues are not used yet in the native path, to optimize exit8671  // values.8672  // TODO: We can't call runPass on the transform yet, due to verifier8673  // failures.8674  DenseMap<VPValue *, VPValue *> IVEndValues;8675  VPlanTransforms::updateScalarResumePhis(*Plan, IVEndValues);8676 8677  assert(verifyVPlanIsValid(*Plan) && "VPlan is invalid");8678  return Plan;8679}8680 8681// Adjust the recipes for reductions. For in-loop reductions the chain of8682// instructions leading from the loop exit instr to the phi need to be converted8683// to reductions, with one operand being vector and the other being the scalar8684// reduction chain. For other reductions, a select is introduced between the phi8685// and users outside the vector region when folding the tail.8686//8687// A ComputeReductionResult recipe is added to the middle block, also for8688// in-loop reductions which compute their result in-loop, because generating8689// the subsequent bc.merge.rdx phi is driven by ComputeReductionResult recipes.8690//8691// Adjust AnyOf reductions; replace the reduction phi for the selected value8692// with a boolean reduction phi node to check if the condition is true in any8693// iteration. The final value is selected by the final ComputeReductionResult.8694void LoopVectorizationPlanner::adjustRecipesForReductions(8695    VPlanPtr &Plan, VPRecipeBuilder &RecipeBuilder, ElementCount MinVF) {8696  using namespace VPlanPatternMatch;8697  VPTypeAnalysis TypeInfo(*Plan);8698  VPRegionBlock *VectorLoopRegion = Plan->getVectorLoopRegion();8699  VPBasicBlock *Header = VectorLoopRegion->getEntryBasicBlock();8700  VPBasicBlock *MiddleVPBB = Plan->getMiddleBlock();8701  SmallVector<VPRecipeBase *> ToDelete;8702 8703  for (VPRecipeBase &R : Header->phis()) {8704    auto *PhiR = dyn_cast<VPReductionPHIRecipe>(&R);8705    if (!PhiR || !PhiR->isInLoop() || (MinVF.isScalar() && !PhiR->isOrdered()))8706      continue;8707 8708    RecurKind Kind = PhiR->getRecurrenceKind();8709    assert(8710        !RecurrenceDescriptor::isAnyOfRecurrenceKind(Kind) &&8711        !RecurrenceDescriptor::isFindIVRecurrenceKind(Kind) &&8712        "AnyOf and FindIV reductions are not allowed for in-loop reductions");8713 8714    bool IsFPRecurrence =8715        RecurrenceDescriptor::isFloatingPointRecurrenceKind(Kind);8716    FastMathFlags FMFs =8717        IsFPRecurrence ? FastMathFlags::getFast() : FastMathFlags();8718 8719    // Collect the chain of "link" recipes for the reduction starting at PhiR.8720    SetVector<VPSingleDefRecipe *> Worklist;8721    Worklist.insert(PhiR);8722    for (unsigned I = 0; I != Worklist.size(); ++I) {8723      VPSingleDefRecipe *Cur = Worklist[I];8724      for (VPUser *U : Cur->users()) {8725        auto *UserRecipe = cast<VPSingleDefRecipe>(U);8726        if (!UserRecipe->getParent()->getEnclosingLoopRegion()) {8727          assert((UserRecipe->getParent() == MiddleVPBB ||8728                  UserRecipe->getParent() == Plan->getScalarPreheader()) &&8729                 "U must be either in the loop region, the middle block or the "8730                 "scalar preheader.");8731          continue;8732        }8733        Worklist.insert(UserRecipe);8734      }8735    }8736 8737    // Visit operation "Links" along the reduction chain top-down starting from8738    // the phi until LoopExitValue. We keep track of the previous item8739    // (PreviousLink) to tell which of the two operands of a Link will remain8740    // scalar and which will be reduced. For minmax by select(cmp), Link will be8741    // the select instructions. Blend recipes of in-loop reduction phi's  will8742    // get folded to their non-phi operand, as the reduction recipe handles the8743    // condition directly.8744    VPSingleDefRecipe *PreviousLink = PhiR; // Aka Worklist[0].8745    for (VPSingleDefRecipe *CurrentLink : drop_begin(Worklist)) {8746      if (auto *Blend = dyn_cast<VPBlendRecipe>(CurrentLink)) {8747        assert(Blend->getNumIncomingValues() == 2 &&8748               "Blend must have 2 incoming values");8749        if (Blend->getIncomingValue(0) == PhiR) {8750          Blend->replaceAllUsesWith(Blend->getIncomingValue(1));8751        } else {8752          assert(Blend->getIncomingValue(1) == PhiR &&8753                 "PhiR must be an operand of the blend");8754          Blend->replaceAllUsesWith(Blend->getIncomingValue(0));8755        }8756        continue;8757      }8758 8759      if (IsFPRecurrence) {8760        FastMathFlags CurFMF =8761            cast<VPRecipeWithIRFlags>(CurrentLink)->getFastMathFlags();8762        if (match(CurrentLink, m_Select(m_VPValue(), m_VPValue(), m_VPValue())))8763          CurFMF |= cast<VPRecipeWithIRFlags>(CurrentLink->getOperand(0))8764                        ->getFastMathFlags();8765        FMFs &= CurFMF;8766      }8767 8768      Instruction *CurrentLinkI = CurrentLink->getUnderlyingInstr();8769 8770      // Index of the first operand which holds a non-mask vector operand.8771      unsigned IndexOfFirstOperand;8772      // Recognize a call to the llvm.fmuladd intrinsic.8773      bool IsFMulAdd = (Kind == RecurKind::FMulAdd);8774      VPValue *VecOp;8775      VPBasicBlock *LinkVPBB = CurrentLink->getParent();8776      if (IsFMulAdd) {8777        assert(8778            RecurrenceDescriptor::isFMulAddIntrinsic(CurrentLinkI) &&8779            "Expected instruction to be a call to the llvm.fmuladd intrinsic");8780        assert(((MinVF.isScalar() && isa<VPReplicateRecipe>(CurrentLink)) ||8781                isa<VPWidenIntrinsicRecipe>(CurrentLink)) &&8782               CurrentLink->getOperand(2) == PreviousLink &&8783               "expected a call where the previous link is the added operand");8784 8785        // If the instruction is a call to the llvm.fmuladd intrinsic then we8786        // need to create an fmul recipe (multiplying the first two operands of8787        // the fmuladd together) to use as the vector operand for the fadd8788        // reduction.8789        VPInstruction *FMulRecipe = new VPInstruction(8790            Instruction::FMul,8791            {CurrentLink->getOperand(0), CurrentLink->getOperand(1)},8792            CurrentLinkI->getFastMathFlags());8793        LinkVPBB->insert(FMulRecipe, CurrentLink->getIterator());8794        VecOp = FMulRecipe;8795      } else if (PhiR->isInLoop() && Kind == RecurKind::AddChainWithSubs &&8796                 match(CurrentLink, m_Sub(m_VPValue(), m_VPValue()))) {8797        Type *PhiTy = TypeInfo.inferScalarType(PhiR);8798        auto *Zero = Plan->getConstantInt(PhiTy, 0);8799        VPWidenRecipe *Sub = new VPWidenRecipe(8800            Instruction::Sub, {Zero, CurrentLink->getOperand(1)}, {},8801            VPIRMetadata(), CurrentLinkI->getDebugLoc());8802        Sub->setUnderlyingValue(CurrentLinkI);8803        LinkVPBB->insert(Sub, CurrentLink->getIterator());8804        VecOp = Sub;8805      } else {8806        if (RecurrenceDescriptor::isMinMaxRecurrenceKind(Kind)) {8807          if (match(CurrentLink, m_Cmp(m_VPValue(), m_VPValue())))8808            continue;8809          assert(isa<VPWidenSelectRecipe>(CurrentLink) &&8810                 "must be a select recipe");8811          IndexOfFirstOperand = 1;8812        } else {8813          assert((MinVF.isScalar() || isa<VPWidenRecipe>(CurrentLink)) &&8814                 "Expected to replace a VPWidenSC");8815          IndexOfFirstOperand = 0;8816        }8817        // Note that for non-commutable operands (cmp-selects), the semantics of8818        // the cmp-select are captured in the recurrence kind.8819        unsigned VecOpId =8820            CurrentLink->getOperand(IndexOfFirstOperand) == PreviousLink8821                ? IndexOfFirstOperand + 18822                : IndexOfFirstOperand;8823        VecOp = CurrentLink->getOperand(VecOpId);8824        assert(VecOp != PreviousLink &&8825               CurrentLink->getOperand(CurrentLink->getNumOperands() - 1 -8826                                       (VecOpId - IndexOfFirstOperand)) ==8827                   PreviousLink &&8828               "PreviousLink must be the operand other than VecOp");8829      }8830 8831      VPValue *CondOp = nullptr;8832      if (CM.blockNeedsPredicationForAnyReason(CurrentLinkI->getParent()))8833        CondOp = RecipeBuilder.getBlockInMask(CurrentLink->getParent());8834 8835      ReductionStyle Style = getReductionStyle(true, PhiR->isOrdered(), 1);8836      auto *RedRecipe =8837          new VPReductionRecipe(Kind, FMFs, CurrentLinkI, PreviousLink, VecOp,8838                                CondOp, Style, CurrentLinkI->getDebugLoc());8839      // Append the recipe to the end of the VPBasicBlock because we need to8840      // ensure that it comes after all of it's inputs, including CondOp.8841      // Delete CurrentLink as it will be invalid if its operand is replaced8842      // with a reduction defined at the bottom of the block in the next link.8843      if (LinkVPBB->getNumSuccessors() == 0)8844        RedRecipe->insertBefore(&*std::prev(std::prev(LinkVPBB->end())));8845      else8846        LinkVPBB->appendRecipe(RedRecipe);8847 8848      CurrentLink->replaceAllUsesWith(RedRecipe);8849      ToDelete.push_back(CurrentLink);8850      PreviousLink = RedRecipe;8851    }8852  }8853  VPBasicBlock *LatchVPBB = VectorLoopRegion->getExitingBasicBlock();8854  Builder.setInsertPoint(&*std::prev(std::prev(LatchVPBB->end())));8855  VPBasicBlock::iterator IP = MiddleVPBB->getFirstNonPhi();8856  for (VPRecipeBase &R :8857       Plan->getVectorLoopRegion()->getEntryBasicBlock()->phis()) {8858    VPReductionPHIRecipe *PhiR = dyn_cast<VPReductionPHIRecipe>(&R);8859    if (!PhiR)8860      continue;8861 8862    const RecurrenceDescriptor &RdxDesc = Legal->getRecurrenceDescriptor(8863        cast<PHINode>(PhiR->getUnderlyingInstr()));8864    Type *PhiTy = TypeInfo.inferScalarType(PhiR);8865    // If tail is folded by masking, introduce selects between the phi8866    // and the users outside the vector region of each reduction, at the8867    // beginning of the dedicated latch block.8868    auto *OrigExitingVPV = PhiR->getBackedgeValue();8869    auto *NewExitingVPV = PhiR->getBackedgeValue();8870    // Don't output selects for partial reductions because they have an output8871    // with fewer lanes than the VF. So the operands of the select would have8872    // different numbers of lanes. Partial reductions mask the input instead.8873    auto *RR = dyn_cast<VPReductionRecipe>(OrigExitingVPV->getDefiningRecipe());8874    if (!PhiR->isInLoop() && CM.foldTailByMasking() &&8875        (!RR || !RR->isPartialReduction())) {8876      VPValue *Cond = RecipeBuilder.getBlockInMask(PhiR->getParent());8877      std::optional<FastMathFlags> FMFs =8878          PhiTy->isFloatingPointTy()8879              ? std::make_optional(RdxDesc.getFastMathFlags())8880              : std::nullopt;8881      NewExitingVPV =8882          Builder.createSelect(Cond, OrigExitingVPV, PhiR, {}, "", FMFs);8883      OrigExitingVPV->replaceUsesWithIf(NewExitingVPV, [](VPUser &U, unsigned) {8884        return isa<VPInstruction>(&U) &&8885               (cast<VPInstruction>(&U)->getOpcode() ==8886                    VPInstruction::ComputeAnyOfResult ||8887                cast<VPInstruction>(&U)->getOpcode() ==8888                    VPInstruction::ComputeReductionResult ||8889                cast<VPInstruction>(&U)->getOpcode() ==8890                    VPInstruction::ComputeFindIVResult);8891      });8892      if (CM.usePredicatedReductionSelect())8893        PhiR->setOperand(1, NewExitingVPV);8894    }8895 8896    // We want code in the middle block to appear to execute on the location of8897    // the scalar loop's latch terminator because: (a) it is all compiler8898    // generated, (b) these instructions are always executed after evaluating8899    // the latch conditional branch, and (c) other passes may add new8900    // predecessors which terminate on this line. This is the easiest way to8901    // ensure we don't accidentally cause an extra step back into the loop while8902    // debugging.8903    DebugLoc ExitDL = OrigLoop->getLoopLatch()->getTerminator()->getDebugLoc();8904 8905    // TODO: At the moment ComputeReductionResult also drives creation of the8906    // bc.merge.rdx phi nodes, hence it needs to be created unconditionally here8907    // even for in-loop reductions, until the reduction resume value handling is8908    // also modeled in VPlan.8909    VPInstruction *FinalReductionResult;8910    VPBuilder::InsertPointGuard Guard(Builder);8911    Builder.setInsertPoint(MiddleVPBB, IP);8912    RecurKind RecurrenceKind = PhiR->getRecurrenceKind();8913    if (RecurrenceDescriptor::isFindIVRecurrenceKind(RecurrenceKind)) {8914      VPValue *Start = PhiR->getStartValue();8915      VPValue *Sentinel = Plan->getOrAddLiveIn(RdxDesc.getSentinelValue());8916      FinalReductionResult =8917          Builder.createNaryOp(VPInstruction::ComputeFindIVResult,8918                               {PhiR, Start, Sentinel, NewExitingVPV}, ExitDL);8919    } else if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RecurrenceKind)) {8920      VPValue *Start = PhiR->getStartValue();8921      FinalReductionResult =8922          Builder.createNaryOp(VPInstruction::ComputeAnyOfResult,8923                               {PhiR, Start, NewExitingVPV}, ExitDL);8924    } else {8925      VPIRFlags Flags =8926          RecurrenceDescriptor::isFloatingPointRecurrenceKind(RecurrenceKind)8927              ? VPIRFlags(RdxDesc.getFastMathFlags())8928              : VPIRFlags();8929      FinalReductionResult =8930          Builder.createNaryOp(VPInstruction::ComputeReductionResult,8931                               {PhiR, NewExitingVPV}, Flags, ExitDL);8932    }8933    // If the vector reduction can be performed in a smaller type, we truncate8934    // then extend the loop exit value to enable InstCombine to evaluate the8935    // entire expression in the smaller type.8936    if (MinVF.isVector() && PhiTy != RdxDesc.getRecurrenceType() &&8937        !RecurrenceDescriptor::isAnyOfRecurrenceKind(RecurrenceKind)) {8938      assert(!PhiR->isInLoop() && "Unexpected truncated inloop reduction!");8939      assert(!RecurrenceDescriptor::isMinMaxRecurrenceKind(RecurrenceKind) &&8940             "Unexpected truncated min-max recurrence!");8941      Type *RdxTy = RdxDesc.getRecurrenceType();8942      VPWidenCastRecipe *Trunc;8943      Instruction::CastOps ExtendOpc =8944          RdxDesc.isSigned() ? Instruction::SExt : Instruction::ZExt;8945      VPWidenCastRecipe *Extnd;8946      {8947        VPBuilder::InsertPointGuard Guard(Builder);8948        Builder.setInsertPoint(8949            NewExitingVPV->getDefiningRecipe()->getParent(),8950            std::next(NewExitingVPV->getDefiningRecipe()->getIterator()));8951        Trunc =8952            Builder.createWidenCast(Instruction::Trunc, NewExitingVPV, RdxTy);8953        Extnd = Builder.createWidenCast(ExtendOpc, Trunc, PhiTy);8954      }8955      if (PhiR->getOperand(1) == NewExitingVPV)8956        PhiR->setOperand(1, Extnd->getVPSingleValue());8957 8958      // Update ComputeReductionResult with the truncated exiting value and8959      // extend its result.8960      FinalReductionResult->setOperand(1, Trunc);8961      FinalReductionResult =8962          Builder.createScalarCast(ExtendOpc, FinalReductionResult, PhiTy, {});8963    }8964 8965    // Update all users outside the vector region. Also replace redundant8966    // ExtractLastElement.8967    for (auto *U : to_vector(OrigExitingVPV->users())) {8968      auto *Parent = cast<VPRecipeBase>(U)->getParent();8969      if (FinalReductionResult == U || Parent->getParent())8970        continue;8971      U->replaceUsesOfWith(OrigExitingVPV, FinalReductionResult);8972      if (match(U, m_CombineOr(m_ExtractLastElement(m_VPValue()),8973                               m_ExtractLane(m_VPValue(), m_VPValue()))))8974        cast<VPInstruction>(U)->replaceAllUsesWith(FinalReductionResult);8975    }8976 8977    // Adjust AnyOf reductions; replace the reduction phi for the selected value8978    // with a boolean reduction phi node to check if the condition is true in8979    // any iteration. The final value is selected by the final8980    // ComputeReductionResult.8981    if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RecurrenceKind)) {8982      auto *Select = cast<VPRecipeBase>(*find_if(PhiR->users(), [](VPUser *U) {8983        return isa<VPWidenSelectRecipe>(U) ||8984               (isa<VPReplicateRecipe>(U) &&8985                cast<VPReplicateRecipe>(U)->getUnderlyingInstr()->getOpcode() ==8986                    Instruction::Select);8987      }));8988      VPValue *Cmp = Select->getOperand(0);8989      // If the compare is checking the reduction PHI node, adjust it to check8990      // the start value.8991      if (VPRecipeBase *CmpR = Cmp->getDefiningRecipe())8992        CmpR->replaceUsesOfWith(PhiR, PhiR->getStartValue());8993      Builder.setInsertPoint(Select);8994 8995      // If the true value of the select is the reduction phi, the new value is8996      // selected if the negated condition is true in any iteration.8997      if (Select->getOperand(1) == PhiR)8998        Cmp = Builder.createNot(Cmp);8999      VPValue *Or = Builder.createOr(PhiR, Cmp);9000      Select->getVPSingleValue()->replaceAllUsesWith(Or);9001      // Delete Select now that it has invalid types.9002      ToDelete.push_back(Select);9003 9004      // Convert the reduction phi to operate on bools.9005      PhiR->setOperand(0, Plan->getFalse());9006      continue;9007    }9008 9009    if (RecurrenceDescriptor::isFindIVRecurrenceKind(9010            RdxDesc.getRecurrenceKind())) {9011      // Adjust the start value for FindFirstIV/FindLastIV recurrences to use9012      // the sentinel value after generating the ResumePhi recipe, which uses9013      // the original start value.9014      PhiR->setOperand(0, Plan->getOrAddLiveIn(RdxDesc.getSentinelValue()));9015    }9016    RecurKind RK = RdxDesc.getRecurrenceKind();9017    if ((!RecurrenceDescriptor::isAnyOfRecurrenceKind(RK) &&9018         !RecurrenceDescriptor::isFindIVRecurrenceKind(RK) &&9019         !RecurrenceDescriptor::isMinMaxRecurrenceKind(RK))) {9020      VPBuilder PHBuilder(Plan->getVectorPreheader());9021      VPValue *Iden = Plan->getOrAddLiveIn(9022          getRecurrenceIdentity(RK, PhiTy, RdxDesc.getFastMathFlags()));9023      // If the PHI is used by a partial reduction, set the scale factor.9024      unsigned ScaleFactor =9025          RecipeBuilder.getScalingForReduction(RdxDesc.getLoopExitInstr())9026              .value_or(1);9027      auto *ScaleFactorVPV = Plan->getConstantInt(32, ScaleFactor);9028      VPValue *StartV = PHBuilder.createNaryOp(9029          VPInstruction::ReductionStartVector,9030          {PhiR->getStartValue(), Iden, ScaleFactorVPV},9031          PhiTy->isFloatingPointTy() ? RdxDesc.getFastMathFlags()9032                                     : FastMathFlags());9033      PhiR->setOperand(0, StartV);9034    }9035  }9036  for (VPRecipeBase *R : ToDelete)9037    R->eraseFromParent();9038 9039  VPlanTransforms::runPass(VPlanTransforms::clearReductionWrapFlags, *Plan);9040}9041 9042void LoopVectorizationPlanner::attachRuntimeChecks(9043    VPlan &Plan, GeneratedRTChecks &RTChecks, bool HasBranchWeights) const {9044  const auto &[SCEVCheckCond, SCEVCheckBlock] = RTChecks.getSCEVChecks();9045  if (SCEVCheckBlock && SCEVCheckBlock->hasNPredecessors(0)) {9046    assert((!CM.OptForSize ||9047            CM.Hints->getForce() == LoopVectorizeHints::FK_Enabled) &&9048           "Cannot SCEV check stride or overflow when optimizing for size");9049    VPlanTransforms::attachCheckBlock(Plan, SCEVCheckCond, SCEVCheckBlock,9050                                      HasBranchWeights);9051  }9052  const auto &[MemCheckCond, MemCheckBlock] = RTChecks.getMemRuntimeChecks();9053  if (MemCheckBlock && MemCheckBlock->hasNPredecessors(0)) {9054    // VPlan-native path does not do any analysis for runtime checks9055    // currently.9056    assert((!EnableVPlanNativePath || OrigLoop->isInnermost()) &&9057           "Runtime checks are not supported for outer loops yet");9058 9059    if (CM.OptForSize) {9060      assert(9061          CM.Hints->getForce() == LoopVectorizeHints::FK_Enabled &&9062          "Cannot emit memory checks when optimizing for size, unless forced "9063          "to vectorize.");9064      ORE->emit([&]() {9065        return OptimizationRemarkAnalysis(DEBUG_TYPE, "VectorizationCodeSize",9066                                          OrigLoop->getStartLoc(),9067                                          OrigLoop->getHeader())9068               << "Code-size may be reduced by not forcing "9069                  "vectorization, or by source-code modifications "9070                  "eliminating the need for runtime checks "9071                  "(e.g., adding 'restrict').";9072      });9073    }9074    VPlanTransforms::attachCheckBlock(Plan, MemCheckCond, MemCheckBlock,9075                                      HasBranchWeights);9076  }9077}9078 9079void LoopVectorizationPlanner::addMinimumIterationCheck(9080    VPlan &Plan, ElementCount VF, unsigned UF,9081    ElementCount MinProfitableTripCount) const {9082  // vscale is not necessarily a power-of-2, which means we cannot guarantee9083  // an overflow to zero when updating induction variables and so an9084  // additional overflow check is required before entering the vector loop.9085  bool IsIndvarOverflowCheckNeededForVF =9086      VF.isScalable() && !TTI.isVScaleKnownToBeAPowerOfTwo() &&9087      !isIndvarOverflowCheckKnownFalse(&CM, VF, UF) &&9088      CM.getTailFoldingStyle() !=9089          TailFoldingStyle::DataAndControlFlowWithoutRuntimeCheck;9090  const uint32_t *BranchWeigths =9091      hasBranchWeightMD(*OrigLoop->getLoopLatch()->getTerminator())9092          ? &MinItersBypassWeights[0]9093          : nullptr;9094  VPlanTransforms::addMinimumIterationCheck(9095      Plan, VF, UF, MinProfitableTripCount,9096      CM.requiresScalarEpilogue(VF.isVector()), CM.foldTailByMasking(),9097      IsIndvarOverflowCheckNeededForVF, OrigLoop, BranchWeigths,9098      OrigLoop->getLoopPredecessor()->getTerminator()->getDebugLoc(),9099      *PSE.getSE());9100}9101 9102void VPDerivedIVRecipe::execute(VPTransformState &State) {9103  assert(!State.Lane && "VPDerivedIVRecipe being replicated.");9104 9105  // Fast-math-flags propagate from the original induction instruction.9106  IRBuilder<>::FastMathFlagGuard FMFG(State.Builder);9107  if (FPBinOp)9108    State.Builder.setFastMathFlags(FPBinOp->getFastMathFlags());9109 9110  Value *Step = State.get(getStepValue(), VPLane(0));9111  Value *Index = State.get(getOperand(1), VPLane(0));9112  Value *DerivedIV = emitTransformedIndex(9113      State.Builder, Index, getStartValue()->getLiveInIRValue(), Step, Kind,9114      cast_if_present<BinaryOperator>(FPBinOp));9115  DerivedIV->setName(Name);9116  State.set(this, DerivedIV, VPLane(0));9117}9118 9119// Determine how to lower the scalar epilogue, which depends on 1) optimising9120// for minimum code-size, 2) predicate compiler options, 3) loop hints forcing9121// predication, and 4) a TTI hook that analyses whether the loop is suitable9122// for predication.9123static ScalarEpilogueLowering getScalarEpilogueLowering(9124    Function *F, Loop *L, LoopVectorizeHints &Hints, bool OptForSize,9125    TargetTransformInfo *TTI, TargetLibraryInfo *TLI,9126    LoopVectorizationLegality &LVL, InterleavedAccessInfo *IAI) {9127  // 1) OptSize takes precedence over all other options, i.e. if this is set,9128  // don't look at hints or options, and don't request a scalar epilogue.9129  if (F->hasOptSize() ||9130      (OptForSize && Hints.getForce() != LoopVectorizeHints::FK_Enabled))9131    return CM_ScalarEpilogueNotAllowedOptSize;9132 9133  // 2) If set, obey the directives9134  if (PreferPredicateOverEpilogue.getNumOccurrences()) {9135    switch (PreferPredicateOverEpilogue) {9136    case PreferPredicateTy::ScalarEpilogue:9137      return CM_ScalarEpilogueAllowed;9138    case PreferPredicateTy::PredicateElseScalarEpilogue:9139      return CM_ScalarEpilogueNotNeededUsePredicate;9140    case PreferPredicateTy::PredicateOrDontVectorize:9141      return CM_ScalarEpilogueNotAllowedUsePredicate;9142    };9143  }9144 9145  // 3) If set, obey the hints9146  switch (Hints.getPredicate()) {9147  case LoopVectorizeHints::FK_Enabled:9148    return CM_ScalarEpilogueNotNeededUsePredicate;9149  case LoopVectorizeHints::FK_Disabled:9150    return CM_ScalarEpilogueAllowed;9151  };9152 9153  // 4) if the TTI hook indicates this is profitable, request predication.9154  TailFoldingInfo TFI(TLI, &LVL, IAI);9155  if (TTI->preferPredicateOverEpilogue(&TFI))9156    return CM_ScalarEpilogueNotNeededUsePredicate;9157 9158  return CM_ScalarEpilogueAllowed;9159}9160 9161// Process the loop in the VPlan-native vectorization path. This path builds9162// VPlan upfront in the vectorization pipeline, which allows to apply9163// VPlan-to-VPlan transformations from the very beginning without modifying the9164// input LLVM IR.9165static bool processLoopInVPlanNativePath(9166    Loop *L, PredicatedScalarEvolution &PSE, LoopInfo *LI, DominatorTree *DT,9167    LoopVectorizationLegality *LVL, TargetTransformInfo *TTI,9168    TargetLibraryInfo *TLI, DemandedBits *DB, AssumptionCache *AC,9169    OptimizationRemarkEmitter *ORE, bool OptForSize, LoopVectorizeHints &Hints,9170    LoopVectorizationRequirements &Requirements) {9171 9172  if (isa<SCEVCouldNotCompute>(PSE.getBackedgeTakenCount())) {9173    LLVM_DEBUG(dbgs() << "LV: cannot compute the outer-loop trip count\n");9174    return false;9175  }9176  assert(EnableVPlanNativePath && "VPlan-native path is disabled.");9177  Function *F = L->getHeader()->getParent();9178  InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL->getLAI());9179 9180  ScalarEpilogueLowering SEL =9181      getScalarEpilogueLowering(F, L, Hints, OptForSize, TTI, TLI, *LVL, &IAI);9182 9183  LoopVectorizationCostModel CM(SEL, L, PSE, LI, LVL, *TTI, TLI, DB, AC, ORE, F,9184                                &Hints, IAI, OptForSize);9185  // Use the planner for outer loop vectorization.9186  // TODO: CM is not used at this point inside the planner. Turn CM into an9187  // optional argument if we don't need it in the future.9188  LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, LVL, CM, IAI, PSE, Hints,9189                               ORE);9190 9191  // Get user vectorization factor.9192  ElementCount UserVF = Hints.getWidth();9193 9194  CM.collectElementTypesForWidening();9195 9196  // Plan how to best vectorize, return the best VF and its cost.9197  const VectorizationFactor VF = LVP.planInVPlanNativePath(UserVF);9198 9199  // If we are stress testing VPlan builds, do not attempt to generate vector9200  // code. Masked vector code generation support will follow soon.9201  // Also, do not attempt to vectorize if no vector code will be produced.9202  if (VPlanBuildStressTest || VectorizationFactor::Disabled() == VF)9203    return false;9204 9205  VPlan &BestPlan = LVP.getPlanFor(VF.Width);9206 9207  {9208    GeneratedRTChecks Checks(PSE, DT, LI, TTI, F->getDataLayout(), CM.CostKind);9209    InnerLoopVectorizer LB(L, PSE, LI, DT, TTI, AC, VF.Width, /*UF=*/1, &CM,9210                           Checks, BestPlan);9211    LLVM_DEBUG(dbgs() << "Vectorizing outer loop in \""9212                      << L->getHeader()->getParent()->getName() << "\"\n");9213    LVP.addMinimumIterationCheck(BestPlan, VF.Width, /*UF=*/1,9214                                 VF.MinProfitableTripCount);9215 9216    LVP.executePlan(VF.Width, /*UF=*/1, BestPlan, LB, DT, false);9217  }9218 9219  reportVectorization(ORE, L, VF, 1);9220 9221  assert(!verifyFunction(*L->getHeader()->getParent(), &dbgs()));9222  return true;9223}9224 9225// Emit a remark if there are stores to floats that required a floating point9226// extension. If the vectorized loop was generated with floating point there9227// will be a performance penalty from the conversion overhead and the change in9228// the vector width.9229static void checkMixedPrecision(Loop *L, OptimizationRemarkEmitter *ORE) {9230  SmallVector<Instruction *, 4> Worklist;9231  for (BasicBlock *BB : L->getBlocks()) {9232    for (Instruction &Inst : *BB) {9233      if (auto *S = dyn_cast<StoreInst>(&Inst)) {9234        if (S->getValueOperand()->getType()->isFloatTy())9235          Worklist.push_back(S);9236      }9237    }9238  }9239 9240  // Traverse the floating point stores upwards searching, for floating point9241  // conversions.9242  SmallPtrSet<const Instruction *, 4> Visited;9243  SmallPtrSet<const Instruction *, 4> EmittedRemark;9244  while (!Worklist.empty()) {9245    auto *I = Worklist.pop_back_val();9246    if (!L->contains(I))9247      continue;9248    if (!Visited.insert(I).second)9249      continue;9250 9251    // Emit a remark if the floating point store required a floating9252    // point conversion.9253    // TODO: More work could be done to identify the root cause such as a9254    // constant or a function return type and point the user to it.9255    if (isa<FPExtInst>(I) && EmittedRemark.insert(I).second)9256      ORE->emit([&]() {9257        return OptimizationRemarkAnalysis(LV_NAME, "VectorMixedPrecision",9258                                          I->getDebugLoc(), L->getHeader())9259               << "floating point conversion changes vector width. "9260               << "Mixed floating point precision requires an up/down "9261               << "cast that will negatively impact performance.";9262      });9263 9264    for (Use &Op : I->operands())9265      if (auto *OpI = dyn_cast<Instruction>(Op))9266        Worklist.push_back(OpI);9267  }9268}9269 9270/// For loops with uncountable early exits, find the cost of doing work when9271/// exiting the loop early, such as calculating the final exit values of9272/// variables used outside the loop.9273/// TODO: This is currently overly pessimistic because the loop may not take9274/// the early exit, but better to keep this conservative for now. In future,9275/// it might be possible to relax this by using branch probabilities.9276static InstructionCost calculateEarlyExitCost(VPCostContext &CostCtx,9277                                              VPlan &Plan, ElementCount VF) {9278  InstructionCost Cost = 0;9279  for (auto *ExitVPBB : Plan.getExitBlocks()) {9280    for (auto *PredVPBB : ExitVPBB->getPredecessors()) {9281      // If the predecessor is not the middle.block, then it must be the9282      // vector.early.exit block, which may contain work to calculate the exit9283      // values of variables used outside the loop.9284      if (PredVPBB != Plan.getMiddleBlock()) {9285        LLVM_DEBUG(dbgs() << "Calculating cost of work in exit block "9286                          << PredVPBB->getName() << ":\n");9287        Cost += PredVPBB->cost(VF, CostCtx);9288      }9289    }9290  }9291  return Cost;9292}9293 9294/// This function determines whether or not it's still profitable to vectorize9295/// the loop given the extra work we have to do outside of the loop:9296///  1. Perform the runtime checks before entering the loop to ensure it's safe9297///     to vectorize.9298///  2. In the case of loops with uncountable early exits, we may have to do9299///     extra work when exiting the loop early, such as calculating the final9300///     exit values of variables used outside the loop.9301///  3. The middle block, if expected TC <= VF.Width.9302static bool isOutsideLoopWorkProfitable(GeneratedRTChecks &Checks,9303                                        VectorizationFactor &VF, Loop *L,9304                                        PredicatedScalarEvolution &PSE,9305                                        VPCostContext &CostCtx, VPlan &Plan,9306                                        ScalarEpilogueLowering SEL,9307                                        std::optional<unsigned> VScale) {9308  InstructionCost TotalCost = Checks.getCost();9309  if (!TotalCost.isValid())9310    return false;9311 9312  // Add on the cost of any work required in the vector early exit block, if9313  // one exists.9314  TotalCost += calculateEarlyExitCost(CostCtx, Plan, VF.Width);9315 9316  // If the expected trip count is less than the VF, the vector loop will only9317  // execute a single iteration. Then the middle block is executed the same9318  // number of times as the vector region.9319  // TODO: Extend logic to always account for the cost of the middle block.9320  auto ExpectedTC = getSmallBestKnownTC(PSE, L);9321  if (ExpectedTC && ElementCount::isKnownLE(*ExpectedTC, VF.Width))9322    TotalCost += Plan.getMiddleBlock()->cost(VF.Width, CostCtx);9323 9324  // When interleaving only scalar and vector cost will be equal, which in turn9325  // would lead to a divide by 0. Fall back to hard threshold.9326  if (VF.Width.isScalar()) {9327    // TODO: Should we rename VectorizeMemoryCheckThreshold?9328    if (TotalCost > VectorizeMemoryCheckThreshold) {9329      LLVM_DEBUG(9330          dbgs()9331          << "LV: Interleaving only is not profitable due to runtime checks\n");9332      return false;9333    }9334    return true;9335  }9336 9337  // The scalar cost should only be 0 when vectorizing with a user specified9338  // VF/IC. In those cases, runtime checks should always be generated.9339  uint64_t ScalarC = VF.ScalarCost.getValue();9340  if (ScalarC == 0)9341    return true;9342 9343  // First, compute the minimum iteration count required so that the vector9344  // loop outperforms the scalar loop.9345  //  The total cost of the scalar loop is9346  //   ScalarC * TC9347  //  where9348  //  * TC is the actual trip count of the loop.9349  //  * ScalarC is the cost of a single scalar iteration.9350  //9351  //  The total cost of the vector loop is9352  //    RtC + VecC * (TC / VF) + EpiC9353  //  where9354  //  * RtC is the sum of the costs cost of9355  //    - the generated runtime checks9356  //    - performing any additional work in the vector.early.exit block for9357  //      loops with uncountable early exits.9358  //    - the middle block, if ExpectedTC <=  VF.Width.9359  //  * VecC is the cost of a single vector iteration.9360  //  * TC is the actual trip count of the loop9361  //  * VF is the vectorization factor9362  //  * EpiCost is the cost of the generated epilogue, including the cost9363  //    of the remaining scalar operations.9364  //9365  // Vectorization is profitable once the total vector cost is less than the9366  // total scalar cost:9367  //   RtC + VecC * (TC / VF) + EpiC <  ScalarC * TC9368  //9369  // Now we can compute the minimum required trip count TC as9370  //   VF * (RtC + EpiC) / (ScalarC * VF - VecC) < TC9371  //9372  // For now we assume the epilogue cost EpiC = 0 for simplicity. Note that9373  // the computations are performed on doubles, not integers and the result9374  // is rounded up, hence we get an upper estimate of the TC.9375  unsigned IntVF = estimateElementCount(VF.Width, VScale);9376  uint64_t RtC = TotalCost.getValue();9377  uint64_t Div = ScalarC * IntVF - VF.Cost.getValue();9378  uint64_t MinTC1 = Div == 0 ? 0 : divideCeil(RtC * IntVF, Div);9379 9380  // Second, compute a minimum iteration count so that the cost of the9381  // runtime checks is only a fraction of the total scalar loop cost. This9382  // adds a loop-dependent bound on the overhead incurred if the runtime9383  // checks fail. In case the runtime checks fail, the cost is RtC + ScalarC9384  // * TC. To bound the runtime check to be a fraction 1/X of the scalar9385  // cost, compute9386  //   RtC < ScalarC * TC * (1 / X)  ==>  RtC * X / ScalarC < TC9387  uint64_t MinTC2 = divideCeil(RtC * 10, ScalarC);9388 9389  // Now pick the larger minimum. If it is not a multiple of VF and a scalar9390  // epilogue is allowed, choose the next closest multiple of VF. This should9391  // partly compensate for ignoring the epilogue cost.9392  uint64_t MinTC = std::max(MinTC1, MinTC2);9393  if (SEL == CM_ScalarEpilogueAllowed)9394    MinTC = alignTo(MinTC, IntVF);9395  VF.MinProfitableTripCount = ElementCount::getFixed(MinTC);9396 9397  LLVM_DEBUG(9398      dbgs() << "LV: Minimum required TC for runtime checks to be profitable:"9399             << VF.MinProfitableTripCount << "\n");9400 9401  // Skip vectorization if the expected trip count is less than the minimum9402  // required trip count.9403  if (auto ExpectedTC = getSmallBestKnownTC(PSE, L)) {9404    if (ElementCount::isKnownLT(*ExpectedTC, VF.MinProfitableTripCount)) {9405      LLVM_DEBUG(dbgs() << "LV: Vectorization is not beneficial: expected "9406                           "trip count < minimum profitable VF ("9407                        << *ExpectedTC << " < " << VF.MinProfitableTripCount9408                        << ")\n");9409 9410      return false;9411    }9412  }9413  return true;9414}9415 9416LoopVectorizePass::LoopVectorizePass(LoopVectorizeOptions Opts)9417    : InterleaveOnlyWhenForced(Opts.InterleaveOnlyWhenForced ||9418                               !EnableLoopInterleaving),9419      VectorizeOnlyWhenForced(Opts.VectorizeOnlyWhenForced ||9420                              !EnableLoopVectorization) {}9421 9422/// Prepare \p MainPlan for vectorizing the main vector loop during epilogue9423/// vectorization. Remove ResumePhis from \p MainPlan for inductions that9424/// don't have a corresponding wide induction in \p EpiPlan.9425static void preparePlanForMainVectorLoop(VPlan &MainPlan, VPlan &EpiPlan) {9426  // Collect PHI nodes of widened phis in the VPlan for the epilogue. Those9427  // will need their resume-values computed in the main vector loop. Others9428  // can be removed from the main VPlan.9429  SmallPtrSet<PHINode *, 2> EpiWidenedPhis;9430  for (VPRecipeBase &R :9431       EpiPlan.getVectorLoopRegion()->getEntryBasicBlock()->phis()) {9432    if (isa<VPCanonicalIVPHIRecipe>(&R))9433      continue;9434    EpiWidenedPhis.insert(9435        cast<PHINode>(R.getVPSingleValue()->getUnderlyingValue()));9436  }9437  for (VPRecipeBase &R :9438       make_early_inc_range(MainPlan.getScalarHeader()->phis())) {9439    auto *VPIRInst = cast<VPIRPhi>(&R);9440    if (EpiWidenedPhis.contains(&VPIRInst->getIRPhi()))9441      continue;9442    // There is no corresponding wide induction in the epilogue plan that would9443    // need a resume value. Remove the VPIRInst wrapping the scalar header phi9444    // together with the corresponding ResumePhi. The resume values for the9445    // scalar loop will be created during execution of EpiPlan.9446    VPRecipeBase *ResumePhi = VPIRInst->getOperand(0)->getDefiningRecipe();9447    VPIRInst->eraseFromParent();9448    ResumePhi->eraseFromParent();9449  }9450  VPlanTransforms::runPass(VPlanTransforms::removeDeadRecipes, MainPlan);9451 9452  using namespace VPlanPatternMatch;9453  // When vectorizing the epilogue, FindFirstIV & FindLastIV reductions can9454  // introduce multiple uses of undef/poison. If the reduction start value may9455  // be undef or poison it needs to be frozen and the frozen start has to be9456  // used when computing the reduction result. We also need to use the frozen9457  // value in the resume phi generated by the main vector loop, as this is also9458  // used to compute the reduction result after the epilogue vector loop.9459  auto AddFreezeForFindLastIVReductions = [](VPlan &Plan,9460                                             bool UpdateResumePhis) {9461    VPBuilder Builder(Plan.getEntry());9462    for (VPRecipeBase &R : *Plan.getMiddleBlock()) {9463      auto *VPI = dyn_cast<VPInstruction>(&R);9464      if (!VPI || VPI->getOpcode() != VPInstruction::ComputeFindIVResult)9465        continue;9466      VPValue *OrigStart = VPI->getOperand(1);9467      if (isGuaranteedNotToBeUndefOrPoison(OrigStart->getLiveInIRValue()))9468        continue;9469      VPInstruction *Freeze =9470          Builder.createNaryOp(Instruction::Freeze, {OrigStart}, {}, "fr");9471      VPI->setOperand(1, Freeze);9472      if (UpdateResumePhis)9473        OrigStart->replaceUsesWithIf(Freeze, [Freeze](VPUser &U, unsigned) {9474          return Freeze != &U && isa<VPPhi>(&U);9475        });9476    }9477  };9478  AddFreezeForFindLastIVReductions(MainPlan, true);9479  AddFreezeForFindLastIVReductions(EpiPlan, false);9480 9481  VPBasicBlock *MainScalarPH = MainPlan.getScalarPreheader();9482  VPValue *VectorTC = &MainPlan.getVectorTripCount();9483  // If there is a suitable resume value for the canonical induction in the9484  // scalar (which will become vector) epilogue loop, use it and move it to the9485  // beginning of the scalar preheader. Otherwise create it below.9486  auto ResumePhiIter =9487      find_if(MainScalarPH->phis(), [VectorTC](VPRecipeBase &R) {9488        return match(&R, m_VPInstruction<Instruction::PHI>(m_Specific(VectorTC),9489                                                           m_ZeroInt()));9490      });9491  VPPhi *ResumePhi = nullptr;9492  if (ResumePhiIter == MainScalarPH->phis().end()) {9493    VPBuilder ScalarPHBuilder(MainScalarPH, MainScalarPH->begin());9494    ResumePhi = ScalarPHBuilder.createScalarPhi(9495        {VectorTC,9496         MainPlan.getVectorLoopRegion()->getCanonicalIV()->getStartValue()},9497        {}, "vec.epilog.resume.val");9498  } else {9499    ResumePhi = cast<VPPhi>(&*ResumePhiIter);9500    if (MainScalarPH->begin() == MainScalarPH->end())9501      ResumePhi->moveBefore(*MainScalarPH, MainScalarPH->end());9502    else if (&*MainScalarPH->begin() != ResumePhi)9503      ResumePhi->moveBefore(*MainScalarPH, MainScalarPH->begin());9504  }9505  // Add a user to to make sure the resume phi won't get removed.9506  VPBuilder(MainScalarPH)9507      .createNaryOp(VPInstruction::ResumeForEpilogue, ResumePhi);9508}9509 9510/// Prepare \p Plan for vectorizing the epilogue loop. That is, re-use expanded9511/// SCEVs from \p ExpandedSCEVs and set resume values for header recipes. Some9512/// reductions require creating new instructions to compute the resume values.9513/// They are collected in a vector and returned. They must be moved to the9514/// preheader of the vector epilogue loop, after created by the execution of \p9515/// Plan.9516static SmallVector<Instruction *> preparePlanForEpilogueVectorLoop(9517    VPlan &Plan, Loop *L, const SCEV2ValueTy &ExpandedSCEVs,9518    EpilogueLoopVectorizationInfo &EPI, LoopVectorizationCostModel &CM,9519    ScalarEvolution &SE) {9520  VPRegionBlock *VectorLoop = Plan.getVectorLoopRegion();9521  VPBasicBlock *Header = VectorLoop->getEntryBasicBlock();9522  Header->setName("vec.epilog.vector.body");9523 9524  VPCanonicalIVPHIRecipe *IV = VectorLoop->getCanonicalIV();9525  // When vectorizing the epilogue loop, the canonical induction needs to be9526  // adjusted by the value after the main vector loop. Find the resume value9527  // created during execution of the main VPlan. It must be the first phi in the9528  // loop preheader. Use the value to increment the canonical IV, and update all9529  // users in the loop region to use the adjusted value.9530  // FIXME: Improve modeling for canonical IV start values in the epilogue9531  // loop.9532  using namespace llvm::PatternMatch;9533  PHINode *EPResumeVal = &*L->getLoopPreheader()->phis().begin();9534  for (Value *Inc : EPResumeVal->incoming_values()) {9535    if (match(Inc, m_SpecificInt(0)))9536      continue;9537    assert(!EPI.VectorTripCount &&9538           "Must only have a single non-zero incoming value");9539    EPI.VectorTripCount = Inc;9540  }9541  // If we didn't find a non-zero vector trip count, all incoming values9542  // must be zero, which also means the vector trip count is zero. Pick the9543  // first zero as vector trip count.9544  // TODO: We should not choose VF * UF so the main vector loop is known to9545  // be dead.9546  if (!EPI.VectorTripCount) {9547    assert(EPResumeVal->getNumIncomingValues() > 0 &&9548           all_of(EPResumeVal->incoming_values(),9549                  [](Value *Inc) { return match(Inc, m_SpecificInt(0)); }) &&9550           "all incoming values must be 0");9551    EPI.VectorTripCount = EPResumeVal->getOperand(0);9552  }9553  VPValue *VPV = Plan.getOrAddLiveIn(EPResumeVal);9554  assert(all_of(IV->users(),9555                [](const VPUser *U) {9556                  return isa<VPScalarIVStepsRecipe>(U) ||9557                         isa<VPDerivedIVRecipe>(U) ||9558                         cast<VPRecipeBase>(U)->isScalarCast() ||9559                         cast<VPInstruction>(U)->getOpcode() ==9560                             Instruction::Add;9561                }) &&9562         "the canonical IV should only be used by its increment or "9563         "ScalarIVSteps when resetting the start value");9564  VPBuilder Builder(Header, Header->getFirstNonPhi());9565  VPInstruction *Add = Builder.createNaryOp(Instruction::Add, {IV, VPV});9566  IV->replaceAllUsesWith(Add);9567  Add->setOperand(0, IV);9568 9569  DenseMap<Value *, Value *> ToFrozen;9570  SmallVector<Instruction *> InstsToMove;9571  // Ensure that the start values for all header phi recipes are updated before9572  // vectorizing the epilogue loop. Skip the canonical IV, which has been9573  // handled above.9574  for (VPRecipeBase &R : drop_begin(Header->phis())) {9575    Value *ResumeV = nullptr;9576    // TODO: Move setting of resume values to prepareToExecute.9577    if (auto *ReductionPhi = dyn_cast<VPReductionPHIRecipe>(&R)) {9578      auto *RdxResult =9579          cast<VPInstruction>(*find_if(ReductionPhi->users(), [](VPUser *U) {9580            auto *VPI = dyn_cast<VPInstruction>(U);9581            return VPI &&9582                   (VPI->getOpcode() == VPInstruction::ComputeAnyOfResult ||9583                    VPI->getOpcode() == VPInstruction::ComputeReductionResult ||9584                    VPI->getOpcode() == VPInstruction::ComputeFindIVResult);9585          }));9586      ResumeV = cast<PHINode>(ReductionPhi->getUnderlyingInstr())9587                    ->getIncomingValueForBlock(L->getLoopPreheader());9588      RecurKind RK = ReductionPhi->getRecurrenceKind();9589      if (RecurrenceDescriptor::isAnyOfRecurrenceKind(RK)) {9590        Value *StartV = RdxResult->getOperand(1)->getLiveInIRValue();9591        // VPReductionPHIRecipes for AnyOf reductions expect a boolean as9592        // start value; compare the final value from the main vector loop9593        // to the start value.9594        BasicBlock *PBB = cast<Instruction>(ResumeV)->getParent();9595        IRBuilder<> Builder(PBB, PBB->getFirstNonPHIIt());9596        ResumeV = Builder.CreateICmpNE(ResumeV, StartV);9597        if (auto *I = dyn_cast<Instruction>(ResumeV))9598          InstsToMove.push_back(I);9599      } else if (RecurrenceDescriptor::isFindIVRecurrenceKind(RK)) {9600        Value *StartV = getStartValueFromReductionResult(RdxResult);9601        ToFrozen[StartV] = cast<PHINode>(ResumeV)->getIncomingValueForBlock(9602            EPI.MainLoopIterationCountCheck);9603 9604        // VPReductionPHIRecipe for FindFirstIV/FindLastIV reductions requires9605        // an adjustment to the resume value. The resume value is adjusted to9606        // the sentinel value when the final value from the main vector loop9607        // equals the start value. This ensures correctness when the start value9608        // might not be less than the minimum value of a monotonically9609        // increasing induction variable.9610        BasicBlock *ResumeBB = cast<Instruction>(ResumeV)->getParent();9611        IRBuilder<> Builder(ResumeBB, ResumeBB->getFirstNonPHIIt());9612        Value *Cmp = Builder.CreateICmpEQ(ResumeV, ToFrozen[StartV]);9613        if (auto *I = dyn_cast<Instruction>(Cmp))9614          InstsToMove.push_back(I);9615        Value *Sentinel = RdxResult->getOperand(2)->getLiveInIRValue();9616        ResumeV = Builder.CreateSelect(Cmp, Sentinel, ResumeV);9617        if (auto *I = dyn_cast<Instruction>(ResumeV))9618          InstsToMove.push_back(I);9619      } else {9620        VPValue *StartVal = Plan.getOrAddLiveIn(ResumeV);9621        auto *PhiR = dyn_cast<VPReductionPHIRecipe>(&R);9622        if (auto *VPI = dyn_cast<VPInstruction>(PhiR->getStartValue())) {9623          assert(VPI->getOpcode() == VPInstruction::ReductionStartVector &&9624                 "unexpected start value");9625          VPI->setOperand(0, StartVal);9626          continue;9627        }9628      }9629    } else {9630      // Retrieve the induction resume values for wide inductions from9631      // their original phi nodes in the scalar loop.9632      PHINode *IndPhi = cast<VPWidenInductionRecipe>(&R)->getPHINode();9633      // Hook up to the PHINode generated by a ResumePhi recipe of main9634      // loop VPlan, which feeds the scalar loop.9635      ResumeV = IndPhi->getIncomingValueForBlock(L->getLoopPreheader());9636    }9637    assert(ResumeV && "Must have a resume value");9638    VPValue *StartVal = Plan.getOrAddLiveIn(ResumeV);9639    cast<VPHeaderPHIRecipe>(&R)->setStartValue(StartVal);9640  }9641 9642  // For some VPValues in the epilogue plan we must re-use the generated IR9643  // values from the main plan. Replace them with live-in VPValues.9644  // TODO: This is a workaround needed for epilogue vectorization and it9645  // should be removed once induction resume value creation is done9646  // directly in VPlan.9647  for (auto &R : make_early_inc_range(*Plan.getEntry())) {9648    // Re-use frozen values from the main plan for Freeze VPInstructions in the9649    // epilogue plan. This ensures all users use the same frozen value.9650    auto *VPI = dyn_cast<VPInstruction>(&R);9651    if (VPI && VPI->getOpcode() == Instruction::Freeze) {9652      VPI->replaceAllUsesWith(Plan.getOrAddLiveIn(9653          ToFrozen.lookup(VPI->getOperand(0)->getLiveInIRValue())));9654      continue;9655    }9656 9657    // Re-use the trip count and steps expanded for the main loop, as9658    // skeleton creation needs it as a value that dominates both the scalar9659    // and vector epilogue loops9660    auto *ExpandR = dyn_cast<VPExpandSCEVRecipe>(&R);9661    if (!ExpandR)9662      continue;9663    VPValue *ExpandedVal =9664        Plan.getOrAddLiveIn(ExpandedSCEVs.lookup(ExpandR->getSCEV()));9665    ExpandR->replaceAllUsesWith(ExpandedVal);9666    if (Plan.getTripCount() == ExpandR)9667      Plan.resetTripCount(ExpandedVal);9668    ExpandR->eraseFromParent();9669  }9670 9671  auto VScale = CM.getVScaleForTuning();9672  unsigned MainLoopStep =9673      estimateElementCount(EPI.MainLoopVF * EPI.MainLoopUF, VScale);9674  unsigned EpilogueLoopStep =9675      estimateElementCount(EPI.EpilogueVF * EPI.EpilogueUF, VScale);9676  VPlanTransforms::addMinimumVectorEpilogueIterationCheck(9677      Plan, EPI.TripCount, EPI.VectorTripCount,9678      CM.requiresScalarEpilogue(EPI.EpilogueVF.isVector()), EPI.EpilogueVF,9679      EPI.EpilogueUF, MainLoopStep, EpilogueLoopStep, SE);9680 9681  return InstsToMove;9682}9683 9684// Generate bypass values from the additional bypass block. Note that when the9685// vectorized epilogue is skipped due to iteration count check, then the9686// resume value for the induction variable comes from the trip count of the9687// main vector loop, passed as the second argument.9688static Value *createInductionAdditionalBypassValues(9689    PHINode *OrigPhi, const InductionDescriptor &II, IRBuilder<> &BypassBuilder,9690    const SCEV2ValueTy &ExpandedSCEVs, Value *MainVectorTripCount,9691    Instruction *OldInduction) {9692  Value *Step = getExpandedStep(II, ExpandedSCEVs);9693  // For the primary induction the additional bypass end value is known.9694  // Otherwise it is computed.9695  Value *EndValueFromAdditionalBypass = MainVectorTripCount;9696  if (OrigPhi != OldInduction) {9697    auto *BinOp = II.getInductionBinOp();9698    // Fast-math-flags propagate from the original induction instruction.9699    if (isa_and_nonnull<FPMathOperator>(BinOp))9700      BypassBuilder.setFastMathFlags(BinOp->getFastMathFlags());9701 9702    // Compute the end value for the additional bypass.9703    EndValueFromAdditionalBypass =9704        emitTransformedIndex(BypassBuilder, MainVectorTripCount,9705                             II.getStartValue(), Step, II.getKind(), BinOp);9706    EndValueFromAdditionalBypass->setName("ind.end");9707  }9708  return EndValueFromAdditionalBypass;9709}9710 9711static void fixScalarResumeValuesFromBypass(BasicBlock *BypassBlock, Loop *L,9712                                            VPlan &BestEpiPlan,9713                                            LoopVectorizationLegality &LVL,9714                                            const SCEV2ValueTy &ExpandedSCEVs,9715                                            Value *MainVectorTripCount) {9716  // Fix reduction resume values from the additional bypass block.9717  BasicBlock *PH = L->getLoopPreheader();9718  for (auto *Pred : predecessors(PH)) {9719    for (PHINode &Phi : PH->phis()) {9720      if (Phi.getBasicBlockIndex(Pred) != -1)9721        continue;9722      Phi.addIncoming(Phi.getIncomingValueForBlock(BypassBlock), Pred);9723    }9724  }9725  auto *ScalarPH = cast<VPIRBasicBlock>(BestEpiPlan.getScalarPreheader());9726  if (ScalarPH->hasPredecessors()) {9727    // If ScalarPH has predecessors, we may need to update its reduction9728    // resume values.9729    for (const auto &[R, IRPhi] :9730         zip(ScalarPH->phis(), ScalarPH->getIRBasicBlock()->phis())) {9731      fixReductionScalarResumeWhenVectorizingEpilog(cast<VPPhi>(&R), IRPhi,9732                                                    BypassBlock);9733    }9734  }9735 9736  // Fix induction resume values from the additional bypass block.9737  IRBuilder<> BypassBuilder(BypassBlock, BypassBlock->getFirstInsertionPt());9738  for (const auto &[IVPhi, II] : LVL.getInductionVars()) {9739    auto *Inc = cast<PHINode>(IVPhi->getIncomingValueForBlock(PH));9740    Value *V = createInductionAdditionalBypassValues(9741        IVPhi, II, BypassBuilder, ExpandedSCEVs, MainVectorTripCount,9742        LVL.getPrimaryInduction());9743    // TODO: Directly add as extra operand to the VPResumePHI recipe.9744    Inc->setIncomingValueForBlock(BypassBlock, V);9745  }9746}9747 9748/// Connect the epilogue vector loop generated for \p EpiPlan to the main vector9749// loop, after both plans have executed, updating branches from the iteration9750// and runtime checks of the main loop, as well as updating various phis. \p9751// InstsToMove contains instructions that need to be moved to the preheader of9752// the epilogue vector loop.9753static void connectEpilogueVectorLoop(9754    VPlan &EpiPlan, Loop *L, EpilogueLoopVectorizationInfo &EPI,9755    DominatorTree *DT, LoopVectorizationLegality &LVL,9756    DenseMap<const SCEV *, Value *> &ExpandedSCEVs, GeneratedRTChecks &Checks,9757    ArrayRef<Instruction *> InstsToMove) {9758  BasicBlock *VecEpilogueIterationCountCheck =9759      cast<VPIRBasicBlock>(EpiPlan.getEntry())->getIRBasicBlock();9760 9761  BasicBlock *VecEpiloguePreHeader =9762      cast<BranchInst>(VecEpilogueIterationCountCheck->getTerminator())9763          ->getSuccessor(1);9764  // Adjust the control flow taking the state info from the main loop9765  // vectorization into account.9766  assert(EPI.MainLoopIterationCountCheck && EPI.EpilogueIterationCountCheck &&9767         "expected this to be saved from the previous pass.");9768  DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Eager);9769  EPI.MainLoopIterationCountCheck->getTerminator()->replaceUsesOfWith(9770      VecEpilogueIterationCountCheck, VecEpiloguePreHeader);9771 9772  DTU.applyUpdates({{DominatorTree::Delete, EPI.MainLoopIterationCountCheck,9773                     VecEpilogueIterationCountCheck},9774                    {DominatorTree::Insert, EPI.MainLoopIterationCountCheck,9775                     VecEpiloguePreHeader}});9776 9777  BasicBlock *ScalarPH =9778      cast<VPIRBasicBlock>(EpiPlan.getScalarPreheader())->getIRBasicBlock();9779  EPI.EpilogueIterationCountCheck->getTerminator()->replaceUsesOfWith(9780      VecEpilogueIterationCountCheck, ScalarPH);9781  DTU.applyUpdates(9782      {{DominatorTree::Delete, EPI.EpilogueIterationCountCheck,9783        VecEpilogueIterationCountCheck},9784       {DominatorTree::Insert, EPI.EpilogueIterationCountCheck, ScalarPH}});9785 9786  // Adjust the terminators of runtime check blocks and phis using them.9787  BasicBlock *SCEVCheckBlock = Checks.getSCEVChecks().second;9788  BasicBlock *MemCheckBlock = Checks.getMemRuntimeChecks().second;9789  if (SCEVCheckBlock) {9790    SCEVCheckBlock->getTerminator()->replaceUsesOfWith(9791        VecEpilogueIterationCountCheck, ScalarPH);9792    DTU.applyUpdates({{DominatorTree::Delete, SCEVCheckBlock,9793                       VecEpilogueIterationCountCheck},9794                      {DominatorTree::Insert, SCEVCheckBlock, ScalarPH}});9795  }9796  if (MemCheckBlock) {9797    MemCheckBlock->getTerminator()->replaceUsesOfWith(9798        VecEpilogueIterationCountCheck, ScalarPH);9799    DTU.applyUpdates(9800        {{DominatorTree::Delete, MemCheckBlock, VecEpilogueIterationCountCheck},9801         {DominatorTree::Insert, MemCheckBlock, ScalarPH}});9802  }9803 9804  // The vec.epilog.iter.check block may contain Phi nodes from inductions9805  // or reductions which merge control-flow from the latch block and the9806  // middle block. Update the incoming values here and move the Phi into the9807  // preheader.9808  SmallVector<PHINode *, 4> PhisInBlock(9809      llvm::make_pointer_range(VecEpilogueIterationCountCheck->phis()));9810 9811  for (PHINode *Phi : PhisInBlock) {9812    Phi->moveBefore(VecEpiloguePreHeader->getFirstNonPHIIt());9813    Phi->replaceIncomingBlockWith(9814        VecEpilogueIterationCountCheck->getSinglePredecessor(),9815        VecEpilogueIterationCountCheck);9816 9817    // If the phi doesn't have an incoming value from the9818    // EpilogueIterationCountCheck, we are done. Otherwise remove the9819    // incoming value and also those from other check blocks. This is needed9820    // for reduction phis only.9821    if (none_of(Phi->blocks(), [&](BasicBlock *IncB) {9822          return EPI.EpilogueIterationCountCheck == IncB;9823        }))9824      continue;9825    Phi->removeIncomingValue(EPI.EpilogueIterationCountCheck);9826    if (SCEVCheckBlock)9827      Phi->removeIncomingValue(SCEVCheckBlock);9828    if (MemCheckBlock)9829      Phi->removeIncomingValue(MemCheckBlock);9830  }9831 9832  auto IP = VecEpiloguePreHeader->getFirstNonPHIIt();9833  for (auto *I : InstsToMove)9834    I->moveBefore(IP);9835 9836  // VecEpilogueIterationCountCheck conditionally skips over the epilogue loop9837  // after executing the main loop. We need to update the resume values of9838  // inductions and reductions during epilogue vectorization.9839  fixScalarResumeValuesFromBypass(VecEpilogueIterationCountCheck, L, EpiPlan,9840                                  LVL, ExpandedSCEVs, EPI.VectorTripCount);9841}9842 9843bool LoopVectorizePass::processLoop(Loop *L) {9844  assert((EnableVPlanNativePath || L->isInnermost()) &&9845         "VPlan-native path is not enabled. Only process inner loops.");9846 9847  LLVM_DEBUG(dbgs() << "\nLV: Checking a loop in '"9848                    << L->getHeader()->getParent()->getName() << "' from "9849                    << L->getLocStr() << "\n");9850 9851  LoopVectorizeHints Hints(L, InterleaveOnlyWhenForced, *ORE, TTI);9852 9853  LLVM_DEBUG(9854      dbgs() << "LV: Loop hints:"9855             << " force="9856             << (Hints.getForce() == LoopVectorizeHints::FK_Disabled9857                     ? "disabled"9858                     : (Hints.getForce() == LoopVectorizeHints::FK_Enabled9859                            ? "enabled"9860                            : "?"))9861             << " width=" << Hints.getWidth()9862             << " interleave=" << Hints.getInterleave() << "\n");9863 9864  // Function containing loop9865  Function *F = L->getHeader()->getParent();9866 9867  // Looking at the diagnostic output is the only way to determine if a loop9868  // was vectorized (other than looking at the IR or machine code), so it9869  // is important to generate an optimization remark for each loop. Most of9870  // these messages are generated as OptimizationRemarkAnalysis. Remarks9871  // generated as OptimizationRemark and OptimizationRemarkMissed are9872  // less verbose reporting vectorized loops and unvectorized loops that may9873  // benefit from vectorization, respectively.9874 9875  if (!Hints.allowVectorization(F, L, VectorizeOnlyWhenForced)) {9876    LLVM_DEBUG(dbgs() << "LV: Loop hints prevent vectorization.\n");9877    return false;9878  }9879 9880  PredicatedScalarEvolution PSE(*SE, *L);9881 9882  // Query this against the original loop and save it here because the profile9883  // of the original loop header may change as the transformation happens.9884  bool OptForSize = llvm::shouldOptimizeForSize(L->getHeader(), PSI, BFI,9885                                                PGSOQueryType::IRPass);9886 9887  // Check if it is legal to vectorize the loop.9888  LoopVectorizationRequirements Requirements;9889  LoopVectorizationLegality LVL(L, PSE, DT, TTI, TLI, F, *LAIs, LI, ORE,9890                                &Requirements, &Hints, DB, AC,9891                                /*AllowRuntimeSCEVChecks=*/!OptForSize, AA);9892  if (!LVL.canVectorize(EnableVPlanNativePath)) {9893    LLVM_DEBUG(dbgs() << "LV: Not vectorizing: Cannot prove legality.\n");9894    Hints.emitRemarkWithHints();9895    return false;9896  }9897 9898  if (LVL.hasUncountableEarlyExit() && !EnableEarlyExitVectorization) {9899    reportVectorizationFailure("Auto-vectorization of loops with uncountable "9900                               "early exit is not enabled",9901                               "UncountableEarlyExitLoopsDisabled", ORE, L);9902    return false;9903  }9904 9905  if (!LVL.getPotentiallyFaultingLoads().empty()) {9906    reportVectorizationFailure("Auto-vectorization of loops with potentially "9907                               "faulting load is not supported",9908                               "PotentiallyFaultingLoadsNotSupported", ORE, L);9909    return false;9910  }9911 9912  // Entrance to the VPlan-native vectorization path. Outer loops are processed9913  // here. They may require CFG and instruction level transformations before9914  // even evaluating whether vectorization is profitable. Since we cannot modify9915  // the incoming IR, we need to build VPlan upfront in the vectorization9916  // pipeline.9917  if (!L->isInnermost())9918    return processLoopInVPlanNativePath(L, PSE, LI, DT, &LVL, TTI, TLI, DB, AC,9919                                        ORE, OptForSize, Hints, Requirements);9920 9921  assert(L->isInnermost() && "Inner loop expected.");9922 9923  InterleavedAccessInfo IAI(PSE, L, DT, LI, LVL.getLAI());9924  bool UseInterleaved = TTI->enableInterleavedAccessVectorization();9925 9926  // If an override option has been passed in for interleaved accesses, use it.9927  if (EnableInterleavedMemAccesses.getNumOccurrences() > 0)9928    UseInterleaved = EnableInterleavedMemAccesses;9929 9930  // Analyze interleaved memory accesses.9931  if (UseInterleaved)9932    IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI));9933 9934  if (LVL.hasUncountableEarlyExit()) {9935    BasicBlock *LoopLatch = L->getLoopLatch();9936    if (IAI.requiresScalarEpilogue() ||9937        any_of(LVL.getCountableExitingBlocks(),9938               [LoopLatch](BasicBlock *BB) { return BB != LoopLatch; })) {9939      reportVectorizationFailure("Auto-vectorization of early exit loops "9940                                 "requiring a scalar epilogue is unsupported",9941                                 "UncountableEarlyExitUnsupported", ORE, L);9942      return false;9943    }9944  }9945 9946  // Check the function attributes and profiles to find out if this function9947  // should be optimized for size.9948  ScalarEpilogueLowering SEL =9949      getScalarEpilogueLowering(F, L, Hints, OptForSize, TTI, TLI, LVL, &IAI);9950 9951  // Check the loop for a trip count threshold: vectorize loops with a tiny trip9952  // count by optimizing for size, to minimize overheads.9953  auto ExpectedTC = getSmallBestKnownTC(PSE, L);9954  if (ExpectedTC && ExpectedTC->isFixed() &&9955      ExpectedTC->getFixedValue() < TinyTripCountVectorThreshold) {9956    LLVM_DEBUG(dbgs() << "LV: Found a loop with a very small trip count. "9957                      << "This loop is worth vectorizing only if no scalar "9958                      << "iteration overheads are incurred.");9959    if (Hints.getForce() == LoopVectorizeHints::FK_Enabled)9960      LLVM_DEBUG(dbgs() << " But vectorizing was explicitly forced.\n");9961    else {9962      LLVM_DEBUG(dbgs() << "\n");9963      // Predicate tail-folded loops are efficient even when the loop9964      // iteration count is low. However, setting the epilogue policy to9965      // `CM_ScalarEpilogueNotAllowedLowTripLoop` prevents vectorizing loops9966      // with runtime checks. It's more effective to let9967      // `isOutsideLoopWorkProfitable` determine if vectorization is9968      // beneficial for the loop.9969      if (SEL != CM_ScalarEpilogueNotNeededUsePredicate)9970        SEL = CM_ScalarEpilogueNotAllowedLowTripLoop;9971    }9972  }9973 9974  // Check the function attributes to see if implicit floats or vectors are9975  // allowed.9976  if (F->hasFnAttribute(Attribute::NoImplicitFloat)) {9977    reportVectorizationFailure(9978        "Can't vectorize when the NoImplicitFloat attribute is used",9979        "loop not vectorized due to NoImplicitFloat attribute",9980        "NoImplicitFloat", ORE, L);9981    Hints.emitRemarkWithHints();9982    return false;9983  }9984 9985  // Check if the target supports potentially unsafe FP vectorization.9986  // FIXME: Add a check for the type of safety issue (denormal, signaling)9987  // for the target we're vectorizing for, to make sure none of the9988  // additional fp-math flags can help.9989  if (Hints.isPotentiallyUnsafe() &&9990      TTI->isFPVectorizationPotentiallyUnsafe()) {9991    reportVectorizationFailure(9992        "Potentially unsafe FP op prevents vectorization",9993        "loop not vectorized due to unsafe FP support.",9994        "UnsafeFP", ORE, L);9995    Hints.emitRemarkWithHints();9996    return false;9997  }9998 9999  bool AllowOrderedReductions;10000  // If the flag is set, use that instead and override the TTI behaviour.10001  if (ForceOrderedReductions.getNumOccurrences() > 0)10002    AllowOrderedReductions = ForceOrderedReductions;10003  else10004    AllowOrderedReductions = TTI->enableOrderedReductions();10005  if (!LVL.canVectorizeFPMath(AllowOrderedReductions)) {10006    ORE->emit([&]() {10007      auto *ExactFPMathInst = Requirements.getExactFPInst();10008      return OptimizationRemarkAnalysisFPCommute(DEBUG_TYPE, "CantReorderFPOps",10009                                                 ExactFPMathInst->getDebugLoc(),10010                                                 ExactFPMathInst->getParent())10011             << "loop not vectorized: cannot prove it is safe to reorder "10012                "floating-point operations";10013    });10014    LLVM_DEBUG(dbgs() << "LV: loop not vectorized: cannot prove it is safe to "10015                         "reorder floating-point operations\n");10016    Hints.emitRemarkWithHints();10017    return false;10018  }10019 10020  // Use the cost model.10021  LoopVectorizationCostModel CM(SEL, L, PSE, LI, &LVL, *TTI, TLI, DB, AC, ORE,10022                                F, &Hints, IAI, OptForSize);10023  // Use the planner for vectorization.10024  LoopVectorizationPlanner LVP(L, LI, DT, TLI, *TTI, &LVL, CM, IAI, PSE, Hints,10025                               ORE);10026 10027  // Get user vectorization factor and interleave count.10028  ElementCount UserVF = Hints.getWidth();10029  unsigned UserIC = Hints.getInterleave();10030  if (UserIC > 1 && !LVL.isSafeForAnyVectorWidth())10031    UserIC = 1;10032 10033  // Plan how to best vectorize.10034  LVP.plan(UserVF, UserIC);10035  VectorizationFactor VF = LVP.computeBestVF();10036  unsigned IC = 1;10037 10038  if (ORE->allowExtraAnalysis(LV_NAME))10039    LVP.emitInvalidCostRemarks(ORE);10040 10041  GeneratedRTChecks Checks(PSE, DT, LI, TTI, F->getDataLayout(), CM.CostKind);10042  if (LVP.hasPlanWithVF(VF.Width)) {10043    // Select the interleave count.10044    IC = LVP.selectInterleaveCount(LVP.getPlanFor(VF.Width), VF.Width, VF.Cost);10045 10046    unsigned SelectedIC = std::max(IC, UserIC);10047    //  Optimistically generate runtime checks if they are needed. Drop them if10048    //  they turn out to not be profitable.10049    if (VF.Width.isVector() || SelectedIC > 1) {10050      Checks.create(L, *LVL.getLAI(), PSE.getPredicate(), VF.Width, SelectedIC);10051 10052      // Bail out early if either the SCEV or memory runtime checks are known to10053      // fail. In that case, the vector loop would never execute.10054      using namespace llvm::PatternMatch;10055      if (Checks.getSCEVChecks().first &&10056          match(Checks.getSCEVChecks().first, m_One()))10057        return false;10058      if (Checks.getMemRuntimeChecks().first &&10059          match(Checks.getMemRuntimeChecks().first, m_One()))10060        return false;10061    }10062 10063    // Check if it is profitable to vectorize with runtime checks.10064    bool ForceVectorization =10065        Hints.getForce() == LoopVectorizeHints::FK_Enabled;10066    VPCostContext CostCtx(CM.TTI, *CM.TLI, LVP.getPlanFor(VF.Width), CM,10067                          CM.CostKind, *CM.PSE.getSE(), L);10068    if (!ForceVectorization &&10069        !isOutsideLoopWorkProfitable(Checks, VF, L, PSE, CostCtx,10070                                     LVP.getPlanFor(VF.Width), SEL,10071                                     CM.getVScaleForTuning())) {10072      ORE->emit([&]() {10073        return OptimizationRemarkAnalysisAliasing(10074                   DEBUG_TYPE, "CantReorderMemOps", L->getStartLoc(),10075                   L->getHeader())10076               << "loop not vectorized: cannot prove it is safe to reorder "10077                  "memory operations";10078      });10079      LLVM_DEBUG(dbgs() << "LV: Too many memory checks needed.\n");10080      Hints.emitRemarkWithHints();10081      return false;10082    }10083  }10084 10085  // Identify the diagnostic messages that should be produced.10086  std::pair<StringRef, std::string> VecDiagMsg, IntDiagMsg;10087  bool VectorizeLoop = true, InterleaveLoop = true;10088  if (VF.Width.isScalar()) {10089    LLVM_DEBUG(dbgs() << "LV: Vectorization is possible but not beneficial.\n");10090    VecDiagMsg = {10091        "VectorizationNotBeneficial",10092        "the cost-model indicates that vectorization is not beneficial"};10093    VectorizeLoop = false;10094  }10095 10096  if (UserIC == 1 && Hints.getInterleave() > 1) {10097    assert(!LVL.isSafeForAnyVectorWidth() &&10098           "UserIC should only be ignored due to unsafe dependencies");10099    LLVM_DEBUG(dbgs() << "LV: Ignoring user-specified interleave count.\n");10100    IntDiagMsg = {"InterleavingUnsafe",10101                  "Ignoring user-specified interleave count due to possibly "10102                  "unsafe dependencies in the loop."};10103    InterleaveLoop = false;10104  } else if (!LVP.hasPlanWithVF(VF.Width) && UserIC > 1) {10105    // Tell the user interleaving was avoided up-front, despite being explicitly10106    // requested.10107    LLVM_DEBUG(dbgs() << "LV: Ignoring UserIC, because vectorization and "10108                         "interleaving should be avoided up front\n");10109    IntDiagMsg = {"InterleavingAvoided",10110                  "Ignoring UserIC, because interleaving was avoided up front"};10111    InterleaveLoop = false;10112  } else if (IC == 1 && UserIC <= 1) {10113    // Tell the user interleaving is not beneficial.10114    LLVM_DEBUG(dbgs() << "LV: Interleaving is not beneficial.\n");10115    IntDiagMsg = {10116        "InterleavingNotBeneficial",10117        "the cost-model indicates that interleaving is not beneficial"};10118    InterleaveLoop = false;10119    if (UserIC == 1) {10120      IntDiagMsg.first = "InterleavingNotBeneficialAndDisabled";10121      IntDiagMsg.second +=10122          " and is explicitly disabled or interleave count is set to 1";10123    }10124  } else if (IC > 1 && UserIC == 1) {10125    // Tell the user interleaving is beneficial, but it explicitly disabled.10126    LLVM_DEBUG(dbgs() << "LV: Interleaving is beneficial but is explicitly "10127                         "disabled.\n");10128    IntDiagMsg = {"InterleavingBeneficialButDisabled",10129                  "the cost-model indicates that interleaving is beneficial "10130                  "but is explicitly disabled or interleave count is set to 1"};10131    InterleaveLoop = false;10132  }10133 10134  // If there is a histogram in the loop, do not just interleave without10135  // vectorizing. The order of operations will be incorrect without the10136  // histogram intrinsics, which are only used for recipes with VF > 1.10137  if (!VectorizeLoop && InterleaveLoop && LVL.hasHistograms()) {10138    LLVM_DEBUG(dbgs() << "LV: Not interleaving without vectorization due "10139                      << "to histogram operations.\n");10140    IntDiagMsg = {10141        "HistogramPreventsScalarInterleaving",10142        "Unable to interleave without vectorization due to constraints on "10143        "the order of histogram operations"};10144    InterleaveLoop = false;10145  }10146 10147  // Override IC if user provided an interleave count.10148  IC = UserIC > 0 ? UserIC : IC;10149 10150  // Emit diagnostic messages, if any.10151  const char *VAPassName = Hints.vectorizeAnalysisPassName();10152  if (!VectorizeLoop && !InterleaveLoop) {10153    // Do not vectorize or interleaving the loop.10154    ORE->emit([&]() {10155      return OptimizationRemarkMissed(VAPassName, VecDiagMsg.first,10156                                      L->getStartLoc(), L->getHeader())10157             << VecDiagMsg.second;10158    });10159    ORE->emit([&]() {10160      return OptimizationRemarkMissed(LV_NAME, IntDiagMsg.first,10161                                      L->getStartLoc(), L->getHeader())10162             << IntDiagMsg.second;10163    });10164    return false;10165  }10166 10167  if (!VectorizeLoop && InterleaveLoop) {10168    LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');10169    ORE->emit([&]() {10170      return OptimizationRemarkAnalysis(VAPassName, VecDiagMsg.first,10171                                        L->getStartLoc(), L->getHeader())10172             << VecDiagMsg.second;10173    });10174  } else if (VectorizeLoop && !InterleaveLoop) {10175    LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width10176                      << ") in " << L->getLocStr() << '\n');10177    ORE->emit([&]() {10178      return OptimizationRemarkAnalysis(LV_NAME, IntDiagMsg.first,10179                                        L->getStartLoc(), L->getHeader())10180             << IntDiagMsg.second;10181    });10182  } else if (VectorizeLoop && InterleaveLoop) {10183    LLVM_DEBUG(dbgs() << "LV: Found a vectorizable loop (" << VF.Width10184                      << ") in " << L->getLocStr() << '\n');10185    LLVM_DEBUG(dbgs() << "LV: Interleave Count is " << IC << '\n');10186  }10187 10188  // Report the vectorization decision.10189  if (VF.Width.isScalar()) {10190    using namespace ore;10191    assert(IC > 1);10192    ORE->emit([&]() {10193      return OptimizationRemark(LV_NAME, "Interleaved", L->getStartLoc(),10194                                L->getHeader())10195             << "interleaved loop (interleaved count: "10196             << NV("InterleaveCount", IC) << ")";10197    });10198  } else {10199    // Report the vectorization decision.10200    reportVectorization(ORE, L, VF, IC);10201  }10202  if (ORE->allowExtraAnalysis(LV_NAME))10203    checkMixedPrecision(L, ORE);10204 10205  // If we decided that it is *legal* to interleave or vectorize the loop, then10206  // do it.10207 10208  VPlan &BestPlan = LVP.getPlanFor(VF.Width);10209  // Consider vectorizing the epilogue too if it's profitable.10210  VectorizationFactor EpilogueVF =10211      LVP.selectEpilogueVectorizationFactor(VF.Width, IC);10212  if (EpilogueVF.Width.isVector()) {10213    std::unique_ptr<VPlan> BestMainPlan(BestPlan.duplicate());10214 10215    // The first pass vectorizes the main loop and creates a scalar epilogue10216    // to be vectorized by executing the plan (potentially with a different10217    // factor) again shortly afterwards.10218    VPlan &BestEpiPlan = LVP.getPlanFor(EpilogueVF.Width);10219    BestEpiPlan.getMiddleBlock()->setName("vec.epilog.middle.block");10220    BestEpiPlan.getVectorPreheader()->setName("vec.epilog.ph");10221    preparePlanForMainVectorLoop(*BestMainPlan, BestEpiPlan);10222    EpilogueLoopVectorizationInfo EPI(VF.Width, IC, EpilogueVF.Width, 1,10223                                      BestEpiPlan);10224    EpilogueVectorizerMainLoop MainILV(L, PSE, LI, DT, TTI, AC, EPI, &CM,10225                                       Checks, *BestMainPlan);10226    auto ExpandedSCEVs = LVP.executePlan(EPI.MainLoopVF, EPI.MainLoopUF,10227                                         *BestMainPlan, MainILV, DT, false);10228    ++LoopsVectorized;10229 10230    // Second pass vectorizes the epilogue and adjusts the control flow10231    // edges from the first pass.10232    EpilogueVectorizerEpilogueLoop EpilogILV(L, PSE, LI, DT, TTI, AC, EPI, &CM,10233                                             Checks, BestEpiPlan);10234    SmallVector<Instruction *> InstsToMove = preparePlanForEpilogueVectorLoop(10235        BestEpiPlan, L, ExpandedSCEVs, EPI, CM, *PSE.getSE());10236    LVP.executePlan(EPI.EpilogueVF, EPI.EpilogueUF, BestEpiPlan, EpilogILV, DT,10237                    true);10238    connectEpilogueVectorLoop(BestEpiPlan, L, EPI, DT, LVL, ExpandedSCEVs,10239                              Checks, InstsToMove);10240    ++LoopsEpilogueVectorized;10241  } else {10242    InnerLoopVectorizer LB(L, PSE, LI, DT, TTI, AC, VF.Width, IC, &CM, Checks,10243                           BestPlan);10244    // TODO: Move to general VPlan pipeline once epilogue loops are also10245    // supported.10246    VPlanTransforms::runPass(10247        VPlanTransforms::materializeConstantVectorTripCount, BestPlan, VF.Width,10248        IC, PSE);10249    LVP.addMinimumIterationCheck(BestPlan, VF.Width, IC,10250                                 VF.MinProfitableTripCount);10251 10252    LVP.executePlan(VF.Width, IC, BestPlan, LB, DT, false);10253    ++LoopsVectorized;10254  }10255 10256  assert(DT->verify(DominatorTree::VerificationLevel::Fast) &&10257         "DT not preserved correctly");10258  assert(!verifyFunction(*F, &dbgs()));10259 10260  return true;10261}10262 10263LoopVectorizeResult LoopVectorizePass::runImpl(Function &F) {10264 10265  // Don't attempt if10266  // 1. the target claims to have no vector registers, and10267  // 2. interleaving won't help ILP.10268  //10269  // The second condition is necessary because, even if the target has no10270  // vector registers, loop vectorization may still enable scalar10271  // interleaving.10272  if (!TTI->getNumberOfRegisters(TTI->getRegisterClassForType(true)) &&10273      TTI->getMaxInterleaveFactor(ElementCount::getFixed(1)) < 2)10274    return LoopVectorizeResult(false, false);10275 10276  bool Changed = false, CFGChanged = false;10277 10278  // The vectorizer requires loops to be in simplified form.10279  // Since simplification may add new inner loops, it has to run before the10280  // legality and profitability checks. This means running the loop vectorizer10281  // will simplify all loops, regardless of whether anything end up being10282  // vectorized.10283  for (const auto &L : *LI)10284    Changed |= CFGChanged |=10285        simplifyLoop(L, DT, LI, SE, AC, nullptr, false /* PreserveLCSSA */);10286 10287  // Build up a worklist of inner-loops to vectorize. This is necessary as10288  // the act of vectorizing or partially unrolling a loop creates new loops10289  // and can invalidate iterators across the loops.10290  SmallVector<Loop *, 8> Worklist;10291 10292  for (Loop *L : *LI)10293    collectSupportedLoops(*L, LI, ORE, Worklist);10294 10295  LoopsAnalyzed += Worklist.size();10296 10297  // Now walk the identified inner loops.10298  while (!Worklist.empty()) {10299    Loop *L = Worklist.pop_back_val();10300 10301    // For the inner loops we actually process, form LCSSA to simplify the10302    // transform.10303    Changed |= formLCSSARecursively(*L, *DT, LI, SE);10304 10305    Changed |= CFGChanged |= processLoop(L);10306 10307    if (Changed) {10308      LAIs->clear();10309 10310#ifndef NDEBUG10311      if (VerifySCEV)10312        SE->verify();10313#endif10314    }10315  }10316 10317  // Process each loop nest in the function.10318  return LoopVectorizeResult(Changed, CFGChanged);10319}10320 10321PreservedAnalyses LoopVectorizePass::run(Function &F,10322                                         FunctionAnalysisManager &AM) {10323  LI = &AM.getResult<LoopAnalysis>(F);10324  // There are no loops in the function. Return before computing other10325  // expensive analyses.10326  if (LI->empty())10327    return PreservedAnalyses::all();10328  SE = &AM.getResult<ScalarEvolutionAnalysis>(F);10329  TTI = &AM.getResult<TargetIRAnalysis>(F);10330  DT = &AM.getResult<DominatorTreeAnalysis>(F);10331  TLI = &AM.getResult<TargetLibraryAnalysis>(F);10332  AC = &AM.getResult<AssumptionAnalysis>(F);10333  DB = &AM.getResult<DemandedBitsAnalysis>(F);10334  ORE = &AM.getResult<OptimizationRemarkEmitterAnalysis>(F);10335  LAIs = &AM.getResult<LoopAccessAnalysis>(F);10336  AA = &AM.getResult<AAManager>(F);10337 10338  auto &MAMProxy = AM.getResult<ModuleAnalysisManagerFunctionProxy>(F);10339  PSI = MAMProxy.getCachedResult<ProfileSummaryAnalysis>(*F.getParent());10340  BFI = nullptr;10341  if (PSI && PSI->hasProfileSummary())10342    BFI = &AM.getResult<BlockFrequencyAnalysis>(F);10343  LoopVectorizeResult Result = runImpl(F);10344  if (!Result.MadeAnyChange)10345    return PreservedAnalyses::all();10346  PreservedAnalyses PA;10347 10348  if (isAssignmentTrackingEnabled(*F.getParent())) {10349    for (auto &BB : F)10350      RemoveRedundantDbgInstrs(&BB);10351  }10352 10353  PA.preserve<LoopAnalysis>();10354  PA.preserve<DominatorTreeAnalysis>();10355  PA.preserve<ScalarEvolutionAnalysis>();10356  PA.preserve<LoopAccessAnalysis>();10357 10358  if (Result.MadeCFGChange) {10359    // Making CFG changes likely means a loop got vectorized. Indicate that10360    // extra simplification passes should be run.10361    // TODO: MadeCFGChanges is not a prefect proxy. Extra passes should only10362    // be run if runtime checks have been added.10363    AM.getResult<ShouldRunExtraVectorPasses>(F);10364    PA.preserve<ShouldRunExtraVectorPasses>();10365  } else {10366    PA.preserveSet<CFGAnalyses>();10367  }10368  return PA;10369}10370 10371void LoopVectorizePass::printPipeline(10372    raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {10373  static_cast<PassInfoMixin<LoopVectorizePass> *>(this)->printPipeline(10374      OS, MapClassName2PassName);10375 10376  OS << '<';10377  OS << (InterleaveOnlyWhenForced ? "" : "no-") << "interleave-forced-only;";10378  OS << (VectorizeOnlyWhenForced ? "" : "no-") << "vectorize-forced-only;";10379  OS << '>';10380}10381