SLPVectorizer.cpp

Go to the documentation of this file.

//===- SLPVectorizer.cpp - A bottom up SLP Vectorizer ---------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This pass implements the Bottom Up SLP vectorizer. It detects consecutive
// stores that can be put together into vector-stores. Next, it attempts to
// construct vectorizable tree using the use-def chains. If a profitable tree
// was found, the SLP vectorizer performs vectorization on the tree.
//
// The pass is inspired by the work described in the paper:
//  "Loop-Aware SLP in GCC" by Ira Rosen, Dorit Nuzman, Ayal Zaks.
//
//===----------------------------------------------------------------------===//
 
#include "llvm/Transforms/Vectorize/SLPVectorizer.h"
#include "SLPVectorizer/SLPCompatibilityAnalysis.h"
#include "SLPVectorizer/SLPCostAnalysis.h"
#include "SLPVectorizer/SLPTypeUtils.h"
#include "SLPVectorizer/SLPUtils.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/ADT/PriorityQueue.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/ScopeExit.h"
#include "llvm/ADT/SetOperations.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/SmallBitVector.h"
#include "llvm/ADT/SmallPtrSet.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallString.h"
#include "llvm/ADT/SmallVectorExtras.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/iterator.h"
#include "llvm/ADT/iterator_range.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/AssumptionCache.h"
#include "llvm/Analysis/CodeMetrics.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Analysis/DemandedBits.h"
#include "llvm/Analysis/DomTreeUpdater.h"
#include "llvm/Analysis/GlobalsModRef.h"
#include "llvm/Analysis/IVDescriptors.h"
#include "llvm/Analysis/Loads.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/MemoryLocation.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/Analysis/TargetLibraryInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/IR/Attributes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/CFG.h"
#include "llvm/IR/Constant.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/MDBuilder.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/Operator.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/IR/ProfDataUtils.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Use.h"
#include "llvm/IR/User.h"
#include "llvm/IR/Value.h"
#include "llvm/IR/ValueHandle.h"
#include "llvm/IR/VectorTypeUtils.h"
#ifdef EXPENSIVE_CHECKS
#include "llvm/IR/Verifier.h"
#endif
#include "llvm/Pass.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/DOTGraphTraits.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/DebugCounter.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GraphWriter.h"
#include "llvm/Support/InstructionCost.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Transforms/Utils/InjectTLIMappings.h"
#include "llvm/Transforms/Utils/Local.h"
#include "llvm/Transforms/Utils/LoopUtils.h"
#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <iterator>
#include <map>
#include <memory>
#include <optional>
#include <set>
#include <string>
#include <tuple>
#include <utility>
 
using namespace llvm;
using namespace llvm::PatternMatch;
using namespace slpvectorizer;
using namespace std::placeholders;
 
#define SV_NAME "slp-vectorizer"
#define DEBUG_TYPE "SLP"
 
STATISTIC(NumVectorInstructions, "Number of vector instructions generated");
STATISTIC(NumStridedStoreChains, "Number of vectorized stride stores");
STATISTIC(NumStoreChains, "Number of vector stores created");
STATISTIC(NumVectorizedStores, "Number of vectorized stores");
 
DEBUG_COUNTER(VectorizedGraphs, "slp-vectorized",
              "Controls which SLP graphs should be vectorized.");
 
static cl::opt<bool>
    RunSLPVectorization("vectorize-slp", cl::init(true), cl::Hidden,
                        cl::desc("Run the SLP vectorization passes"));
 
static cl::opt<bool>
    SLPReVec("slp-revec", cl::init(false), cl::Hidden,
             cl::desc("Enable vectorization for wider vector utilization"));
 
static cl::opt<int>
    SLPCostThreshold("slp-threshold", cl::init(0), cl::Hidden,
                     cl::desc("Only vectorize if you gain more than this "
                              "number "));
 
static cl::opt<bool>
ShouldVectorizeHor("slp-vectorize-hor", cl::init(true), cl::Hidden,
                   cl::desc("Attempt to vectorize horizontal reductions"));
 
static cl::opt<bool> ShouldStartVectorizeHorAtStore(
    "slp-vectorize-hor-store", cl::init(false), cl::Hidden,
    cl::desc(
        "Attempt to vectorize horizontal reductions feeding into a store"));
 
static cl::opt<bool> SplitAlternateInstructions(
    "slp-split-alternate-instructions", cl::init(true), cl::Hidden,
    cl::desc("Improve the code quality by splitting alternate instructions"));
 
static cl::opt<bool> SLPInstCountCheck(
    "slp-inst-count-check", cl::init(true), cl::Hidden,
    cl::desc("Reject vectorization if vector instruction count exceeds "
             "scalar instruction count"));
 
static cl::opt<int>
MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden,
    cl::desc("Attempt to vectorize for this register size in bits"));
 
static cl::opt<unsigned>
MaxVFOption("slp-max-vf", cl::init(0), cl::Hidden,
    cl::desc("Maximum SLP vectorization factor (0=unlimited)"));
 
/// Limits the size of scheduling regions in a block.
/// It avoid long compile times for _very_ large blocks where vector
/// instructions are spread over a wide range.
/// This limit is way higher than needed by real-world functions.
static cl::opt<int>
ScheduleRegionSizeBudget("slp-schedule-budget", cl::init(100000), cl::Hidden,
    cl::desc("Limit the size of the SLP scheduling region per block"));
 
static cl::opt<int> MinVectorRegSizeOption(
    "slp-min-reg-size", cl::init(128), cl::Hidden,
    cl::desc("Attempt to vectorize for this register size in bits"));
 
static cl::opt<unsigned> RecursionMaxDepth(
    "slp-recursion-max-depth", cl::init(12), cl::Hidden,
    cl::desc("Limit the recursion depth when building a vectorizable tree"));
 
static cl::opt<unsigned> MinTreeSize(
    "slp-min-tree-size", cl::init(3), cl::Hidden,
    cl::desc("Only vectorize small trees if they are fully vectorizable"));
 
static cl::opt<unsigned> PHINodeVectorizationBudget(
    "slp-phi-vectorization-budget", cl::init(1024), cl::Hidden,
    cl::desc("Do not vectorize a bundle of PHI nodes if the product of the "
             "bundle size and the number of incoming values exceeds this "
             "value, to limit the compile time spent on wide PHIs"));
 
// The maximum depth that the look-ahead score heuristic will explore.
// The higher this value, the higher the compilation time overhead.
static cl::opt<int> LookAheadMaxDepth(
    "slp-max-look-ahead-depth", cl::init(2), cl::Hidden,
    cl::desc("The maximum look-ahead depth for operand reordering scores"));
 
// The maximum depth that the look-ahead score heuristic will explore
// when it probing among candidates for vectorization tree roots.
// The higher this value, the higher the compilation time overhead but unlike
// similar limit for operands ordering this is less frequently used, hence
// impact of higher value is less noticeable.
static cl::opt<int> RootLookAheadMaxDepth(
    "slp-max-root-look-ahead-depth", cl::init(2), cl::Hidden,
    cl::desc("The maximum look-ahead depth for searching best rooting option"));
 
static cl::opt<unsigned> MinProfitableStridedLoads(
    "slp-min-strided-loads", cl::init(2), cl::Hidden,
    cl::desc("The minimum number of loads, which should be considered strided, "
             "if the stride is > 1 or is runtime value"));
 
static cl::opt<unsigned> MinProfitableStridedStores(
    "slp-min-strided-stores", cl::init(2), cl::Hidden,
    cl::desc(
        "The minimum number of stores, which should be considered strided, "
        "if the stride is > 1 or is runtime value"));
 
static cl::opt<unsigned> MaxProfitableStride(
    "slp-max-stride", cl::init(8), cl::Hidden,
    cl::desc("The maximum stride, considered to be profitable."));
 
static cl::opt<bool>
    EnableStridedStores("slp-enable-strided-stores", cl::init(false),
                        cl::Hidden,
                        cl::desc("Enable SLP trees to be built from strided "
                                 "store chains."));
 
static cl::opt<bool> EnableMaskedStores(
    "slp-enable-masked-stores", cl::init(true), cl::Hidden,
    cl::desc("Enable vectorization of non-consecutive stores as a single "
             "masked store, when the target supports masked stores."));
 
static cl::opt<bool>
    DisableTreeReorder("slp-disable-tree-reorder", cl::init(false), cl::Hidden,
                       cl::desc("Disable tree reordering even if it is "
                                "profitable. Used for testing only."));
 
static cl::opt<bool>
    ForceStridedLoads("slp-force-strided-loads", cl::init(false), cl::Hidden,
                      cl::desc("Generate strided loads even if they are not "
                               "profitable. Used for testing only."));
 
static cl::opt<bool>
    ViewSLPTree("view-slp-tree", cl::Hidden,
                cl::desc("Display the SLP trees with Graphviz"));
 
static cl::opt<bool> VectorizeNonPowerOf2(
    "slp-vectorize-non-power-of-2", cl::init(false), cl::Hidden,
    cl::desc("Try to vectorize with non-power-of-2 number of elements."));
 
static cl::opt<bool> ForcePostProcessStoresOperands(
    "slp-postprocess-stores-operands", cl::init(false), cl::Hidden,
    cl::desc("Force vectorization of non-vectorizable stores operands."));
 
static cl::opt<bool> NonVectReductions(
    "slp-non-vectorizables-as-reductions", cl::init(false), cl::Hidden,
    cl::desc(
        "Use  non-vectorizable instructions as potential reduction roots."));
 
static cl::opt<bool> VectorizePoorThroughput(
    "slp-vectorize-poor-throughput", cl::init(true), cl::Hidden,
    cl::desc("Use poor-throughput instructions (e.g. fdiv, frem, fsqrt) as "
             "standalone vectorization seeds."));
 
static cl::opt<bool> VectorizeOnceUsed(
    "slp-vectorize-once-used", cl::init(true), cl::Hidden,
    cl::desc("Use instructions with the single user as standalone "
             "vectorization seeds."));
 
/// True when \p slp-vectorize-non-power-of-2 is enabled and \p NumElts is a
/// supported non-power-of-2 width: \p NumElts + 1 must be a power of two
/// (e.g. 3 or 7 lanes, i.e. almost a full power-of-2 register).
static bool isAllowedNonPowerOf2VF(unsigned NumElts) {
  return VectorizeNonPowerOf2 && has_single_bit(NumElts + 1);
}
 
/// Enables vectorization of copyable elements.
static cl::opt<bool> VectorizeCopyableElements(
    "slp-copyable-elements", cl::init(true), cl::Hidden,
    cl::desc("Try to replace values with the idempotent instructions for "
             "better vectorization."));
 
/// Gather operands of associative single-use binary chains into one node.
static cl::opt<bool> VectorizeReassociatedOps(
    "slp-reassociate-ops", cl::init(true), cl::Hidden,
    cl::desc("Gather operands of associative binary chains into one node."));
 
/// The family-realigned seed already groups the vectorizable columns; the
/// VLOperands polish on top is quadratic in the column count, so past this
/// many columns keep the seed instead.
static cl::opt<unsigned> ReassocReorderColumnLimit(
    "slp-reassociate-reorder-limit", cl::init(32), cl::Hidden,
    cl::desc("Max flattened operand columns for which associative-chain "
             "reordering runs the full operand reorder."));
 
static cl::opt<unsigned> LoopAwareTripCount(
    "slp-cost-loop-trip-count", cl::init(2), cl::Hidden,
    cl::desc("Loop trip count, considered by the cost model during "
             "modeling (0=loops are ignored and considered flat code)"));
 
/// Refine the loop-aware cost scaling of gather/buildvector tree entries by
/// using the per-lane execution scale of the operand that feeds each lane,
/// instead of a single whole-entry scale. This matches the LICM hoisting
/// performed by optimizeGatherSequence() at codegen time: lanes whose
/// operands are loop-invariant in an inner loop contribute the outer loop's
/// execution scale rather than the inner loop's, which avoids over-costing
/// buildvectors that bridge values from outer loop nests into an inner loop.
static cl::opt<bool> PerLaneGatherScale(
    "slp-per-lane-gather-scale", cl::init(true), cl::Hidden,
    cl::desc("Use per-lane execution scale for gather/buildvector tree "
             "entries to model LICM-hoistable buildvector sequences."));
 
/// Enable versioning of a basic block with runtime alias checks.
static cl::opt<bool> SLPEnableRuntimeAliasChecks(
    "slp-vectorize-with-runtime-alias-checks", cl::init(true), cl::Hidden,
    cl::desc("Allow SLP to version a block with runtime alias checks to "
             "vectorize trees blocked by may-alias memory dependencies."));
 
/// Maximum number of runtime alias checks (one per pair of base objects) that
/// may guard a single versioned region.
static cl::opt<unsigned> SLPMaxRuntimeAliasChecks(
    "slp-max-runtime-alias-checks", cl::init(8), cl::Hidden,
    cl::desc("The maximum number of runtime alias checks generated to guard a "
             "single SLP-vectorized region."));
 
/// The runtime checks and the guard branch execute on both the vector and the
/// scalar fallback path, so they add overhead to the scalar code.
static cl::opt<unsigned> SLPRuntimeAliasChecksMaxScalarCostPercent(
    "slp-runtime-alias-checks-max-scalar-cost-percent", cl::init(25),
    cl::Hidden,
    cl::desc("Maximum SLP runtime alias check cost, as a percentage of the "
             "guarded scalar region cost, before versioning is rejected to "
             "avoid pessimizing the scalar fallback path."));
 
// Limit the number of alias checks. The limit is chosen so that
// it has no negative effect on the llvm benchmarks.
static const unsigned AliasedCheckLimit = 10;
 
// Another limit for the alias checks: The maximum distance between load/store
// instructions where alias checks are done.
// This limit is useful for very large basic blocks.
static const unsigned MaxMemDepDistance = 160;
 
/// If the ScheduleRegionSizeBudget is exhausted, we allow small scheduling
/// regions to be handled.
static const int MinScheduleRegionSize = 16;
 
/// Maximum allowed number of operands in the PHI nodes.
static const unsigned MaxPHINumOperands = 128;
 
/// Checks if the vector of instructions can be represented as a shuffle, like:
/// %x0 = extractelement <4 x i8> %x, i32 0
/// %x3 = extractelement <4 x i8> %x, i32 3
/// %y1 = extractelement <4 x i8> %y, i32 1
/// %y2 = extractelement <4 x i8> %y, i32 2
/// %x0x0 = mul i8 %x0, %x0
/// %x3x3 = mul i8 %x3, %x3
/// %y1y1 = mul i8 %y1, %y1
/// %y2y2 = mul i8 %y2, %y2
/// %ins1 = insertelement <4 x i8> poison, i8 %x0x0, i32 0
/// %ins2 = insertelement <4 x i8> %ins1, i8 %x3x3, i32 1
/// %ins3 = insertelement <4 x i8> %ins2, i8 %y1y1, i32 2
/// %ins4 = insertelement <4 x i8> %ins3, i8 %y2y2, i32 3
/// ret <4 x i8> %ins4
/// can be transformed into:
/// %1 = shufflevector <4 x i8> %x, <4 x i8> %y, <4 x i32> <i32 0, i32 3, i32 5,
///                                                         i32 6>
/// %2 = mul <4 x i8> %1, %1
/// ret <4 x i8> %2
/// Mask will return the Shuffle Mask equivalent to the extracted elements.
/// TODO: Can we split off and reuse the shuffle mask detection from
/// ShuffleVectorInst/getShuffleCost?
static std::optional<TargetTransformInfo::ShuffleKind>
isFixedVectorShuffle(ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask,
                     AssumptionCache *AC) {
  const auto *It = find_if(VL, IsaPred<ExtractElementInst>);
  if (It == VL.end())
    return std::nullopt;
  unsigned Size = accumulate(VL, 0u, [](unsigned S, Value *V) {
    auto *EI = dyn_cast<ExtractElementInst>(V);
    if (!EI)
      return S;
    auto *VTy = dyn_cast<FixedVectorType>(EI->getVectorOperandType());
    if (!VTy)
      return S;
    return std::max(S, VTy->getNumElements());
  });
 
  Value *Vec1 = nullptr;
  Value *Vec2 = nullptr;
  bool HasNonUndefVec = any_of(VL, [&](Value *V) {
    auto *EE = dyn_cast<ExtractElementInst>(V);
    if (!EE)
      return false;
    Value *Vec = EE->getVectorOperand();
    if (isa<UndefValue>(Vec))
      return false;
    return isGuaranteedNotToBePoison(Vec, AC);
  });
  enum ShuffleMode { Unknown, Select, Permute };
  ShuffleMode CommonShuffleMode = Unknown;
  Mask.assign(VL.size(), PoisonMaskElem);
  for (unsigned I = 0, E = VL.size(); I < E; ++I) {
    // Undef, or a copyable lane modeled on an extract main op, can be
    // represented as an undef element in a vector.
    if (isa<UndefValue>(VL[I]))
      continue;
    auto *EI = dyn_cast<ExtractElementInst>(VL[I]);
    if (!EI)
      continue;
    if (isa<ScalableVectorType>(EI->getVectorOperandType()))
      return std::nullopt;
    auto *Vec = EI->getVectorOperand();
    // We can extractelement from undef or poison vector.
    if (isUndefVector</*isPoisonOnly=*/true>(Vec).all())
      continue;
    // All vector operands must have the same number of vector elements.
    if (isa<UndefValue>(Vec)) {
      Mask[I] = I;
    } else {
      if (isa<UndefValue>(EI->getIndexOperand()))
        continue;
      auto *Idx = dyn_cast<ConstantInt>(EI->getIndexOperand());
      if (!Idx)
        return std::nullopt;
      // Undefined behavior if Idx is negative or >= Size.
      if (Idx->getValue().uge(Size))
        continue;
      unsigned IntIdx = Idx->getValue().getZExtValue();
      Mask[I] = IntIdx;
    }
    if (isUndefVector(Vec).all() && HasNonUndefVec)
      continue;
    // For correct shuffling we have to have at most 2 different vector operands
    // in all extractelement instructions.
    if (!Vec1 || Vec1 == Vec) {
      Vec1 = Vec;
    } else if (!Vec2 || Vec2 == Vec) {
      Vec2 = Vec;
      Mask[I] += Size;
    } else {
      return std::nullopt;
    }
    if (CommonShuffleMode == Permute)
      continue;
    // If the extract index is not the same as the operation number, it is a
    // permutation.
    if (Mask[I] % Size != I) {
      CommonShuffleMode = Permute;
      continue;
    }
    CommonShuffleMode = Select;
  }
  // If we're not crossing lanes in different vectors, consider it as blending.
  if (CommonShuffleMode == Select && Vec2)
    return TargetTransformInfo::SK_Select;
  // If Vec2 was never used, we have a permutation of a single vector, otherwise
  // we have permutation of 2 vectors.
  return Vec2 ? TargetTransformInfo::SK_PermuteTwoSrc
              : TargetTransformInfo::SK_PermuteSingleSrc;
}
 
/// Returns number of parts, the type \p VecTy will be split at the codegen
/// phase. If the type is going to be scalarized or does not uses whole
/// registers, returns 1.
static unsigned
getNumberOfParts(const TargetTransformInfo &TTI, Type *VecTy, Type *ScalarTy,
                 const unsigned Limit = std::numeric_limits<unsigned>::max()) {
  if (isa<StructType>(VecTy))
    return 1;
  unsigned NumParts = TTI.getNumberOfParts(VecTy);
  if (NumParts == 0 || NumParts >= Limit)
    return 1;
  unsigned Sz = getNumElements(VecTy);
  unsigned ScalarSz = getNumElements(ScalarTy);
  Type *ElementTy = toScalarizedTy(VecTy);
  unsigned PWSz = getFullVectorNumberOfElements(TTI, ElementTy, Sz, SLPReVec);
  if (NumParts >= Sz || PWSz % NumParts != 0 ||
      (PWSz / NumParts) % ScalarSz != 0 ||
      !hasFullVectorsOrPowerOf2(TTI, ElementTy, PWSz / NumParts, SLPReVec))
    return 1;
  const unsigned NumElts = PWSz / NumParts;
  if (divideCeil(Sz, NumElts) != NumParts)
    return 1;
  return NumParts;
}
 
/// Bottom Up SLP Vectorizer.
class slpvectorizer::BoUpSLP {
  class TreeEntry;
  class ScheduleEntity;
  class ScheduleData;
  class ScheduleCopyableData;
  class ScheduleBundle;
  class ShuffleCostEstimator;
  class ShuffleInstructionBuilder;
 
public:
  /// If we decide to generate strided load / store, this struct contains all
  /// the necessary info. It's fields are calculated by analyzeRtStrideCandidate
  /// and analyzeConstantStrideCandidate. Note that Stride can be given either
  /// as a SCEV or as a Value if it already exists. To get the stride in bytes,
  /// StrideVal (or value obtained from StrideSCEV) has to by multiplied by the
  /// size of element of FixedVectorType.
  struct StridedPtrInfo {
    Value *StrideVal = nullptr;
    const SCEV *StrideSCEV = nullptr;
    FixedVectorType *Ty = nullptr;
  };
 
  /// Tracks the state we can represent the loads in the given sequence.
  enum class LoadsState {
    Gather,
    Vectorize,
    ScatterVectorize,
    StridedVectorize,
    CompressVectorize,
    BlendedLoadVectorize
  };
 
  using ValueList = SmallVector<Value *, 8>;
  using InstrList = SmallVector<Instruction *, 16>;
  using ValueSet = SmallPtrSet<Value *, 16>;
  using StoreList = SmallVector<StoreInst *, 8>;
  using ExtraValueToDebugLocsMap = SmallDenseSet<Value *, 4>;
  using OrdersType = SmallVector<unsigned, 4>;
 
  BoUpSLP(Function *Func, ScalarEvolution *Se, TargetTransformInfo *Tti,
          TargetLibraryInfo *TLi, AAResults *Aa, LoopInfo *Li,
          DominatorTree *Dt, AssumptionCache *AC, DemandedBits *DB,
          const DataLayout *DL, OptimizationRemarkEmitter *ORE)
      : BatchAA(*Aa), F(Func), SE(Se), TTI(Tti), TLI(TLi), LI(Li), DT(Dt),
        AC(AC), DB(DB), DL(DL), ORE(ORE), CostKind(getSLPCostKind(Func)),
        Builder(Se->getContext(), TargetFolder(*DL)) {
    CodeMetrics::collectEphemeralValues(F, AC, EphValues);
    // Use the vector register size specified by the target unless overridden
    // by a command-line option.
    // TODO: It would be better to limit the vectorization factor based on
    //       data type rather than just register size. For example, x86 AVX has
    //       256-bit registers, but it does not support integer operations
    //       at that width (that requires AVX2).
    if (MaxVectorRegSizeOption.getNumOccurrences())
      MaxVecRegSize = MaxVectorRegSizeOption;
    else
      MaxVecRegSize =
          TTI->getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
              .getFixedValue();
 
    if (MinVectorRegSizeOption.getNumOccurrences())
      MinVecRegSize = MinVectorRegSizeOption;
    else
      MinVecRegSize = TTI->getMinVectorRegisterBitWidth();
  }
 
  /// Vectorize the tree that starts with the elements in \p VL.
  /// Returns the vectorized root.
  Value *vectorizeTree();
 
  /// Vectorize the tree but with the list of externally used values \p
  /// ExternallyUsedValues. Values in this MapVector can be replaced but the
  /// generated extractvalue instructions.
  Value *
  vectorizeTree(const ExtraValueToDebugLocsMap &ExternallyUsedValues,
                Instruction *ReductionRoot = nullptr,
                ArrayRef<std::tuple<WeakTrackingVH, unsigned, bool, bool>>
                    VectorValuesAndScales = {});
 
  /// \returns the cost incurred by unwanted spills and fills, caused by
  /// holding live values over call sites.
  InstructionCost getSpillCost();
 
  TargetTransformInfo::TargetCostKind getCostKind() const { return CostKind; }
 
  /// Calculates the cost of the subtrees, trims non-profitable ones and returns
  /// final cost.
  InstructionCost
  calculateTreeCostAndTrimNonProfitable(ArrayRef<Value *> VectorizedVals = {},
                                        Instruction *RdxRoot = nullptr);
 
  /// \returns the vectorization cost of the subtree that starts at \p VL.
  /// A negative number means that this is profitable.
  InstructionCost getTreeCost(InstructionCost TreeCost,
                              ArrayRef<Value *> VectorizedVals = {},
                              InstructionCost ReductionCost = TTI::TCC_Free,
                              Instruction *RdxRoot = nullptr);
 
  /// Construct a vectorizable tree that starts at \p Roots, ignoring users for
  /// the purpose of scheduling and extraction in the \p UserIgnoreLst.
  void buildTree(ArrayRef<Value *> Roots,
                 const SmallDenseSet<Value *> &UserIgnoreLst);
 
  /// Construct a vectorizable tree that starts at \p Roots.
  void buildTree(ArrayRef<Value *> Roots);
 
  /// Sets the narrowed reduction chain instructions, dropped together with
  /// the reduction.
  void setNarrowedChainInsts(ArrayRef<Instruction *> Insts) {
    NarrowedChainInsts.insert(Insts.begin(), Insts.end());
  }
 
  /// Returns true if the last buildTree() observed a may-alias memory
  /// dependency between two distinct, range-checkable base objects, i.e. a
  /// dependency that could be turned into a runtime alias check.
  bool hasRuntimeCheckableBlockers() const {
    return HasRuntimeCheckableBlockers;
  }
 
  /// Records whether a may-alias dependency between distinct, range-checkable
  /// base objects has been observed, so the caller can decide to retry with
  /// runtime alias checks enabled.
  void setHasRuntimeCheckableBlockers(bool V) {
    HasRuntimeCheckableBlockers = V;
  }
 
  /// Returns true if the last buildTree() kept a may-alias memory dependency
  /// that is not runtime-checkable (call or a non-simple mem access). Such a
  /// dependency cannot be dropped, so a runtime-checks retry cannot unblock the
  /// region and would be pure overhead.
  bool hasNonCheckableMemBlocker() const { return HasNonCheckableMemBlocker; }
 
  /// Records that a non-runtime-checkable may-alias dependency was kept.
  void setHasNonCheckableMemBlocker(bool V) { HasNonCheckableMemBlocker = V; }
 
  /// Returns true if the current vectorization attempt may drop
  /// runtime-checkable may-alias dependencies and guard the region with
  /// runtime alias checks.
  bool isTryingRuntimeAliasChecks() const { return TryRuntimeAliasChecks; }
 
  /// Enables or disables dropping runtime-checkable may-alias dependencies in
  /// favor of runtime alias checks for the current vectorization attempt.
  void setTryRuntimeAliasChecks(bool V) { TryRuntimeAliasChecks = V; }
 
  /// Resets the runtime alias check data.
  void resetRuntimeAliasCheckState() {
    HasRuntimeCheckableBlockers = false;
    HasNonCheckableMemBlocker = false;
    RTChecksFinalized = false;
    RTChecks.clear();
    RTOrigBodyOrder.clear();
  }
 
  /// Snapshots RTChecks.BB's body (non-PHI, non-terminator) into
  /// RTOrigBodyOrder in program order, for the scalar fallback.
  void captureRuntimeCheckBodySnapshot();
 
  /// Returns true if \p BB satisfies the block-level preconditions for runtime
  /// alias check versioning (straight-line, outside any loop, duplicable, not a
  /// scalar fallback, function not optimized for size). These checks do not
  /// depend on the collected checks, so they can gate the (expensive)
  /// optimistic retry before any tree is rebuilt.
  bool canVersionBlockForRuntimeChecks(BasicBlock *BB) const;
 
  /// Returns true if the runtime alias checks can be safely emitted to guard
  /// the vectorized region.
  bool canVersionForRuntimeChecks();
 
  /// Returns true if \p BB is a scalar fallback block created by runtime alias
  /// check versioning.
  bool isScalarFallbackBlock(BasicBlock *BB) const {
    return ScalarFallbackBlocks.contains(BB);
  }
 
  /// Returns true if an optimistic runtime-checks versioning attempt already
  /// failed for \p BB, so further retries in the same block can be skipped.
  bool runtimeChecksFailedForBlock(BasicBlock *BB) const {
    return FailedRuntimeChecksBlocks.contains(BB);
  }
 
  /// Records that an optimistic runtime-checks versioning attempt failed for
  /// \p BB.
  void markRuntimeChecksFailedForBlock(BasicBlock *BB) {
    FailedRuntimeChecksBlocks.insert(BB);
  }
 
  /// Returns the modeled cost of the runtime alias checks collected during the
  /// last (optimistic) buildTree().
  InstructionCost getRuntimeChecksCost() const;
 
  /// Returns true if the last (optimistic) buildTree() collected any runtime
  /// alias checks that must guard the vectorized region.
  bool hasRuntimeAliasChecks() const { return !RTChecks.BasePairs.empty(); }
 
  /// Returns true if vectorization changed the CFG (i.e. a block was versioned
  /// with runtime alias checks). When true, CFG analyses must not be preserved.
  bool isCFGChanged() const { return CFGChanged; }
 
  TreeEntry &getRootNode() {
    assert(!VectorizableTree.empty() && "No graph to get the first node from");
    return *VectorizableTree.front();
  }
 
  const TreeEntry &getRootNode() const {
    assert(!VectorizableTree.empty() && "No graph to get the first node from");
    return *VectorizableTree.front();
  }
 
  /// Returns the scalars of the root node.
  ArrayRef<Value *> getRootNodeScalars() const { return getRootNode().Scalars; }
 
  /// Returns the lane the given value is vectorized to in the root node.
  unsigned findRootLaneForValue(Value *V) const {
    return getRootNode().findLaneForValue(V);
  }
 
  /// Returns the type/is-signed info for the root node in the graph without
  /// casting.
  std::optional<std::pair<Type *, bool>> getRootNodeTypeWithNoCast() const {
    const TreeEntry &Root = getRootNode();
    if (Root.State != TreeEntry::Vectorize || Root.isAltShuffle() ||
        !Root.Scalars.front()->getType()->isIntegerTy())
      return std::nullopt;
    auto It = MinBWs.find(&Root);
    if (It != MinBWs.end())
      return std::make_pair(IntegerType::get(Root.Scalars.front()->getContext(),
                                             It->second.first),
                            It->second.second);
    if (Root.getOpcode() == Instruction::ZExt ||
        Root.getOpcode() == Instruction::SExt)
      return std::make_pair(cast<CastInst>(Root.getMainOp())->getSrcTy(),
                            Root.getOpcode() == Instruction::SExt);
    return std::nullopt;
  }
 
  /// Checks if the root graph node can be emitted with narrower bitwidth at
  /// codegen and returns it signedness, if so.
  bool isSignedMinBitwidthRootNode() const {
    return MinBWs.at(&getRootNode()).second;
  }
 
  /// Returns reduction type after minbitdth analysis.
  FixedVectorType *getReductionType() const {
    if (ReductionBitWidth == 0 ||
        !getRootNodeScalars().front()->getType()->isIntegerTy() ||
        ReductionBitWidth >=
            DL->getTypeSizeInBits(getRootNodeScalars().front()->getType()))
      return cast<FixedVectorType>(
          getWidenedType(getRootNodeScalars().front()->getType(),
                         getRootNode().getVectorFactor()));
    return cast<FixedVectorType>(getWidenedType(
        IntegerType::get(getRootNodeScalars().front()->getContext(),
                         ReductionBitWidth),
        getRootNode().getVectorFactor()));
  }
 
  /// Returns true if the tree results in one of the reduced bitcasts variants.
  bool isReducedBitcastRoot() const {
    return getRootNode().hasState() &&
           (getRootNode().CombinedOp == TreeEntry::ReducedBitcast ||
            getRootNode().CombinedOp == TreeEntry::ReducedBitcastBSwap ||
            getRootNode().CombinedOp == TreeEntry::ReducedBitcastLoads ||
            getRootNode().CombinedOp == TreeEntry::ReducedBitcastBSwapLoads) &&
           getRootNode().State == TreeEntry::Vectorize;
  }
 
  /// Returns true if the tree results in the reduced cmp bitcast root.
  bool isReducedCmpBitcastRoot() const {
    return getRootNode().hasState() &&
           getRootNode().CombinedOp == TreeEntry::ReducedCmpBitcast &&
           getRootNode().State == TreeEntry::Vectorize;
  }
 
  /// Returns true if the tree is a reduction tree.
  bool isReductionTree() const { return UserIgnoreList != nullptr; }
 
  /// Builds external uses of the vectorized scalars, i.e. the list of
  /// vectorized scalars to be extracted, their lanes and their scalar users. \p
  /// ExternallyUsedValues contains additional list of external uses to handle
  /// vectorization of reductions.
  void
  buildExternalUses(const ExtraValueToDebugLocsMap &ExternallyUsedValues = {});
 
  /// Transforms graph nodes to target specific representations, if profitable.
  void transformNodes();
 
  /// Clear the internal data structures that are created by 'buildTree'.
  void deleteTree() {
    VectorizableTree.clear();
    ScalarToTreeEntries.clear();
    DeletedNodes.clear();
    TransformedToGatherNodes.clear();
    OperandsToTreeEntry.clear();
    ScalarsInSplitNodes.clear();
    MustGather.clear();
    ReassocScalarToTreeEntries.clear();
    KeptReassocScalars.clear();
    NonScheduledFirst.clear();
    EntryToLastInstruction.clear();
    LastInstructionToPos.clear();
    LoadEntriesToVectorize.clear();
    IsGraphTransformMode = false;
    GatheredLoadsEntriesFirst.reset();
    SplatGatheredScalarsRoots.clear();
    CompressEntryToData.clear();
    ExternalUses.clear();
    ExternalUsesAsOriginalScalar.clear();
    ExternalUsesWithNonUsers.clear();
    ExternalUseReplacements.clear();
    RTChecks.clear();
    HasRuntimeCheckableBlockers = false;
    HasNonCheckableMemBlocker = false;
    RTChecksFinalized = false;
    for (auto &Iter : BlocksSchedules) {
      BlockScheduling *BS = Iter.second.get();
      BS->clear();
    }
    MinBWs.clear();
    ReductionBitWidth = 0;
    BaseGraphSize = 1;
    CastMaxMinBWSizes.reset();
    ExtraBitWidthNodes.clear();
    InstrElementSize.clear();
    UserIgnoreList = nullptr;
    NarrowedChainInsts.clear();
    PostponedGathers.clear();
    ValueToGatherNodes.clear();
    TreeEntryToStridedPtrInfoMap.clear();
    CurrentLoopNest.clear();
    MergedLoopBTCs.clear();
  }
 
  unsigned getTreeSize() const { return VectorizableTree.size(); }
 
  /// Returns the base graph size, before any transformations.
  unsigned getCanonicalGraphSize() const { return BaseGraphSize; }
 
  /// Perform LICM and CSE on the newly generated gather sequences.
  void optimizeGatherSequence();
 
  /// Does this non-empty order represent an identity order?  Identity
  /// should be represented as an empty order, so this is used to
  /// decide if we can canonicalize a computed order.  Undef elements
  /// (represented as size) are ignored.
  static bool isIdentityOrder(ArrayRef<unsigned> Order) {
    assert(!Order.empty() && "expected non-empty order");
    const unsigned Sz = Order.size();
    return all_of(enumerate(Order), [&](const auto &P) {
      return P.value() == P.index() || P.value() == Sz;
    });
  }
 
  /// Checks if the specified gather tree entry \p TE can be represented as a
  /// shuffled vector entry + (possibly) permutation with other gathers. It
  /// implements the checks only for possibly ordered scalars (Loads,
  /// ExtractElement, ExtractValue), which can be part of the graph.
  /// \param TopToBottom If true, used for the whole tree rotation, false - for
  /// sub-tree rotations. \param IgnoreReorder true, if the order of the root
  /// node might be ignored.
  std::optional<OrdersType> findReusedOrderedScalars(const TreeEntry &TE,
                                                     bool TopToBottom,
                                                     bool IgnoreReorder);
 
  /// Sort loads into increasing pointers offsets to allow greater clustering.
  std::optional<OrdersType> findPartiallyOrderedLoads(const TreeEntry &TE);
 
  /// Gets reordering data for the given tree entry. If the entry is vectorized
  /// - just return ReorderIndices, otherwise check if the scalars can be
  /// reordered and return the most optimal order.
  /// \return std::nullopt if ordering is not important, empty order, if
  /// identity order is important, or the actual order.
  /// \param TopToBottom If true, include the order of vectorized stores and
  /// insertelement nodes, otherwise skip them.
  /// \param IgnoreReorder true, if the root node order can be ignored.
  std::optional<OrdersType>
  getReorderingData(const TreeEntry &TE, bool TopToBottom, bool IgnoreReorder);
 
  /// Checks if it is profitable to reorder the current tree.
  /// If the tree does not contain many profitable reordable nodes, better to
  /// skip it to save compile time.
  bool isProfitableToReorder() const;
 
  /// Reorders the current graph to the most profitable order starting from the
  /// root node to the leaf nodes. The best order is chosen only from the nodes
  /// of the same size (vectorization factor). Smaller nodes are considered
  /// parts of subgraph with smaller VF and they are reordered independently. We
  /// can make it because we still need to extend smaller nodes to the wider VF
  /// and we can merge reordering shuffles with the widening shuffles.
  void reorderTopToBottom();
 
  /// Reorders the current graph to the most profitable order starting from
  /// leaves to the root. It allows to rotate small subgraphs and reduce the
  /// number of reshuffles if the leaf nodes use the same order. In this case we
  /// can merge the orders and just shuffle user node instead of shuffling its
  /// operands. Plus, even the leaf nodes have different orders, it allows to
  /// sink reordering in the graph closer to the root node and merge it later
  /// during analysis.
  void reorderBottomToTop(bool IgnoreReorder = false);
 
  /// Marks the schedule data of the copyable-modeled operands of \p TE for
  /// dependency recalculation at the next bundle scheduling.
  void markCopyableDepsForRecalc(TreeEntry &TE);
 
  /// \return The vector element size in bits to use when vectorizing the
  /// expression tree ending at \p V. If V is a store, the size is the width of
  /// the stored value. Otherwise, the size is the width of the largest loaded
  /// value reaching V. This method is used by the vectorizer to calculate
  /// vectorization factors.
  unsigned getVectorElementSize(Value *V);
 
  /// Compute the minimum type sizes required to represent the entries in a
  /// vectorizable tree.
  void computeMinimumValueSizes();
 
  // \returns maximum vector register size as set by TTI or overridden by cl::opt.
  unsigned getMaxVecRegSize() const {
    return MaxVecRegSize;
  }
 
  // \returns minimum vector register size as set by cl::opt.
  unsigned getMinVecRegSize() const {
    return MinVecRegSize;
  }
 
  /// \returns the number of parts, the type \p VecTy is split at the codegen
  /// phase. The type legalization queries are repeated for the very same types
  /// during the analysis, so the results are cached for the function.
  unsigned getNumberOfParts(
      Type *VecTy, Type *ScalarTy,
      unsigned Limit = std::numeric_limits<unsigned>::max()) const {
    auto [It, Inserted] =
        NumberOfPartsCache.try_emplace(std::make_tuple(VecTy, ScalarTy, Limit));
    if (Inserted)
      It->second = ::getNumberOfParts(*TTI, VecTy, ScalarTy, Limit);
    return It->second;
  }
 
  unsigned getMinVF(unsigned Sz) const {
    return std::max(2U, getMinVecRegSize() / Sz);
  }
 
  unsigned getMaximumVF(unsigned ElemWidth, unsigned Opcode) const {
    unsigned MaxVF = MaxVFOption.getNumOccurrences() ?
      MaxVFOption : TTI->getMaximumVF(ElemWidth, Opcode);
    return MaxVF ? MaxVF : UINT_MAX;
  }
 
  /// Check if homogeneous aggregate is isomorphic to some VectorType.
  /// Accepts homogeneous multidimensional aggregate of scalars/vectors like
  /// {[4 x i16], [4 x i16]}, { <2 x float>, <2 x float> },
  /// {{{i16, i16}, {i16, i16}}, {{i16, i16}, {i16, i16}}} and so on.
  ///
  /// \returns number of elements in vector if isomorphism exists, 0 otherwise.
  unsigned canMapToVector(Type *T) const;
 
  /// \returns true if the vectorized insertvalue result can be stored directly
  /// as a vector, i.e. every insertvalue with an external user is consumed by a
  /// single store only.
  bool canVectorStoreInsertValue(const TreeEntry *E) const;
 
  /// \returns the source vector type for an InsertElement/InsertValue
  /// buildvector node \p E: the inserted vector type for insertelement, or a
  /// vector of the inserted scalar type wide enough to cover the highest
  /// inserted index for insertvalue.
  FixedVectorType *getInsertBuildVectorSrcTy(const TreeEntry *E) const;
 
  /// \returns True if the VectorizableTree is both tiny and not fully
  /// vectorizable. We do not vectorize such trees.
  bool isTreeTinyAndNotFullyVectorizable(bool ForReduction = false) const;
 
  /// Checks if the graph and all its subgraphs cannot be better vectorized.
  /// It may happen, if all gather nodes are loads and they cannot be
  /// "clusterized". In this case even subgraphs cannot be vectorized more
  /// effectively than the base graph.
  bool isTreeNotExtendable() const;
 
  bool isStridedLoad(ArrayRef<Value *> PointerOps, Type *ScalarTy,
                     Align Alignment, const int64_t Diff,
                     const size_t Sz) const;
 
  /// Return true if an array of scalar loads can be replaced with a strided
  ///  load (with constant stride).
  ///
  ///  It is possible that the load gets "widened". Suppose that originally each
  ///  load loads `k` bytes and `PointerOps` can be arranged as follows (`%s` is
  ///  constant): %b + 0 * %s + 0 %b + 0 * %s + 1 %b + 0 * %s + 2
  ///  ...
  ///  %b + 0 * %s + (w - 1)
  ///
  ///  %b + 1 * %s + 0
  ///  %b + 1 * %s + 1
  ///  %b + 1 * %s + 2
  ///  ...
  ///  %b + 1 * %s + (w - 1)
  ///  ...
  ///
  ///  %b + (n - 1) * %s + 0
  ///  %b + (n - 1) * %s + 1
  ///  %b + (n - 1) * %s + 2
  ///  ...
  ///  %b + (n - 1) * %s + (w - 1)
  ///
  /// In this case we will generate a strided load of type `<n x (k * w)>`.
  ///
  /// \param PointerOps list of pointer arguments of loads.
  /// \param ElemTy original scalar type of loads.
  /// \param Alignment alignment of the first load.
  /// \param SortedIndices is the order of PointerOps as returned by
  /// `sortPtrAccesses`
  /// \param Diff Pointer difference between the lowest and the highes pointer
  /// in `PointerOps` as returned by `getPointersDiff`.
  /// \param Ptr0 first pointer in `PointersOps`.
  /// \param PtrN last pointer in `PointersOps`.
  /// \param SPtrInfo If the function return `true`, it also sets all the fields
  /// of `SPtrInfo` necessary to generate the strided load later.
  bool analyzeConstantStrideCandidate(
      const ArrayRef<Value *> PointerOps, Type *ElemTy, Align Alignment,
      const SmallVectorImpl<unsigned> &SortedIndices, const int64_t Diff,
      Value *Ptr0, StridedPtrInfo &SPtrInfo) const;
 
  /// Return true if an array of scalar loads can be replaced with a strided
  /// load (with run-time stride).
  /// \param PointerOps list of pointer arguments of loads.
  /// \param ScalarTy type of loads.
  /// \param CommonAlignment common alignement of loads as computed by
  /// `computeCommonAlignment<LoadInst>`.
  /// \param SortedIndicies is a list of indicies computed by this function such
  /// that the sequence `PointerOps[SortedIndices[0]],
  /// PointerOps[SortedIndicies[1]], ..., PointerOps[SortedIndices[n]]` is
  /// ordered by the coefficient of the stride. For example, if PointerOps is
  /// `%base + %stride, %base, %base + 2 * stride` the `SortedIndices` will be
  /// `[1, 0, 2]`. We follow the convention that if `SortedIndices` has to be
  /// `0, 1, 2, 3, ...` we return empty vector for `SortedIndicies`.
  /// \param SPtrInfo If the function return `true`, it also sets all the fields
  /// of `SPtrInfo` necessary to generate the strided load later.
  /// \param IsLoad Is this a strided load (true) or strided store (false)
  bool analyzeRtStrideCandidate(ArrayRef<Value *> PointerOps, Type *ScalarTy,
                                Align CommonAlignment,
                                SmallVectorImpl<unsigned> &SortedIndices,
                                StridedPtrInfo &SPtrInfo, bool IsLoad) const;
 
  /// Checks if the given array of loads can be represented as a vectorized,
  /// scatter or just simple gather.
  /// \param VL list of loads.
  /// \param VL0 main load value.
  /// \param Order returned order of load instructions.
  /// \param PointerOps returned list of pointer operands.
  /// \param BestVF return best vector factor, if recursive check found better
  /// vectorization sequences rather than masked gather.
  /// \param TryRecursiveCheck used to check if long masked gather can be
  /// represented as a serie of loads/insert subvector, if profitable.
  LoadsState canVectorizeLoads(ArrayRef<Value *> VL, const Value *VL0,
                               SmallVectorImpl<unsigned> &Order,
                               SmallVectorImpl<Value *> &PointerOps,
                               StridedPtrInfo &SPtrInfo,
                               unsigned *BestVF = nullptr,
                               bool TryRecursiveCheck = true) const;
 
  /// Checks whether some existing tree entry has scalars equal to \p VL.
  /// \p S is the common opcode of \p VL when one exists; an empty \p S means
  /// the values have no common opcode (mixed buildvector/gather candidates).
  bool hasSameNode(const InstructionsState &S, ArrayRef<Value *> VL) const {
    auto IsSame = [&](const TreeEntry *TE) { return TE->isSame(VL); };
    if (S) {
      // Any vectorized or gather entry equal to VL must contain S.getMainOp()
      // (the representative instruction, which is also the recorded scalar
      // for copyable-elements bundles), so probing the MainOp-indexed maps
      // is sufficient and avoids scanning the whole tree.
      return any_of(getTreeEntries(S.getMainOp()), IsSame) ||
             any_of(ValueToGatherNodes.lookup(S.getMainOp()), IsSame);
    }
    // No common opcode: only gather entries can match. Each non-constant
    // value in VL has to be in the gather entry's scalar list and is
    // therefore present in ValueToGatherNodes. Probe by VL members instead
    // of scanning the whole tree (O(tree) -> O(|VL|)).
    SmallPtrSet<const TreeEntry *, 4> Visited;
    for (Value *V : VL) {
      // Constants/poisons are not tracked in ValueToGatherNodes.
      if (isConstant(V))
        continue;
      for (const TreeEntry *TE : ValueToGatherNodes.lookup(V)) {
        if (!Visited.insert(TE).second)
          continue;
        if (IsSame(TE))
          return true;
      }
    }
    return false;
  }
 
  /// Registers non-vectorizable sequence of loads
  template <typename T> void registerNonVectorizableLoads(ArrayRef<T *> VL) {
    ListOfKnonwnNonVectorizableLoads.insert(hash_value(VL));
  }
 
  /// Checks if the given loads sequence is known as not vectorizable
  template <typename T>
  bool areKnownNonVectorizableLoads(ArrayRef<T *> VL) const {
    return ListOfKnonwnNonVectorizableLoads.contains(hash_value(VL));
  }
 
  OptimizationRemarkEmitter *getORE() { return ORE; }
 
  /// This structure holds any data we need about the edges being traversed
  /// during buildTreeRec(). We keep track of:
  /// (i) the user TreeEntry index, and
  /// (ii) the index of the edge.
  struct EdgeInfo {
    EdgeInfo() = default;
    EdgeInfo(TreeEntry *UserTE, unsigned EdgeIdx)
        : UserTE(UserTE), EdgeIdx(EdgeIdx) {}
    /// The user TreeEntry.
    TreeEntry *UserTE = nullptr;
    /// The operand index of the use.
    unsigned EdgeIdx = UINT_MAX;
#ifndef NDEBUG
    friend inline raw_ostream &operator<<(raw_ostream &OS,
                                          const BoUpSLP::EdgeInfo &EI) {
      EI.dump(OS);
      return OS;
    }
    /// Debug print.
    void dump(raw_ostream &OS) const {
      OS << "{User:" << (UserTE ? std::to_string(UserTE->Idx) : "null")
         << " EdgeIdx:" << EdgeIdx << "}";
    }
    LLVM_DUMP_METHOD void dump() const { dump(dbgs()); }
#endif
    bool operator == (const EdgeInfo &Other) const {
      return UserTE == Other.UserTE && EdgeIdx == Other.EdgeIdx;
    }
 
    operator bool() const { return UserTE != nullptr; }
  };
  friend struct DenseMapInfo<EdgeInfo>;
 
  /// A helper class used for scoring candidates for two consecutive lanes.
  class LookAheadHeuristics {
    const TargetLibraryInfo &TLI;
    const DataLayout &DL;
    ScalarEvolution &SE;
    const BoUpSLP &R;
    int NumLanes; // Total number of lanes (aka vectorization factor).
    int MaxLevel; // The maximum recursion depth for accumulating score.
 
  public:
    LookAheadHeuristics(const TargetLibraryInfo &TLI, const DataLayout &DL,
                        ScalarEvolution &SE, const BoUpSLP &R, int NumLanes,
                        int MaxLevel)
        : TLI(TLI), DL(DL), SE(SE), R(R), NumLanes(NumLanes),
          MaxLevel(MaxLevel) {}
 
    // The hard-coded scores listed here are not very important, though it shall
    // be higher for better matches to improve the resulting cost. When
    // computing the scores of matching one sub-tree with another, we are
    // basically counting the number of values that are matching. So even if all
    // scores are set to 1, we would still get a decent matching result.
    // However, sometimes we have to break ties. For example we may have to
    // choose between matching loads vs matching opcodes. This is what these
    // scores are helping us with: they provide the order of preference. Also,
    // this is important if the scalar is externally used or used in another
    // tree entry node in the different lane.
 
    /// Loads from consecutive memory addresses, e.g. load(A[i]), load(A[i+1]).
    static constexpr int ScoreConsecutiveLoads = 40;
    /// The same load multiple times. This should have a better score than
    /// `ScoreSplat` because it in x86 for a 2-lane vector we can represent it
    /// with `movddup (%reg), xmm0` which has a throughput of 0.5 versus 0.5 for
    /// a vector load and 1.0 for a broadcast.
    static constexpr int ScoreSplatLoads = 30;
    /// Loads from reversed memory addresses, e.g. load(A[i+1]), load(A[i]).
    static constexpr int ScoreReversedLoads = 30;
    /// A load candidate for masked gather.
    static constexpr int ScoreMaskedGatherCandidate = 10;
    /// ExtractElementInst from same vector and consecutive indexes.
    static constexpr int ScoreConsecutiveExtracts = 40;
    /// ExtractElementInst from same vector and reversed indices.
    static constexpr int ScoreReversedExtracts = 30;
    /// Constants.
    static constexpr int ScoreConstants = 15;
    /// Same constants.
    static constexpr int ScoreSameConstants = 17;
    /// Instructions with the same opcode.
    static constexpr int ScoreSameOpcode = 20;
    /// Instructions with alt opcodes (e.g, add + sub).
    static constexpr int ScoreAltOpcodes = 10;
    /// Identical instructions (a.k.a. splat or broadcast).
    static constexpr int ScoreSplat = 10;
    /// Matching with an undef is preferable to failing.
    static constexpr int ScoreUndef = 10;
    /// Score for failing to find a decent match.
    static constexpr int ScoreFail = 0;
    /// Score if all users are vectorized.
    static constexpr int ScoreAllUserVectorized = 10;
 
    /// \returns the score of placing \p V1 and \p V2 in consecutive lanes.
    /// \p U1 and \p U2 are the users of \p V1 and \p V2.
    /// Also, checks if \p V1 and \p V2 are compatible with instructions in \p
    /// MainAltOps.
    int getShallowScore(Value *V1, Value *V2, Instruction *U1, Instruction *U2,
                        ArrayRef<Value *> MainAltOps) const {
      if (!isValidElementType(V1->getType(), SLPReVec) ||
          !isValidElementType(V2->getType(), SLPReVec))
        return LookAheadHeuristics::ScoreFail;
 
      if (V1 == V2) {
        if (isa<LoadInst>(V1)) {
          // Retruns true if the users of V1 and V2 won't need to be extracted.
          auto AllUsersAreInternal = [U1, U2, this](Value *V1, Value *V2) {
            // Bail out if we have too many uses to save compilation time.
            if (V1->hasNUsesOrMore(UsesLimit) || V2->hasNUsesOrMore(UsesLimit))
              return false;
 
            auto AllUsersVectorized = [U1, U2, this](Value *V) {
              return llvm::all_of(V->users(), [U1, U2, this](Value *U) {
                return U == U1 || U == U2 || R.isVectorized(U);
              });
            };
            return AllUsersVectorized(V1) && AllUsersVectorized(V2);
          };
          // A broadcast of a load can be cheaper on some targets.
          if (R.TTI->isLegalBroadcastLoad(V1->getType(),
                                          ElementCount::getFixed(NumLanes)) &&
              ((int)V1->getNumUses() == NumLanes ||
               AllUsersAreInternal(V1, V2)))
            return LookAheadHeuristics::ScoreSplatLoads;
        }
        if (isa<UndefValue>(V1))
          return LookAheadHeuristics::ScoreUndef;
        if (isConstant(V1))
          return LookAheadHeuristics::ScoreSameConstants;
        return LookAheadHeuristics::ScoreSplat;
      }
 
      auto CheckSameEntryOrFail = [&]() {
        if (ArrayRef<TreeEntry *> TEs1 = R.getTreeEntries(V1); !TEs1.empty()) {
          SmallPtrSet<TreeEntry *, 4> Set(llvm::from_range, TEs1);
          if (ArrayRef<TreeEntry *> TEs2 = R.getTreeEntries(V2);
              !TEs2.empty() &&
              any_of(TEs2, [&](TreeEntry *E) { return Set.contains(E); }))
            return LookAheadHeuristics::ScoreSplatLoads;
        }
        return LookAheadHeuristics::ScoreFail;
      };
 
      auto *LI1 = dyn_cast<LoadInst>(V1);
      auto *LI2 = dyn_cast<LoadInst>(V2);
      if (LI1 && LI2) {
        if (LI1->getParent() != LI2->getParent() || !LI1->isSimple() ||
            !LI2->isSimple())
          return CheckSameEntryOrFail();
 
        std::optional<int64_t> Dist = getPointersDiff(
            LI1->getType(), LI1->getPointerOperand(), LI2->getType(),
            LI2->getPointerOperand(), DL, SE, /*StrictCheck=*/true);
        if (!Dist || *Dist == 0) {
          if (getUnderlyingObject(LI1->getPointerOperand()) ==
                  getUnderlyingObject(LI2->getPointerOperand()) &&
              R.TTI->isLegalMaskedGather(
                  getWidenedType(LI1->getType(), NumLanes), LI1->getAlign()))
            return LookAheadHeuristics::ScoreMaskedGatherCandidate;
          return CheckSameEntryOrFail();
        }
        // The distance is too large - still may be profitable to use masked
        // loads/gathers.
        if (std::abs(*Dist) > NumLanes / 2)
          return LookAheadHeuristics::ScoreMaskedGatherCandidate;
        // This still will detect consecutive loads, but we might have "holes"
        // in some cases. It is ok for non-power-2 vectorization and may produce
        // better results. It should not affect current vectorization.
        return (*Dist > 0) ? LookAheadHeuristics::ScoreConsecutiveLoads
                           : LookAheadHeuristics::ScoreReversedLoads;
      }
 
      auto *C1 = dyn_cast<Constant>(V1);
      auto *C2 = dyn_cast<Constant>(V2);
      if (C1 && C2)
        return LookAheadHeuristics::ScoreConstants;
 
      // Consider constants and buildvector compatible.
      if ((C1 && isa<InsertElementInst>(V2)) ||
          (C2 && isa<InsertElementInst>(V1)))
        return LookAheadHeuristics::ScoreSameOpcode;
 
      // Extracts from consecutive indexes of the same vector better score as
      // the extracts could be optimized away.
      Value *EV1;
      ConstantInt *Ex1Idx;
      if (match(V1, m_ExtractElt(m_Value(EV1), m_ConstantInt(Ex1Idx)))) {
        // Undefs are always profitable for extractelements.
        // Compiler can easily combine poison and extractelement <non-poison> or
        // undef and extractelement <poison>. But combining undef +
        // extractelement <non-poison-but-may-produce-poison> requires some
        // extra operations.
        if (isa<UndefValue>(V2))
          return (isa<PoisonValue>(V2) || isUndefVector(EV1).all())
                     ? LookAheadHeuristics::ScoreConsecutiveExtracts
                     : LookAheadHeuristics::ScoreSameOpcode;
        Value *EV2 = nullptr;
        ConstantInt *Ex2Idx = nullptr;
        if (match(V2,
                  m_ExtractElt(m_Value(EV2), m_CombineOr(m_ConstantInt(Ex2Idx),
                                                         m_Undef())))) {
          // Undefs are always profitable for extractelements.
          if (!Ex2Idx)
            return LookAheadHeuristics::ScoreConsecutiveExtracts;
          if (isUndefVector(EV2).all() && EV2->getType() == EV1->getType())
            return LookAheadHeuristics::ScoreConsecutiveExtracts;
          if (EV2 == EV1) {
            int Idx1 = Ex1Idx->getZExtValue();
            int Idx2 = Ex2Idx->getZExtValue();
            int Dist = Idx2 - Idx1;
            // The distance is too large - still may be profitable to use
            // shuffles.
            if (std::abs(Dist) == 0)
              return LookAheadHeuristics::ScoreSplat;
            if (std::abs(Dist) > NumLanes / 2)
              return LookAheadHeuristics::ScoreSameOpcode;
            return (Dist > 0) ? LookAheadHeuristics::ScoreConsecutiveExtracts
                              : LookAheadHeuristics::ScoreReversedExtracts;
          }
          return LookAheadHeuristics::ScoreAltOpcodes;
        }
        return CheckSameEntryOrFail();
      }
 
      auto *I1 = dyn_cast<Instruction>(V1);
      auto *I2 = dyn_cast<Instruction>(V2);
      if (I1 && I2) {
        if (I1->getParent() != I2->getParent())
          return CheckSameEntryOrFail();
        Value *V;
        Value *Cond;
        // ZExt i1 to something must be considered same opcode for select i1
        // cmp, x, y
        // Required to better match the transformation after
        // BoUpSLP::matchesInversedZExtSelect analysis.
        if ((match(I1, m_ZExt(m_Value(V))) &&
             match(I2, m_Select(m_Value(Cond), m_Value(), m_Value())) &&
             V->getType() == Cond->getType()) ||
            (match(I2, m_ZExt(m_Value(V))) &&
             match(I1, m_Select(m_Value(Cond), m_Value(), m_Value())) &&
             V->getType() == Cond->getType()))
          return LookAheadHeuristics::ScoreSameOpcode;
        SmallVector<Value *, 4> Ops(MainAltOps);
        Ops.push_back(I1);
        Ops.push_back(I2);
        InstructionsState S = getSameOpcode(Ops, TLI);
        // Note: Only consider instructions with <= 2 operands to avoid
        // complexity explosion.
        if (S &&
            (S.getMainOp()->getNumOperands() <= 2 || !MainAltOps.empty() ||
             !S.isAltShuffle()) &&
            all_of(Ops, [&S](Value *V) {
              return isa<PoisonValue>(V) ||
                     cast<Instruction>(V)->getNumOperands() ==
                         S.getMainOp()->getNumOperands();
            }))
          return S.isAltShuffle() ? LookAheadHeuristics::ScoreAltOpcodes
                                  : LookAheadHeuristics::ScoreSameOpcode;
      }
 
      if (I1 && isa<PoisonValue>(V2))
        return LookAheadHeuristics::ScoreSameOpcode;
 
      if (isa<UndefValue>(V2))
        return LookAheadHeuristics::ScoreUndef;
 
      return CheckSameEntryOrFail();
    }
 
    /// Go through the operands of \p LHS and \p RHS recursively until
    /// MaxLevel, and return the cummulative score. \p U1 and \p U2 are
    /// the users of \p LHS and \p RHS (that is \p LHS and \p RHS are operands
    /// of \p U1 and \p U2), except at the beginning of the recursion where
    /// these are set to nullptr.
    ///
    /// For example:
    /// \verbatim
    ///  A[0]  B[0]  A[1]  B[1]  C[0] D[0]  B[1] A[1]
    ///     \ /         \ /         \ /        \ /
    ///      +           +           +          +
    ///     G1          G2          G3         G4
    /// \endverbatim
    /// The getScoreAtLevelRec(G1, G2) function will try to match the nodes at
    /// each level recursively, accumulating the score. It starts from matching
    /// the additions at level 0, then moves on to the loads (level 1). The
    /// score of G1 and G2 is higher than G1 and G3, because {A[0],A[1]} and
    /// {B[0],B[1]} match with LookAheadHeuristics::ScoreConsecutiveLoads, while
    /// {A[0],C[0]} has a score of LookAheadHeuristics::ScoreFail.
    /// Please note that the order of the operands does not matter, as we
    /// evaluate the score of all profitable combinations of operands. In
    /// other words the score of G1 and G4 is the same as G1 and G2. This
    /// heuristic is based on ideas described in:
    ///   Look-ahead SLP: Auto-vectorization in the presence of commutative
    ///   operations, CGO 2018 by Vasileios Porpodas, Rodrigo C. O. Rocha,
    ///   Luís F. W. Góes
    int getScoreAtLevelRec(Value *LHS, Value *RHS, Instruction *U1,
                           Instruction *U2, int CurrLevel,
                           ArrayRef<Value *> MainAltOps) const {
 
      // Get the shallow score of V1 and V2.
      int ShallowScoreAtThisLevel =
          getShallowScore(LHS, RHS, U1, U2, MainAltOps);
 
      // If reached MaxLevel,
      //  or if V1 and V2 are not instructions,
      //  or if they are SPLAT,
      //  or if they are not consecutive,
      //  or if profitable to vectorize loads or extractelements, early return
      //  the current cost.
      auto *I1 = dyn_cast<Instruction>(LHS);
      auto *I2 = dyn_cast<Instruction>(RHS);
      if (CurrLevel == MaxLevel || !(I1 && I2) || I1 == I2 ||
          ShallowScoreAtThisLevel == LookAheadHeuristics::ScoreFail ||
          (((isa<LoadInst>(I1) && isa<LoadInst>(I2)) ||
            (I1->getNumOperands() > 2 && I2->getNumOperands() > 2) ||
            (isa<ExtractElementInst>(I1) && isa<ExtractElementInst>(I2))) &&
           ShallowScoreAtThisLevel))
        return ShallowScoreAtThisLevel;
      assert(I1 && I2 && "Should have early exited.");
 
      // Contains the I2 operand indexes that got matched with I1 operands.
      SmallSet<unsigned, 4> Op2Used;
 
      // Recursion towards the operands of I1 and I2. We are trying all possible
      // operand pairs, and keeping track of the best score.
      if (I1->getNumOperands() != I2->getNumOperands())
        return LookAheadHeuristics::ScoreSameOpcode;
      for (unsigned OpIdx1 = 0, NumOperands1 = I1->getNumOperands();
           OpIdx1 != NumOperands1; ++OpIdx1) {
        // Try to pair op1I with the best operand of I2.
        int MaxTmpScore = 0;
        unsigned MaxOpIdx2 = 0;
        bool FoundBest = false;
        // If I2 is commutative try all combinations.
        unsigned FromIdx = isCommutative(I2) ? 0 : OpIdx1;
        unsigned ToIdx = isCommutative(I2)
                             ? I2->getNumOperands()
                             : std::min(I2->getNumOperands(), OpIdx1 + 1);
        assert(FromIdx <= ToIdx && "Bad index");
        for (unsigned OpIdx2 = FromIdx; OpIdx2 != ToIdx; ++OpIdx2) {
          // Skip operands already paired with OpIdx1.
          if (Op2Used.count(OpIdx2))
            continue;
          // Recursively calculate the cost at each level
          int TmpScore =
              getScoreAtLevelRec(I1->getOperand(OpIdx1), I2->getOperand(OpIdx2),
                                 I1, I2, CurrLevel + 1, {});
          // Look for the best score.
          if (TmpScore > LookAheadHeuristics::ScoreFail &&
              TmpScore > MaxTmpScore) {
            MaxTmpScore = TmpScore;
            MaxOpIdx2 = OpIdx2;
            FoundBest = true;
          }
        }
        if (FoundBest) {
          // Pair {OpIdx1, MaxOpIdx2} was found to be best. Never revisit it.
          Op2Used.insert(MaxOpIdx2);
          ShallowScoreAtThisLevel += MaxTmpScore;
        }
      }
      return ShallowScoreAtThisLevel;
    }
  };
  /// A helper data structure to hold the operands of a vector of instructions.
  /// This supports a fixed vector length for all operand vectors.
  class VLOperands {
    /// For each operand we need (i) the value, and (ii) the opcode that it
    /// would be attached to if the expression was in a left-linearized form.
    /// This is required to avoid illegal operand reordering.
    /// For example:
    /// \verbatim
    ///                         0 Op1
    ///                         |/
    /// Op1 Op2   Linearized    + Op2
    ///   \ /     ---------->   |/
    ///    -                    -
    ///
    /// Op1 - Op2            (0 + Op1) - Op2
    /// \endverbatim
    ///
    /// Value Op1 is attached to a '+' operation, and Op2 to a '-'.
    ///
    /// Another way to think of this is to track all the operations across the
    /// path from the operand all the way to the root of the tree and to
    /// calculate the operation that corresponds to this path. For example, the
    /// path from Op2 to the root crosses the RHS of the '-', therefore the
    /// corresponding operation is a '-' (which matches the one in the
    /// linearized tree, as shown above).
    ///
    /// For lack of a better term, we refer to this operation as Accumulated
    /// Path Operation (APO).
    struct OperandData {
      OperandData() = default;
      OperandData(Value *V, bool APO, bool IsUsed)
          : V(V), APO(APO), IsUsed(IsUsed) {}
      /// The operand value.
      Value *V = nullptr;
      /// TreeEntries only allow a single opcode, or an alternate sequence of
      /// them (e.g, +, -). Therefore, we can safely use a boolean value for the
      /// APO. It is set to 'true' if 'V' is attached to an inverse operation
      /// in the left-linearized form (e.g., Sub/Div), and 'false' otherwise
      /// (e.g., Add/Mul)
      bool APO = false;
      /// Helper data for the reordering function.
      bool IsUsed = false;
    };
 
    /// During operand reordering, we are trying to select the operand at lane
    /// that matches best with the operand at the neighboring lane. Our
    /// selection is based on the type of value we are looking for. For example,
    /// if the neighboring lane has a load, we need to look for a load that is
    /// accessing a consecutive address. These strategies are summarized in the
    /// 'ReorderingMode' enumerator.
    enum class ReorderingMode {
      Load,     ///< Matching loads to consecutive memory addresses
      Opcode,   ///< Matching instructions based on opcode (same or alternate)
      Constant, ///< Matching constants
      Splat,    ///< Matching the same instruction multiple times (broadcast)
      Failed,   ///< We failed to create a vectorizable group
    };
 
    using OperandDataVec = SmallVector<OperandData, 2>;
 
    /// A vector of operand vectors.
    SmallVector<OperandDataVec, 4> OpsVec;
    /// When VL[0] is IntrinsicInst, ArgSize is CallBase::arg_size. When VL[0]
    /// is not IntrinsicInst, ArgSize is User::getNumOperands.
    unsigned ArgSize = 0;
 
    const TargetLibraryInfo &TLI;
    const DataLayout &DL;
    ScalarEvolution &SE;
    const BoUpSLP &R;
    const Loop *L = nullptr;
 
    /// \returns the operand data at \p OpIdx and \p Lane.
    OperandData &getData(unsigned OpIdx, unsigned Lane) {
      return OpsVec[OpIdx][Lane];
    }
 
    /// \returns the operand data at \p OpIdx and \p Lane. Const version.
    const OperandData &getData(unsigned OpIdx, unsigned Lane) const {
      return OpsVec[OpIdx][Lane];
    }
 
    /// Clears the used flag for all entries.
    void clearUsed() {
      for (unsigned OpIdx = 0, NumOperands = getNumOperands();
           OpIdx != NumOperands; ++OpIdx)
        for (unsigned Lane = 0, NumLanes = getNumLanes(); Lane != NumLanes;
             ++Lane)
          OpsVec[OpIdx][Lane].IsUsed = false;
    }
 
    /// Swap the operand at \p OpIdx1 with that one at \p OpIdx2.
    void swap(unsigned OpIdx1, unsigned OpIdx2, unsigned Lane) {
      std::swap(OpsVec[OpIdx1][Lane], OpsVec[OpIdx2][Lane]);
    }
 
    /// \param Lane lane of the operands under analysis.
    /// \param OpIdx operand index in \p Lane lane we're looking the best
    /// candidate for.
    /// \param Idx operand index of the current candidate value.
    /// \returns The additional score due to possible broadcasting of the
    /// elements in the lane. It is more profitable to have power-of-2 unique
    /// elements in the lane, it will be vectorized with higher probability
    /// after removing duplicates. Currently the SLP vectorizer supports only
    /// vectorization of the power-of-2 number of unique scalars.
    int getSplatScore(unsigned Lane, unsigned OpIdx, unsigned Idx,
                      const SmallBitVector &UsedLanes) const {
      Value *IdxLaneV = getData(Idx, Lane).V;
      if (!isa<Instruction>(IdxLaneV) || IdxLaneV == getData(OpIdx, Lane).V ||
          isa<ExtractElementInst>(IdxLaneV))
        return 0;
      SmallDenseMap<Value *, unsigned, 4> Uniques;
      for (unsigned Ln : seq<unsigned>(getNumLanes())) {
        if (Ln == Lane)
          continue;
        Value *OpIdxLnV = getData(OpIdx, Ln).V;
        if (!isa<Instruction>(OpIdxLnV))
          return 0;
        Uniques.try_emplace(OpIdxLnV, Ln);
      }
      unsigned UniquesCount = Uniques.size();
      auto IdxIt = Uniques.find(IdxLaneV);
      unsigned UniquesCntWithIdxLaneV =
          IdxIt != Uniques.end() ? UniquesCount : UniquesCount + 1;
      Value *OpIdxLaneV = getData(OpIdx, Lane).V;
      auto OpIdxIt = Uniques.find(OpIdxLaneV);
      unsigned UniquesCntWithOpIdxLaneV =
          OpIdxIt != Uniques.end() ? UniquesCount : UniquesCount + 1;
      if (UniquesCntWithIdxLaneV == UniquesCntWithOpIdxLaneV)
        return 0;
      return std::min(bit_ceil(UniquesCntWithOpIdxLaneV) -
                          UniquesCntWithOpIdxLaneV,
                      UniquesCntWithOpIdxLaneV -
                          bit_floor(UniquesCntWithOpIdxLaneV)) -
             ((IdxIt != Uniques.end() && UsedLanes.test(IdxIt->second))
                  ? UniquesCntWithIdxLaneV - bit_floor(UniquesCntWithIdxLaneV)
                  : bit_ceil(UniquesCntWithIdxLaneV) - UniquesCntWithIdxLaneV);
    }
 
    /// \param Lane lane of the operands under analysis.
    /// \param OpIdx operand index in \p Lane lane we're looking the best
    /// candidate for.
    /// \param Idx operand index of the current candidate value.
    /// \returns The additional score for the scalar which users are all
    /// vectorized.
    int getExternalUseScore(unsigned Lane, unsigned OpIdx, unsigned Idx) const {
      Value *IdxLaneV = getData(Idx, Lane).V;
      Value *OpIdxLaneV = getData(OpIdx, Lane).V;
      // Do not care about number of uses for vector-like instructions
      // (extractelement/extractvalue with constant indices), they are extracts
      // themselves and already externally used. Vectorization of such
      // instructions does not add extra extractelement instruction, just may
      // remove it.
      if (isVectorLikeInstWithConstOps(IdxLaneV) &&
          isVectorLikeInstWithConstOps(OpIdxLaneV))
        return LookAheadHeuristics::ScoreAllUserVectorized;
      auto *IdxLaneI = dyn_cast<Instruction>(IdxLaneV);
      if (!IdxLaneI || !isa<Instruction>(OpIdxLaneV))
        return 0;
      return R.areAllUsersVectorized(IdxLaneI)
                 ? LookAheadHeuristics::ScoreAllUserVectorized
                 : 0;
    }
 
    /// Score scaling factor for fully compatible instructions but with
    /// different number of external uses. Allows better selection of the
    /// instructions with less external uses.
    static constexpr int ScoreScaleFactor = 10;
    /// Scale factor for constants only.
    static constexpr int ScoreConstantScaleFactor = 6;
 
    /// \Returns the look-ahead score, which tells us how much the sub-trees
    /// rooted at \p LHS and \p RHS match, the more they match the higher the
    /// score. This helps break ties in an informed way when we cannot decide on
    /// the order of the operands by just considering the immediate
    /// predecessors.
    int getLookAheadScore(Value *LHS, Value *RHS, ArrayRef<Value *> MainAltOps,
                          int Lane, unsigned OpIdx, unsigned Idx,
                          bool &IsUsed, const SmallBitVector &UsedLanes) {
      LookAheadHeuristics LookAhead(TLI, DL, SE, R, getNumLanes(),
                                    LookAheadMaxDepth);
      // Keep track of the instruction stack as we recurse into the operands
      // during the look-ahead score exploration.
      int Score =
          LookAhead.getScoreAtLevelRec(LHS, RHS, /*U1=*/nullptr, /*U2=*/nullptr,
                                       /*CurrLevel=*/1, MainAltOps);
      if (Score) {
        int SplatScore =
            getSplatScore(Lane, OpIdx, Idx, UsedLanes) * ScoreScaleFactor;
        if (Score <= -SplatScore) {
          // Failed score.
          Score = 0;
        } else {
          Score += SplatScore;
          // Scale score to see the difference between different operands
          // and similar operands but all vectorized/not all vectorized
          // uses. It does not affect actual selection of the best
          // compatible operand in general, just allows to select the
          // operand with all vectorized uses.
          const int SF = (LHS == RHS && isConstant(LHS))
                             ? ScoreConstantScaleFactor
                             : ScoreScaleFactor;
          Score *= SF;
          Score += getExternalUseScore(Lane, OpIdx, Idx);
          IsUsed = true;
        }
      }
      return Score;
    }
 
    /// Best defined scores per lanes between the passes. Used to choose the
    /// best operand (with the highest score) between the passes.
    /// The key - {Operand Index, Lane}.
    /// The value - the best score between the passes for the lane and the
    /// operand.
    SmallDenseMap<std::pair<unsigned, unsigned>, unsigned, 8>
        BestScoresPerLanes;
 
    // Search all operands in Ops[*][Lane] for the one that matches best
    // Ops[OpIdx][LastLane] and return its opreand index.
    // If no good match can be found, return std::nullopt.
    std::optional<unsigned>
    getBestOperand(unsigned OpIdx, int Lane, int LastLane,
                   ArrayRef<ReorderingMode> ReorderingModes,
                   ArrayRef<Value *> MainAltOps,
                   const SmallBitVector &UsedLanes) {
      unsigned NumOperands = getNumOperands();
 
      // The operand of the previous lane at OpIdx.
      Value *OpLastLane = getData(OpIdx, LastLane).V;
 
      // Our strategy mode for OpIdx.
      ReorderingMode RMode = ReorderingModes[OpIdx];
      if (RMode == ReorderingMode::Failed)
        return std::nullopt;
 
      // The linearized opcode of the operand at OpIdx, Lane.
      bool OpIdxAPO = getData(OpIdx, Lane).APO;
 
      // The best operand index and its score.
      // Sometimes we have more than one option (e.g., Opcode and Undefs), so we
      // are using the score to differentiate between the two.
      struct BestOpData {
        std::optional<unsigned> Idx;
        unsigned Score = 0;
      } BestOp;
      BestOp.Score =
          BestScoresPerLanes.try_emplace(std::make_pair(OpIdx, Lane), 0)
              .first->second;
 
      // Track if the operand must be marked as used. If the operand is set to
      // Score 1 explicitly (because of non power-of-2 unique scalars, we may
      // want to reestimate the operands again on the following iterations).
      bool IsUsed = RMode == ReorderingMode::Splat ||
                    RMode == ReorderingMode::Constant ||
                    RMode == ReorderingMode::Load;
      // Iterate through all unused operands and look for the best.
      for (unsigned Idx = 0; Idx != NumOperands; ++Idx) {
        // Get the operand at Idx and Lane.
        OperandData &OpData = getData(Idx, Lane);
        Value *Op = OpData.V;
        bool OpAPO = OpData.APO;
 
        // Skip already selected operands.
        if (OpData.IsUsed)
          continue;
 
        // Skip if we are trying to move the operand to a position with a
        // different opcode in the linearized tree form. This would break the
        // semantics.
        if (OpAPO != OpIdxAPO)
          continue;
 
        // Look for an operand that matches the current mode.
        switch (RMode) {
        case ReorderingMode::Load:
        case ReorderingMode::Opcode: {
          bool LeftToRight = Lane > LastLane;
          Value *OpLeft = (LeftToRight) ? OpLastLane : Op;
          Value *OpRight = (LeftToRight) ? Op : OpLastLane;
          int Score = getLookAheadScore(OpLeft, OpRight, MainAltOps, Lane,
                                        OpIdx, Idx, IsUsed, UsedLanes);
          if (Score > static_cast<int>(BestOp.Score) ||
              (Score > 0 && Score == static_cast<int>(BestOp.Score) &&
               Idx == OpIdx)) {
            BestOp.Idx = Idx;
            BestOp.Score = Score;
            BestScoresPerLanes[std::make_pair(OpIdx, Lane)] = Score;
          }
          break;
        }
        case ReorderingMode::Constant:
          if (isa<Constant>(Op) ||
              (!BestOp.Score && L && L->isLoopInvariant(Op))) {
            BestOp.Idx = Idx;
            if (isa<Constant>(Op)) {
              BestOp.Score = LookAheadHeuristics::ScoreConstants;
              BestScoresPerLanes[std::make_pair(OpIdx, Lane)] =
                  LookAheadHeuristics::ScoreConstants;
            }
            if (isa<UndefValue>(Op) || !isa<Constant>(Op))
              IsUsed = false;
          }
          break;
        case ReorderingMode::Splat:
          if (Op == OpLastLane || (!BestOp.Score && isa<Constant>(Op))) {
            IsUsed = Op == OpLastLane;
            if (Op == OpLastLane) {
              BestOp.Score = LookAheadHeuristics::ScoreSplat;
              BestScoresPerLanes[std::make_pair(OpIdx, Lane)] =
                  LookAheadHeuristics::ScoreSplat;
            }
            BestOp.Idx = Idx;
          }
          break;
        case ReorderingMode::Failed:
          llvm_unreachable("Not expected Failed reordering mode.");
        }
      }
 
      if (BestOp.Idx) {
        getData(*BestOp.Idx, Lane).IsUsed = IsUsed;
        return BestOp.Idx;
      }
      // If we could not find a good match return std::nullopt.
      return std::nullopt;
    }
 
    /// Helper for reorderOperandVecs.
    /// \returns the lane that we should start reordering from. This is the one
    /// which has the least number of operands that can freely move about or
    /// less profitable because it already has the most optimal set of operands.
    unsigned getBestLaneToStartReordering() const {
      unsigned Min = UINT_MAX;
      unsigned SameOpNumber = 0;
      // std::pair<unsigned, unsigned> is used to implement a simple voting
      // algorithm and choose the lane with the least number of operands that
      // can freely move about or less profitable because it already has the
      // most optimal set of operands. The first unsigned is a counter for
      // voting, the second unsigned is the counter of lanes with instructions
      // with same/alternate opcodes and same parent basic block.
      MapVector<unsigned, std::pair<unsigned, unsigned>> HashMap;
      // Try to be closer to the original results, if we have multiple lanes
      // with same cost. If 2 lanes have the same cost, use the one with the
      // highest index.
      for (int I = getNumLanes(); I > 0; --I) {
        unsigned Lane = I - 1;
        OperandsOrderData NumFreeOpsHash =
            getMaxNumOperandsThatCanBeReordered(Lane);
        // Compare the number of operands that can move and choose the one with
        // the least number.
        if (NumFreeOpsHash.NumOfAPOs < Min) {
          Min = NumFreeOpsHash.NumOfAPOs;
          SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent;
          HashMap.clear();
          HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane);
        } else if (NumFreeOpsHash.NumOfAPOs == Min &&
                   NumFreeOpsHash.NumOpsWithSameOpcodeParent < SameOpNumber) {
          // Select the most optimal lane in terms of number of operands that
          // should be moved around.
          SameOpNumber = NumFreeOpsHash.NumOpsWithSameOpcodeParent;
          HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane);
        } else if (NumFreeOpsHash.NumOfAPOs == Min &&
                   NumFreeOpsHash.NumOpsWithSameOpcodeParent == SameOpNumber) {
          auto [It, Inserted] =
              HashMap.try_emplace(NumFreeOpsHash.Hash, 1, Lane);
          if (!Inserted)
            ++It->second.first;
        }
      }
      // Select the lane with the minimum counter.
      unsigned BestLane = 0;
      unsigned CntMin = UINT_MAX;
      for (const auto &Data : reverse(HashMap)) {
        if (Data.second.first < CntMin) {
          CntMin = Data.second.first;
          BestLane = Data.second.second;
        }
      }
      return BestLane;
    }
 
    /// Data structure that helps to reorder operands.
    struct OperandsOrderData {
      /// The best number of operands with the same APOs, which can be
      /// reordered.
      unsigned NumOfAPOs = UINT_MAX;
      /// Number of operands with the same/alternate instruction opcode and
      /// parent.
      unsigned NumOpsWithSameOpcodeParent = 0;
      /// Hash for the actual operands ordering.
      /// Used to count operands, actually their position id and opcode
      /// value. It is used in the voting mechanism to find the lane with the
      /// least number of operands that can freely move about or less profitable
      /// because it already has the most optimal set of operands. Can be
      /// replaced with SmallVector<unsigned> instead but hash code is faster
      /// and requires less memory.
      unsigned Hash = 0;
    };
    /// \returns the maximum number of operands that are allowed to be reordered
    /// for \p Lane and the number of compatible instructions(with the same
    /// parent/opcode). This is used as a heuristic for selecting the first lane
    /// to start operand reordering.
    OperandsOrderData getMaxNumOperandsThatCanBeReordered(unsigned Lane) const {
      unsigned CntTrue = 0;
      unsigned NumOperands = getNumOperands();
      // Operands with the same APO can be reordered. We therefore need to count
      // how many of them we have for each APO, like this: Cnt[APO] = x.
      // Since we only have two APOs, namely true and false, we can avoid using
      // a map. Instead we can simply count the number of operands that
      // correspond to one of them (in this case the 'true' APO), and calculate
      // the other by subtracting it from the total number of operands.
      // Operands with the same instruction opcode and parent are more
      // profitable since we don't need to move them in many cases, with a high
      // probability such lane already can be vectorized effectively.
      bool AllUndefs = true;
      unsigned NumOpsWithSameOpcodeParent = 0;
      Instruction *OpcodeI = nullptr;
      BasicBlock *Parent = nullptr;
      unsigned Hash = 0;
      for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) {
        const OperandData &OpData = getData(OpIdx, Lane);
        if (OpData.APO)
          ++CntTrue;
        // Use Boyer-Moore majority voting for finding the majority opcode and
        // the number of times it occurs.
        if (auto *I = dyn_cast<Instruction>(OpData.V)) {
          if (!OpcodeI || !getSameOpcode({OpcodeI, I}, TLI) ||
              I->getParent() != Parent) {
            if (NumOpsWithSameOpcodeParent == 0) {
              NumOpsWithSameOpcodeParent = 1;
              OpcodeI = I;
              Parent = I->getParent();
            } else {
              --NumOpsWithSameOpcodeParent;
            }
          } else {
            ++NumOpsWithSameOpcodeParent;
          }
        }
        Hash = hash_combine(
            Hash, hash_value((OpIdx + 1) * (OpData.V->getValueID() + 1)));
        AllUndefs = AllUndefs && isa<UndefValue>(OpData.V);
      }
      if (AllUndefs)
        return {};
      OperandsOrderData Data;
      Data.NumOfAPOs = std::max(CntTrue, NumOperands - CntTrue);
      Data.NumOpsWithSameOpcodeParent = NumOpsWithSameOpcodeParent;
      Data.Hash = Hash;
      return Data;
    }
 
    /// Go through the instructions in VL and append their operands.
    void appendOperands(ArrayRef<Value *> VL, ArrayRef<ValueList> Operands,
                        const InstructionsState &S) {
      assert(!Operands.empty() && !VL.empty() && "Bad list of operands");
      assert((empty() || all_of(Operands,
                                [this](const ValueList &VL) {
                                  return VL.size() == getNumLanes();
                                })) &&
             "Expected same number of lanes");
      assert(S.valid() && "InstructionsState is invalid.");
      // IntrinsicInst::isCommutative returns true if swapping the first "two"
      // arguments to the intrinsic produces the same result.
      Instruction *MainOp = S.getMainOp();
      ArgSize = getNumberOfPotentiallyCommutativeOps(MainOp);
      OpsVec.resize(ArgSize);
      unsigned NumLanes = VL.size();
      for (OperandDataVec &Ops : OpsVec)
        Ops.resize(NumLanes);
      for (unsigned Lane : seq<unsigned>(NumLanes)) {
        // Our tree has just 3 nodes: the root and two operands.
        // It is therefore trivial to get the APO. We only need to check the
        // opcode of V and whether the operand at OpIdx is the LHS or RHS
        // operand. The LHS operand of both add and sub is never attached to an
        // inversese operation in the linearized form, therefore its APO is
        // false. The RHS is true only if V is an inverse operation.
 
        // Since operand reordering is performed on groups of commutative
        // operations or alternating sequences (e.g., +, -), we can safely tell
        // the inverse operations by checking commutativity.
        auto *I = dyn_cast<Instruction>(VL[Lane]);
        if (!I && isa<PoisonValue>(VL[Lane])) {
          for (unsigned OpIdx : seq<unsigned>(ArgSize))
            OpsVec[OpIdx][Lane] = {Operands[OpIdx][Lane], true, false};
          continue;
        }
        bool IsInverseOperation = false;
        if (S.isCopyableElement(VL[Lane])) {
          // The value is a copyable element.
          IsInverseOperation =
              !isCommutative(MainOp, VL[Lane], /*IsCopyable=*/true);
        } else {
          assert(I && "Expected instruction");
          auto [SelectedOp, Ops] = convertTo(I, S);
          // We cannot check commutativity by the converted instruction
          // (SelectedOp) because isCommutative also examines def-use
          // relationships.
          IsInverseOperation = !isCommutative(SelectedOp, I);
        }
        for (unsigned OpIdx : seq<unsigned>(ArgSize)) {
          bool APO = (OpIdx == 0) ? false : IsInverseOperation;
          OpsVec[OpIdx][Lane] = {Operands[OpIdx][Lane], APO, false};
        }
      }
    }
 
    /// \returns the number of operands.
    unsigned getNumOperands() const { return ArgSize; }
 
    /// \returns the number of lanes.
    unsigned getNumLanes() const { return OpsVec[0].size(); }
 
    /// \returns the operand value at \p OpIdx and \p Lane.
    Value *getValue(unsigned OpIdx, unsigned Lane) const {
      return getData(OpIdx, Lane).V;
    }
 
    /// \returns true if the data structure is empty.
    bool empty() const { return OpsVec.empty(); }
 
    /// Clears the data.
    void clear() { OpsVec.clear(); }
 
    /// \Returns true if there are enough operands identical to \p Op to fill
    /// the whole vector (it is mixed with constants or loop invariant values).
    /// Note: This modifies the 'IsUsed' flag, so a cleanUsed() must follow.
    bool shouldBroadcast(Value *Op, unsigned OpIdx, unsigned Lane) {
      assert(Op == getValue(OpIdx, Lane) &&
             "Op is expected to be getValue(OpIdx, Lane).");
      // Small number of loads - try load matching.
      if (isa<LoadInst>(Op) && getNumLanes() == 2 && getNumOperands() == 2)
        return false;
      bool OpAPO = getData(OpIdx, Lane).APO;
      bool IsInvariant = L && L->isLoopInvariant(Op);
      unsigned Cnt = 0;
      for (unsigned Ln = 0, Lns = getNumLanes(); Ln != Lns; ++Ln) {
        if (Ln == Lane)
          continue;
        // This is set to true if we found a candidate for broadcast at Lane.
        bool FoundCandidate = false;
        for (unsigned OpI = 0, OpE = getNumOperands(); OpI != OpE; ++OpI) {
          OperandData &Data = getData(OpI, Ln);
          if (Data.APO != OpAPO || Data.IsUsed)
            continue;
          Value *OpILane = getValue(OpI, Lane);
          bool IsConstantOp = isa<Constant>(OpILane);
          // Consider the broadcast candidate if:
          // 1. Same value is found in one of the operands.
          if (Data.V == Op ||
              // 2. The operand in the given lane is not constant but there is a
              // constant operand in another lane (which can be moved to the
              // given lane). In this case we can represent it as a simple
              // permutation of constant and broadcast.
              (!IsConstantOp &&
               ((Lns > 2 && isa<Constant>(Data.V)) ||
                // 2.1. If we have only 2 lanes, need to check that value in the
                // next lane does not build same opcode sequence.
                (Lns == 2 &&
                 !getSameOpcode({Op, getValue((OpI + 1) % OpE, Ln)}, TLI) &&
                 isa<Constant>(Data.V)))) ||
              // 3. The operand in the current lane is loop invariant (can be
              // hoisted out) and another operand is also a loop invariant
              // (though not a constant). In this case the whole vector can be
              // hoisted out.
              // FIXME: need to teach the cost model about this case for better
              // estimation.
              (IsInvariant && !isa<Constant>(Data.V) &&
               !getSameOpcode({Op, Data.V}, TLI) &&
               L->isLoopInvariant(Data.V))) {
            FoundCandidate = true;
            Data.IsUsed = Data.V == Op;
            if (Data.V == Op)
              ++Cnt;
            break;
          }
        }
        if (!FoundCandidate)
          return false;
      }
      return getNumLanes() == 2 || Cnt > 1;
    }
 
    /// Checks if there is at least single compatible operand in lanes other
    /// than \p Lane, compatible with the operand \p Op.
    bool canBeVectorized(Instruction *Op, unsigned OpIdx, unsigned Lane) const {
      assert(Op == getValue(OpIdx, Lane) &&
             "Op is expected to be getValue(OpIdx, Lane).");
      bool OpAPO = getData(OpIdx, Lane).APO;
      for (unsigned Ln = 0, Lns = getNumLanes(); Ln != Lns; ++Ln) {
        if (Ln == Lane)
          continue;
        if (any_of(seq<unsigned>(getNumOperands()), [&](unsigned OpI) {
              const OperandData &Data = getData(OpI, Ln);
              if (Data.APO != OpAPO || Data.IsUsed)
                return true;
              Value *OpILn = getValue(OpI, Ln);
              return (L && L->isLoopInvariant(OpILn)) ||
                     (getSameOpcode({Op, OpILn}, TLI) &&
                      allSameBlock({Op, OpILn}));
            }))
          return true;
      }
      return false;
    }
 
  public:
    /// Initialize with all the operands of the instruction vector \p RootVL.
    VLOperands(ArrayRef<Value *> RootVL, ArrayRef<ValueList> Operands,
               const InstructionsState &S, const BoUpSLP &R)
        : TLI(*R.TLI), DL(*R.DL), SE(*R.SE), R(R),
          L(R.LI->getLoopFor(S.getMainOp()->getParent())) {
      // Append all the operands of RootVL.
      appendOperands(RootVL, Operands, S);
    }
 
    /// Initialize with flattened operand columns of an associative node.
    /// ArgSize is taken from \p Operands, APO is always false.
    VLOperands(ArrayRef<ValueList> Operands, const BasicBlock *BB,
               const BoUpSLP &R)
        : TLI(*R.TLI), DL(*R.DL), SE(*R.SE), R(R), L(R.LI->getLoopFor(BB)) {
      assert(!Operands.empty() && "Expected at least one operand column");
      ArgSize = Operands.size();
      OpsVec.resize(ArgSize);
      unsigned NumLanes = Operands.front().size();
      for (auto [OpIdx, Ops] : enumerate(OpsVec)) {
        Ops.resize(NumLanes);
        for (unsigned Lane : seq<unsigned>(NumLanes))
          Ops[Lane] = OperandData(Operands[OpIdx][Lane], /*APO=*/false,
                                  /*IsUsed=*/false);
      }
    }
 
    /// \Returns a value vector with the operands across all lanes for the
    /// opearnd at \p OpIdx.
    ValueList getVL(unsigned OpIdx) const {
      ValueList OpVL(OpsVec[OpIdx].size());
      assert(OpsVec[OpIdx].size() == getNumLanes() &&
             "Expected same num of lanes across all operands");
      for (unsigned Lane = 0, Lanes = getNumLanes(); Lane != Lanes; ++Lane)
        OpVL[Lane] = OpsVec[OpIdx][Lane].V;
      return OpVL;
    }
 
    // Performs operand reordering for 2 or more operands.
    // The original operands are in OrigOps[OpIdx][Lane].
    // The reordered operands are returned in 'SortedOps[OpIdx][Lane]'.
    void reorder() {
      unsigned NumOperands = getNumOperands();
      unsigned NumLanes = getNumLanes();
      // Each operand has its own mode. We are using this mode to help us select
      // the instructions for each lane, so that they match best with the ones
      // we have selected so far.
      SmallVector<ReorderingMode, 2> ReorderingModes(NumOperands);
 
      // This is a greedy single-pass algorithm. We are going over each lane
      // once and deciding on the best order right away with no back-tracking.
      // However, in order to increase its effectiveness, we start with the lane
      // that has operands that can move the least. For example, given the
      // following lanes:
      //  Lane 0 : A[0] = B[0] + C[0]   // Visited 3rd
      //  Lane 1 : A[1] = C[1] - B[1]   // Visited 1st
      //  Lane 2 : A[2] = B[2] + C[2]   // Visited 2nd
      //  Lane 3 : A[3] = C[3] - B[3]   // Visited 4th
      // we will start at Lane 1, since the operands of the subtraction cannot
      // be reordered. Then we will visit the rest of the lanes in a circular
      // fashion. That is, Lanes 2, then Lane 0, and finally Lane 3.
 
      // Find the first lane that we will start our search from.
      unsigned FirstLane = getBestLaneToStartReordering();
 
      // Initialize the modes.
      for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) {
        Value *OpLane0 = getValue(OpIdx, FirstLane);
        // Keep track if we have instructions with all the same opcode on one
        // side.
        if (auto *OpILane0 = dyn_cast<Instruction>(OpLane0)) {
          // Check if OpLane0 should be broadcast.
          if (shouldBroadcast(OpLane0, OpIdx, FirstLane) ||
              !canBeVectorized(OpILane0, OpIdx, FirstLane))
            ReorderingModes[OpIdx] = ReorderingMode::Splat;
          else if (isa<LoadInst>(OpILane0))
            ReorderingModes[OpIdx] = ReorderingMode::Load;
          else
            ReorderingModes[OpIdx] = ReorderingMode::Opcode;
        } else if (isa<Constant>(OpLane0)) {
          ReorderingModes[OpIdx] = ReorderingMode::Constant;
        } else if (isa<Argument>(OpLane0)) {
          // Our best hope is a Splat. It may save some cost in some cases.
          ReorderingModes[OpIdx] = ReorderingMode::Splat;
        } else {
          llvm_unreachable("Unexpected value kind.");
        }
      }
 
      // Check that we don't have same operands. No need to reorder if operands
      // are just perfect diamond or shuffled diamond match. Do not do it only
      // for possible broadcasts.
      auto &&SkipReordering = [this]() {
        SmallPtrSet<Value *, 4> UniqueValues;
        ArrayRef<OperandData> Op0 = OpsVec.front();
        for (const OperandData &Data : Op0)
          UniqueValues.insert(Data.V);
        for (ArrayRef<OperandData> Op :
             ArrayRef(OpsVec).slice(1, getNumOperands() - 1)) {
          if (any_of(Op, [&UniqueValues](const OperandData &Data) {
                return !UniqueValues.contains(Data.V);
              }))
            return false;
        }
        return UniqueValues.size() != 2;
      };
 
      // If the initial strategy fails for any of the operand indexes, then we
      // perform reordering again in a second pass. This helps avoid assigning
      // high priority to the failed strategy, and should improve reordering for
      // the non-failed operand indexes.
      for (int Pass = 0; Pass != 2; ++Pass) {
        // Check if no need to reorder operands since they're are perfect or
        // shuffled diamond match.
        // Need to do it to avoid extra external use cost counting for
        // shuffled matches, which may cause regressions.
        if (SkipReordering())
          break;
        // Skip the second pass if the first pass did not fail.
        bool StrategyFailed = false;
        // Mark all operand data as free to use.
        clearUsed();
        // We keep the original operand order for the FirstLane, so reorder the
        // rest of the lanes. We are visiting the nodes in a circular fashion,
        // using FirstLane as the center point and increasing the radius
        // distance.
        SmallVector<SmallVector<Value *, 2>> MainAltOps(NumOperands);
        for (unsigned I = 0; I < NumOperands; ++I)
          MainAltOps[I].push_back(getData(I, FirstLane).V);
 
        SmallBitVector UsedLanes(NumLanes);
        UsedLanes.set(FirstLane);
        for (unsigned Distance = 1; Distance != NumLanes; ++Distance) {
          // Visit the lane on the right and then the lane on the left.
          for (int Direction : {+1, -1}) {
            int Lane = FirstLane + Direction * Distance;
            if (Lane < 0 || Lane >= (int)NumLanes)
              continue;
            UsedLanes.set(Lane);
            int LastLane = Lane - Direction;
            assert(LastLane >= 0 && LastLane < (int)NumLanes &&
                   "Out of bounds");
            // Look for a good match for each operand.
            for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) {
              // Search for the operand that matches SortedOps[OpIdx][Lane-1].
              std::optional<unsigned> BestIdx =
                  getBestOperand(OpIdx, Lane, LastLane, ReorderingModes,
                                 MainAltOps[OpIdx], UsedLanes);
              // By not selecting a value, we allow the operands that follow to
              // select a better matching value. We will get a non-null value in
              // the next run of getBestOperand().
              if (BestIdx) {
                // Swap the current operand with the one returned by
                // getBestOperand().
                swap(OpIdx, *BestIdx, Lane);
              } else {
                // Enable the second pass.
                StrategyFailed = true;
              }
              // Try to get the alternate opcode and follow it during analysis.
              if (MainAltOps[OpIdx].size() != 2) {
                OperandData &AltOp = getData(OpIdx, Lane);
                InstructionsState OpS =
                    getSameOpcode({MainAltOps[OpIdx].front(), AltOp.V}, TLI);
                if (OpS && OpS.isAltShuffle())
                  MainAltOps[OpIdx].push_back(AltOp.V);
              }
            }
          }
        }
        // Skip second pass if the strategy did not fail.
        if (!StrategyFailed)
          break;
      }
    }
 
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    LLVM_DUMP_METHOD static StringRef getModeStr(ReorderingMode RMode) {
      switch (RMode) {
      case ReorderingMode::Load:
        return "Load";
      case ReorderingMode::Opcode:
        return "Opcode";
      case ReorderingMode::Constant:
        return "Constant";
      case ReorderingMode::Splat:
        return "Splat";
      case ReorderingMode::Failed:
        return "Failed";
      }
      llvm_unreachable("Unimplemented Reordering Type");
    }
 
    LLVM_DUMP_METHOD static raw_ostream &printMode(ReorderingMode RMode,
                                                   raw_ostream &OS) {
      return OS << getModeStr(RMode);
    }
 
    /// Debug print.
    LLVM_DUMP_METHOD static void dumpMode(ReorderingMode RMode) {
      printMode(RMode, dbgs());
    }
 
    friend raw_ostream &operator<<(raw_ostream &OS, ReorderingMode RMode) {
      return printMode(RMode, OS);
    }
 
    LLVM_DUMP_METHOD raw_ostream &print(raw_ostream &OS) const {
      const unsigned Indent = 2;
      unsigned Cnt = 0;
      for (const OperandDataVec &OpDataVec : OpsVec) {
        OS << "Operand " << Cnt++ << "\n";
        for (const OperandData &OpData : OpDataVec) {
          OS.indent(Indent) << "{";
          if (Value *V = OpData.V)
            OS << *V;
          else
            OS << "null";
          OS << ", APO:" << OpData.APO << "}\n";
        }
        OS << "\n";
      }
      return OS;
    }
 
    /// Debug print.
    LLVM_DUMP_METHOD void dump() const { print(dbgs()); }
#endif
  };
 
  /// Evaluate each pair in \p Candidates and return index into \p Candidates
  /// for a pair which have highest score deemed to have best chance to form
  /// root of profitable tree to vectorize. Return std::nullopt if no candidate
  /// scored above the LookAheadHeuristics::ScoreFail. \param Limit Lower limit
  /// of the cost, considered to be good enough score.
  std::pair<std::optional<int>, int>
  findBestRootPair(ArrayRef<std::pair<Value *, Value *>> Candidates,
                   int Limit = LookAheadHeuristics::ScoreFail) const {
    LookAheadHeuristics LookAhead(*TLI, *DL, *SE, *this, /*NumLanes=*/2,
                                  RootLookAheadMaxDepth);
    int BestScore = Limit;
    std::optional<int> Index;
    for (int I : seq<int>(0, Candidates.size())) {
      int Score = LookAhead.getScoreAtLevelRec(Candidates[I].first,
                                               Candidates[I].second,
                                               /*U1=*/nullptr, /*U2=*/nullptr,
                                               /*CurrLevel=*/1, {});
      if (Score > BestScore) {
        BestScore = Score;
        Index = I;
      }
    }
    return std::make_pair(Index, BestScore);
  }
 
  /// Checks if the instruction is marked for deletion.
  bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); }
 
  /// Checks if the value is used only by the assume-like intrinsics.
  bool isEphemeralValue(const Value *V) const { return EphValues.contains(V); }
 
  /// Removes an instruction from its block and eventually deletes it.
  /// It's like Instruction::eraseFromParent() except that the actual deletion
  /// is delayed until BoUpSLP is destructed.
  void eraseInstruction(Instruction *I) {
    DeletedInstructions.insert(I);
  }
 
  /// Remove instructions from the parent function and clear the operands of \p
  /// DeadVals instructions, marking for deletion trivially dead operands.
  template <typename T>
  void removeInstructionsAndOperands(
      ArrayRef<T *> DeadVals,
      ArrayRef<std::tuple<WeakTrackingVH, unsigned, bool, bool>>
          VectorValuesAndScales) {
    SmallVector<WeakTrackingVH> DeadInsts;
    for (T *V : DeadVals) {
      auto *I = cast<Instruction>(V);
      eraseInstruction(I);
    }
    DenseSet<Value *> Processed;
    for (T *V : DeadVals) {
      if (!V || !Processed.insert(V).second)
        continue;
      auto *I = cast<Instruction>(V);
      salvageDebugInfo(*I);
      ArrayRef<TreeEntry *> Entries = getTreeEntries(I);
      for (Use &U : I->operands()) {
        if (auto *OpI = dyn_cast_if_present<Instruction>(U.get());
            OpI && !DeletedInstructions.contains(OpI) && OpI->hasOneUser() &&
            wouldInstructionBeTriviallyDead(OpI, TLI) &&
            !ExternalUseReplacements.contains(OpI) &&
            (Entries.empty() || none_of(Entries, [&](const TreeEntry *Entry) {
               return Entry->VectorizedValue == OpI;
             })))
          DeadInsts.push_back(OpI);
      }
      I->dropAllReferences();
    }
    for (T *V : DeadVals) {
      auto *I = cast<Instruction>(V);
      if (!I->getParent())
        continue;
      assert((I->use_empty() || all_of(I->uses(),
                                       [&](Use &U) {
                                         return isDeleted(
                                             cast<Instruction>(U.getUser()));
                                       })) &&
             "trying to erase instruction with users.");
      I->removeFromParent();
      SE->forgetValue(I);
    }
    // Process the dead instruction list until empty.
    while (!DeadInsts.empty()) {
      Value *V = DeadInsts.pop_back_val();
      Instruction *VI = cast_or_null<Instruction>(V);
      if (!VI || !VI->getParent())
        continue;
      assert(isInstructionTriviallyDead(VI, TLI) &&
             "Live instruction found in dead worklist!");
      assert(VI->use_empty() && "Instructions with uses are not dead.");
 
      // Don't lose the debug info while deleting the instructions.
      salvageDebugInfo(*VI);
 
      // Null out all of the instruction's operands to see if any operand
      // becomes dead as we go.
      for (Use &OpU : VI->operands()) {
        Value *OpV = OpU.get();
        if (!OpV)
          continue;
        OpU.set(nullptr);
 
        if (!OpV->use_empty())
          continue;
 
        // If the operand is an instruction that became dead as we nulled out
        // the operand, and if it is 'trivially' dead, delete it in a future
        // loop iteration.
        if (auto *OpI = dyn_cast<Instruction>(OpV))
          if (!DeletedInstructions.contains(OpI) &&
              !ExternalUseReplacements.contains(OpI) &&
              (!OpI->getType()->isVectorTy() ||
               none_of(
                   VectorValuesAndScales,
                   [&](const std::tuple<WeakTrackingVH, unsigned, bool, bool>
                           &V) { return std::get<0>(V) == OpI; })) &&
              isInstructionTriviallyDead(OpI, TLI))
            DeadInsts.push_back(OpI);
      }
 
      VI->removeFromParent();
      eraseInstruction(VI);
      SE->forgetValue(VI);
    }
  }
 
  /// Checks if the instruction was already analyzed for being possible
  /// reduction root.
  bool isAnalyzedReductionRoot(Instruction *I) const {
    return AnalyzedReductionsRoots.count(I);
  }
  /// Register given instruction as already analyzed for being possible
  /// reduction root.
  void analyzedReductionRoot(Instruction *I) {
    AnalyzedReductionsRoots.insert(I);
  }
  /// Checks if the provided list of reduced values was checked already for
  /// vectorization.
  bool areAnalyzedReductionVals(ArrayRef<Value *> VL) const {
    return AnalyzedReductionVals.contains(hash_value(VL));
  }
  /// Adds the list of reduced values to list of already checked values for the
  /// vectorization.
  void analyzedReductionVals(ArrayRef<Value *> VL) {
    AnalyzedReductionVals.insert(hash_value(VL));
  }
  /// Checks if the value was already a part of the analyzed vector node.
  bool isAnalyzedScalar(const Value *V) const {
    return AnalyzedScalars.contains(V);
  }
  /// Checks if the given bundle was already rejected as non-vectorizable.
  bool isAnalyzedBundle(ArrayRef<Value *> VL) const {
    return AnalyzedBundles.contains(hash_value(VL));
  }
  /// Registers the bundle as rejected for the vectorization.
  void analyzedBundle(ArrayRef<Value *> VL) {
    AnalyzedBundles.insert(hash_value(VL));
  }
  /// Clear the list of the analyzed reduction root instructions.
  void clearReductionData() {
    AnalyzedReductionsRoots.clear();
    AnalyzedReductionVals.clear();
    AnalyzedBundles.clear();
    AnalyzedMinBWVals.clear();
  }
  /// Checks if the given value is gathered in one of the nodes.
  bool isAnyGathered(const SmallDenseSet<Value *> &Vals) const {
    return any_of(MustGather, [&](Value *V) { return Vals.contains(V); });
  }
  /// Checks if the given value is gathered in one of the nodes.
  bool isGathered(const Value *V) const {
    return MustGather.contains(V);
  }
  /// Checks if the specified value was not schedule.
  bool isNotScheduled(const Value *V) const {
    return NonScheduledFirst.contains(V);
  }
 
  /// Check if \p V is a peeled reassociated scalar still owned by a live
  /// (non-deleted, non-gathered) tree entry.
  bool isReassocScalarVectorized(const Value *V) const {
    auto It = ReassocScalarToTreeEntries.find(V);
    return It != ReassocScalarToTreeEntries.end() &&
           any_of(It->second, [&](const TreeEntry *E) {
             return !DeletedNodes.contains(E) &&
                    !TransformedToGatherNodes.contains(E);
           });
  }
 
  /// Check if the value is vectorized in the tree.
  bool isVectorized(const Value *V) const {
    assert(V && "V cannot be nullptr.");
    if (isReassocScalarVectorized(V))
      return true;
    return any_of(getTreeEntries(V), [&](const TreeEntry *E) {
      return !DeletedNodes.contains(E) && !TransformedToGatherNodes.contains(E);
    });
  }
 
  /// Returns true if the role of \p I is already decided by its user: a deleted
  /// user was folded into some other vector by an earlier attempt.
  bool hasResolvedUser(Instruction *I) const {
    return any_of(I->users(), [&](User *U) {
      auto *UI = dyn_cast<Instruction>(U);
      return UI && isDeleted(UI);
    });
  }
 
  /// Checks if it is legal and profitable to build SplitVectorize node for the
  /// given \p VL.
  /// \param Op1 first homogeneous scalars.
  /// \param Op2 second homogeneous scalars.
  /// \param ReorderIndices indices to reorder the scalars.
  /// \returns true if the node was successfully built.
  bool canBuildSplitNode(ArrayRef<Value *> VL,
                         const InstructionsState &LocalState,
                         SmallVectorImpl<Value *> &Op1,
                         SmallVectorImpl<Value *> &Op2,
                         OrdersType &ReorderIndices) const;
 
  ~BoUpSLP();
 
private:
  /// Determine if a node \p E in can be demoted to a smaller type with a
  /// truncation. We collect the entries that will be demoted in ToDemote.
  /// \param E Node for analysis
  /// \param ToDemote indices of the nodes to be demoted.
  bool collectValuesToDemote(
      const TreeEntry &E, bool IsProfitableToDemoteRoot, unsigned &BitWidth,
      SmallVectorImpl<unsigned> &ToDemote, DenseSet<const TreeEntry *> &Visited,
      const SmallDenseSet<unsigned, 8> &NodesToKeepBWs, unsigned &MaxDepthLevel,
      bool &IsProfitableToDemote, bool IsTruncRoot) const;
 
  /// Builds the list of reorderable operands on the edges \p Edges of the \p
  /// UserTE, which allow reordering (i.e. the operands can be reordered because
  /// they have only one user and reordarable).
  /// \param ReorderableGathers List of all gather nodes that require reordering
  /// (e.g., gather of extractlements or partially vectorizable loads).
  /// \param GatherOps List of gather operand nodes for \p UserTE that require
  /// reordering, subset of \p NonVectorized.
  void buildReorderableOperands(
      TreeEntry *UserTE,
      SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges,
      const SmallPtrSetImpl<const TreeEntry *> &ReorderableGathers,
      SmallVectorImpl<TreeEntry *> &GatherOps);
 
  /// Checks if the given \p TE is a gather node with clustered reused scalars
  /// and reorders it per given \p Mask.
  void reorderNodeWithReuses(TreeEntry &TE, ArrayRef<int> Mask) const;
 
  /// Checks if all users of \p I are the part of the vectorization tree.
  bool areAllUsersVectorized(
      Instruction *I,
      const SmallDenseSet<Value *> *VectorizedVals = nullptr) const;
 
  /// Estimates the number of scalar instructions in the tree, each weighted by
  /// its loop-nest trip count (nest-invariant entries are dropped when
  /// \p TreeLoop is non-null).
  uint64_t getNumScalarInsts(bool HasTreeLoop);
 
  /// Estimates the number of vector instructions (including buildvectors,
  /// shuffles, and extracts) the tree produces, weighted like
  /// getNumScalarInsts().
  uint64_t getNumVectorInsts(bool HasTreeLoop);
 
  /// Return information about the vector formed for the specified index
  /// of a vector of (the same) instruction.
  TargetTransformInfo::OperandValueInfo
  getOperandInfo(ArrayRef<Value *> Ops) const;
 
  /// \returns the graph entry for the \p Idx operand of the \p E entry.
  const TreeEntry *getOperandEntry(const TreeEntry *E, unsigned Idx) const;
  TreeEntry *getOperandEntry(TreeEntry *E, unsigned Idx) {
    return const_cast<TreeEntry *>(
        getOperandEntry(const_cast<const TreeEntry *>(E), Idx));
  }
 
  /// Gets the root instruction for the given node. If the node is a strided
  /// load/store node with the reverse order, the root instruction is the last
  /// one.
  Instruction *getRootEntryInstruction(const TreeEntry &Entry) const;
 
  /// \returns Cast context for the given graph node.
  TargetTransformInfo::CastContextHint
  getCastContextHint(const TreeEntry &TE) const;
 
  /// \returns the scale of the given tree entry to the loop iteration.
  /// \p Scalar is the scalar value from the entry, if using the parent for the
  /// external use.
  /// \p U is the user of the vectorized value from the entry, if using the
  /// parent for the external use.
  uint64_t getScaleToLoopIterations(const TreeEntry &TE,
                                    Value *Scalar = nullptr,
                                    Instruction *U = nullptr);
 
  /// \returns the product of trip counts of the loop \p L and all of its
  /// enclosing loops. Unlike the state kept by getScaleToLoopIterations(),
  /// this helper depends only on the loop structure and is independent of
  /// per-entry operand invariance. Returns 1 when loop-aware cost modeling
  /// is disabled or \p L is null.
  uint64_t getLoopNestScale(const Loop *L);
 
  /// \returns a refined execution scale for a gather/buildvector tree entry
  /// \p TE. The scale is computed as the average of per-lane execution
  /// scales: each lane's scale is the loop-nest scale of the loop that
  /// contains the lane's defining instruction (or 1 if the lane is a
  /// constant / loop-invariant non-instruction value). This models the
  /// LICM hoisting that optimizeGatherSequence() performs after vectorization
  /// for inserts with loop-invariant operands. Falls back to the whole-entry
  /// scale when per-lane information is unavailable or the feature is off.
  uint64_t getGatherNodeEffectiveScale(const TreeEntry &TE,
                                       Instruction *U = nullptr);
 
  /// \returns the loop-nest execution scale of \p TE.
  uint64_t getEntryEffectiveScale(const TreeEntry &TE,
                                  Instruction *U = nullptr);
 
  /// Get the loop nest for the given loop \p L.
  ArrayRef<const Loop *> getLoopNest(const Loop *L);
 
  /// \returns the cost of the vectorizable entry.
  InstructionCost getEntryCost(const TreeEntry *E,
                               ArrayRef<Value *> VectorizedVals,
                               SmallPtrSetImpl<Value *> &CheckedExtracts);
 
  /// Estimates spill/reload cost from vector register pressure for \p E at the
  /// point of emitting its vector result type \p FinalVecTy. \p ScalarTy is the
  /// scalar/slot type used to widen into \p VecTy/\p FinalVecTy and may itself
  /// be a FixedVectorType in ReVec mode or an adjusted type due to MinBWs.
  InstructionCost
  getVectorSpillReloadCost(const TreeEntry *E, Type *ScalarTy, Type *VecTy,
                           Type *FinalVecTy,
                           const TTI::TargetCostKind CostKind) const;
 
  /// This is the recursive part of buildTree.
  void buildTreeRec(ArrayRef<Value *> Roots, unsigned Depth, const EdgeInfo &EI,
                    unsigned InterleaveFactor = 0);
 
  /// \returns true if the ExtractElement/ExtractValue instructions in \p VL can
  /// be vectorized to use the original vector (or aggregate "bitcast" to a
  /// vector) and sets \p CurrentOrder to the identity permutation; otherwise
  /// returns false, setting \p CurrentOrder to either an empty vector or a
  /// non-identity permutation that allows to reuse extract instructions.
  /// \param ResizeAllowed indicates whether it is allowed to handle subvector
  /// extract order.
  bool canReuseExtract(ArrayRef<Value *> VL,
                       SmallVectorImpl<unsigned> &CurrentOrder,
                       bool ResizeAllowed = false) const;
 
  /// Vectorize a single entry in the tree.
  Value *vectorizeTree(TreeEntry *E);
 
  /// Vectorize a single entry in the tree, the \p Idx-th operand of the entry
  /// \p E.
  Value *vectorizeOperand(TreeEntry *E, unsigned NodeIdx);
 
  /// Create a new vector from a list of scalar values.  Produces a sequence
  /// which exploits values reused across lanes, and arranges the inserts
  /// for ease of later optimization.
  template <typename BVTy, typename ResTy, typename... Args>
  ResTy processBuildVector(const TreeEntry *E, Type *ScalarTy, Args &...Params);
 
  /// Create a new vector from a list of scalar values.  Produces a sequence
  /// which exploits values reused across lanes, and arranges the inserts
  /// for ease of later optimization.
  Value *createBuildVector(const TreeEntry *E, Type *ScalarTy);
 
  /// Returns the instruction in the bundle, which can be used as a base point
  /// for scheduling. Usually it is the last instruction in the bundle, except
  /// for the case when all operands are external (in this case, it is the first
  /// instruction in the list).
  Instruction &getLastInstructionInBundle(const TreeEntry *E);
 
  /// Tries to find extractelement instructions with constant indices from fixed
  /// vector type and gather such instructions into a bunch, which highly likely
  /// might be detected as a shuffle of 1 or 2 input vectors. If this attempt
  /// was successful, the matched scalars are replaced by poison values in \p VL
  /// for future analysis.
  std::optional<TargetTransformInfo::ShuffleKind>
  tryToGatherSingleRegisterExtractElements(MutableArrayRef<Value *> VL,
                                           SmallVectorImpl<int> &Mask) const;
 
  /// Tries to find extractelement instructions with constant indices from fixed
  /// vector type and gather such instructions into a bunch, which highly likely
  /// might be detected as a shuffle of 1 or 2 input vectors. If this attempt
  /// was successful, the matched scalars are replaced by poison values in \p VL
  /// for future analysis.
  SmallVector<std::optional<TargetTransformInfo::ShuffleKind>>
  tryToGatherExtractElements(SmallVectorImpl<Value *> &VL,
                             SmallVectorImpl<int> &Mask,
                             unsigned NumParts) const;
 
  /// Checks if the gathered \p VL can be represented as a single register
  /// shuffle(s) of previous tree entries.
  /// \param TE Tree entry checked for permutation.
  /// \param VL List of scalars (a subset of the TE scalar), checked for
  /// permutations. Must form single-register vector.
  /// \param ForOrder Tries to fetch the best candidates for ordering info. Also
  /// commands to build the mask using the original vector value, without
  /// relying on the potential reordering.
  /// \returns ShuffleKind, if gathered values can be represented as shuffles of
  /// previous tree entries. \p Part of \p Mask is filled with the shuffle mask.
  std::optional<TargetTransformInfo::ShuffleKind>
  isGatherShuffledSingleRegisterEntry(
      const TreeEntry *TE, ArrayRef<Value *> VL, MutableArrayRef<int> Mask,
      SmallVectorImpl<const TreeEntry *> &Entries, unsigned Part, bool ForOrder,
      unsigned SliceSize);
 
  /// Checks if the gathered \p VL can be represented as multi-register
  /// shuffle(s) of previous tree entries.
  /// \param TE Tree entry checked for permutation.
  /// \param VL List of scalars (a subset of the TE scalar), checked for
  /// permutations.
  /// \param ForOrder Tries to fetch the best candidates for ordering info. Also
  /// commands to build the mask using the original vector value, without
  /// relying on the potential reordering.
  /// \returns per-register series of ShuffleKind, if gathered values can be
  /// represented as shuffles of previous tree entries. \p Mask is filled with
  /// the shuffle mask (also on per-register base).
  SmallVector<std::optional<TargetTransformInfo::ShuffleKind>>
  isGatherShuffledEntry(
      const TreeEntry *TE, ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask,
      SmallVectorImpl<SmallVector<const TreeEntry *>> &Entries,
      unsigned NumParts, bool ForOrder = false);
 
  /// \returns the cost of gathering (inserting) the values in \p VL into a
  /// vector.
  /// \param ForPoisonSrc true if initial vector is poison, false otherwise.
  InstructionCost getGatherCost(ArrayRef<Value *> VL, bool ForPoisonSrc,
                                Type *ScalarTy) const;
 
  /// Set the Builder insert point to one after the last instruction in
  /// the bundle
  void setInsertPointAfterBundle(const TreeEntry *E);
 
  /// \returns a vector from a collection of scalars in \p VL. if \p Root is not
  /// specified, the starting vector value is poison.
  Value *
  gather(ArrayRef<Value *> VL, Value *Root, Type *ScalarTy,
         function_ref<Value *(Value *, Value *, ArrayRef<int>)> CreateShuffle);
 
  /// \returns whether the VectorizableTree is fully vectorizable and will
  /// be beneficial even the tree height is tiny.
  bool isFullyVectorizableTinyTree(bool ForReduction) const;
 
  /// Run through the list of all gathered loads in the graph and try to find
  /// vector loads/masked gathers instead of regular gathers. Later these loads
  /// are reshufled to build final gathered nodes.
  void tryToVectorizeGatheredLoads(
      const SmallMapVector<
          std::tuple<BasicBlock *, Value *, Type *>,
          SmallVector<SmallVector<std::pair<LoadInst *, int64_t>>>, 8>
          &GatheredLoads);
 
  /// Run through the gather nodes that are splats of the same instruction and
  /// try to vectorize the unique splatted values together as a separate
  /// subtree. The splat gathers are then emitted as broadcasts of the
  /// vectorized subtree instead of insertion sequences.
  void tryToVectorizeSplatGatheredScalars();
 
  /// Helper for `findExternalStoreUsersReorderIndices()`. It iterates over the
  /// users of \p TE and collects the stores. It returns the map from the store
  /// pointers to the collected stores.
  SmallVector<SmallVector<StoreInst *>>
  collectUserStores(const BoUpSLP::TreeEntry *TE) const;
 
  /// Helper for `findExternalStoreUsersReorderIndices()`. It checks if the
  /// stores in \p StoresVec can form a vector instruction. If so it returns
  /// true and populates \p ReorderIndices with the shuffle indices of the
  /// stores when compared to the sorted vector.
  bool canFormVector(ArrayRef<StoreInst *> StoresVec,
                     OrdersType &ReorderIndices) const;
 
  /// Iterates through the users of \p TE, looking for scalar stores that can be
  /// potentially vectorized in a future SLP-tree. If found, it keeps track of
  /// their order and builds an order index vector for each store bundle. It
  /// returns all these order vectors found.
  /// We run this after the tree has formed, otherwise we may come across user
  /// instructions that are not yet in the tree.
  SmallVector<OrdersType, 1>
  findExternalStoreUsersReorderIndices(TreeEntry *TE) const;
 
  /// Tries to reorder the gathering node for better vectorization
  /// opportunities.
  void reorderGatherNode(TreeEntry &TE);
 
  /// Checks if the tree represents disjoint or reduction of shl(zext, (0, 8,
  /// .., 56))-like pattern.
  /// If the int shifts unique, also strided, but not ordered, sets \p Order.
  /// If the node can be represented as a bitcast + bswap, sets \p IsBSwap.
  /// If the root nodes are loads, sets \p ForLoads to true.
  bool matchesShlZExt(const TreeEntry &TE, OrdersType &Order, bool &IsBSwap,
                      bool &ForLoads) const;
 
  /// Checks if the \p SelectTE matches zext+selects, which can be inversed for
  /// better codegen in case like zext (icmp ne), select (icmp eq), ....
  bool matchesInversedZExtSelect(
      const TreeEntry &SelectTE,
      SmallVectorImpl<unsigned> &InversedCmpsIndices) const;
 
  /// Checks if the tree is reduction or of bit selects, like select %cmp, <1,
  /// 2, 4, 8, ..>, zeroinitializer, which can be reduced just to a bitcast %cmp
  /// to in.
  bool matchesSelectOfBits(const TreeEntry &SelectTE) const;
 
  class TreeEntry {
  public:
    using VecTreeTy = SmallVector<std::unique_ptr<TreeEntry>, 8>;
    TreeEntry(VecTreeTy &Container) : Container(Container) {}
 
    /// \returns Common mask for reorder indices and reused scalars.
    SmallVector<int> getCommonMask() const {
      if (State == TreeEntry::SplitVectorize)
        return {};
      SmallVector<int> Mask;
      inversePermutation(ReorderIndices, Mask);
      addMask(Mask, ReuseShuffleIndices);
      return Mask;
    }
 
    /// \returns The mask for split nodes.
    SmallVector<int> getSplitMask() const {
      assert(State == TreeEntry::SplitVectorize && !ReorderIndices.empty() &&
             "Expected only split vectorize node.");
      unsigned CommonVF = std::max<unsigned>(
          CombinedEntriesWithIndices.back().second,
          Scalars.size() - CombinedEntriesWithIndices.back().second);
      const unsigned Scale = getNumElements(Scalars.front()->getType());
      CommonVF *= Scale;
      SmallVector<int> Mask(getVectorFactor() * Scale, PoisonMaskElem);
      for (auto [Idx, I] : enumerate(ReorderIndices)) {
        for (unsigned K : seq<unsigned>(Scale)) {
          Mask[Scale * I + K] =
              Scale * Idx + K +
              (Idx >= CombinedEntriesWithIndices.back().second
                   ? CommonVF - CombinedEntriesWithIndices.back().second * Scale
                   : 0);
        }
      }
      return Mask;
    }
 
    /// Updates (reorders) SplitVectorize node according to the given mask \p
    /// Mask and order \p MaskOrder.
    void reorderSplitNode(unsigned Idx, ArrayRef<int> Mask,
                          ArrayRef<int> MaskOrder);
 
    /// \returns true if the scalars in VL are equal to this entry.
    bool isSame(ArrayRef<Value *> VL) const {
      auto &&IsSame = [VL](ArrayRef<Value *> Scalars, ArrayRef<int> Mask) {
        if (Mask.size() != VL.size() && VL.size() == Scalars.size())
          return std::equal(VL.begin(), VL.end(), Scalars.begin());
        return VL.size() == Mask.size() &&
               std::equal(VL.begin(), VL.end(), Mask.begin(),
                          [Scalars](Value *V, int Idx) {
                            return isa<PoisonValue>(V) ||
                                   (Idx != PoisonMaskElem && V == Scalars[Idx]);
                          });
      };
      if (!ReorderIndices.empty()) {
        // TODO: implement matching if the nodes are just reordered, still can
        // treat the vector as the same if the list of scalars matches VL
        // directly, without reordering.
        SmallVector<int> Mask;
        inversePermutation(ReorderIndices, Mask);
        if (VL.size() == Scalars.size())
          return IsSame(Scalars, Mask);
        if (VL.size() == ReuseShuffleIndices.size()) {
          addMask(Mask, ReuseShuffleIndices);
          return IsSame(Scalars, Mask);
        }
        return false;
      }
      return IsSame(Scalars, ReuseShuffleIndices);
    }
 
    /// \returns true if current entry has same operands as \p TE.
    bool hasEqualOperands(const TreeEntry &TE) const {
      if (TE.getNumOperands() != getNumOperands())
        return false;
      SmallBitVector Used(getNumOperands());
      for (unsigned I = 0, E = getNumOperands(); I < E; ++I) {
        unsigned PrevCount = Used.count();
        for (unsigned K = 0; K < E; ++K) {
          if (Used.test(K))
            continue;
          if (getOperand(K) == TE.getOperand(I)) {
            Used.set(K);
            break;
          }
        }
        // Check if we actually found the matching operand.
        if (PrevCount == Used.count())
          return false;
      }
      return true;
    }
 
    /// \return Final vectorization factor for the node. Defined by the total
    /// number of vectorized scalars, including those, used several times in the
    /// entry and counted in the \a ReuseShuffleIndices, if any.
    unsigned getVectorFactor() const {
      if (!ReuseShuffleIndices.empty())
        return ReuseShuffleIndices.size();
      return Scalars.size();
    };
 
    /// Checks if the current node is a gather node.
    bool isGather() const { return State == NeedToGather; }
 
    /// A vector of scalars.
    ValueList Scalars;
 
    /// The Scalars are vectorized into this value. It is initialized to Null.
    WeakTrackingVH VectorizedValue = nullptr;
 
    /// Do we need to gather this sequence or vectorize it
    /// (either with vector instruction or with scatter/gather
    /// intrinsics for store/load)?
    enum EntryState {
      Vectorize,            ///< The node is regularly vectorized.
      ScatterVectorize,     ///< Masked scatter/gather node.
      StridedVectorize,     ///< Strided loads (and stores)
      ExpandVectorize,      ///< Masked stores, the values are expanded into
                            ///< a wider vector and vectorized with a mask.
      CompressVectorize,    ///< (Masked) load with compress.
      BlendedLoadVectorize, ///< (Masked) loads blended via `select` from two
                            ///< candidate base pointers.
      NeedToGather,         ///< Gather/buildvector node.
      CombinedVectorize, ///< Vectorized node, combined with its user into more
                         ///< complex node like select/cmp to minmax, mul/add to
                         ///< fma, etc. Must be used for the following nodes in
                         ///< the pattern, not the very first one.
      SplitVectorize,    ///< Splits the node into 2 subnodes, vectorizes them
                         ///< independently and then combines back.
    };
    EntryState State;
 
    /// List of combined opcodes supported by the vectorizer.
    enum CombinedOpcode {
      NotCombinedOp = -1,
      MinMax = Instruction::OtherOpsEnd + 1,
      FMulAdd,
      ReducedBitcast,
      ReducedBitcastBSwap,
      ReducedBitcastLoads,
      ReducedBitcastBSwapLoads,
      ReducedCmpBitcast,
    };
    CombinedOpcode CombinedOp = NotCombinedOp;
 
    /// Does this sequence require some shuffling?
    SmallVector<int, 4> ReuseShuffleIndices;
 
    /// Does this entry require reordering?
    SmallVector<unsigned, 4> ReorderIndices;
 
    /// Points back to the VectorizableTree.
    ///
    /// Only used for Graphviz right now.  Unfortunately GraphTrait::NodeRef has
    /// to be a pointer and needs to be able to initialize the child iterator.
    /// Thus we need a reference back to the container to translate the indices
    /// to entries.
    VecTreeTy &Container;
 
    /// The TreeEntry index containing the user of this entry.
    EdgeInfo UserTreeIndex;
 
    /// The index of this treeEntry in VectorizableTree.
    unsigned Idx = 0;
 
    /// For gather/buildvector/alt opcode nodes, which are combined from
    /// other nodes as a series of insertvector instructions.
    SmallVector<std::pair<unsigned, unsigned>, 2> CombinedEntriesWithIndices;
 
    /// For ExtractValue entries that are vectorized via the struct-call path
    /// (checkEVsForVecCalls succeeded during tree building), stores the common
    /// field-index path shared by all scalars in the bundle. Empty for all
    /// other entry kinds.
    SmallVector<unsigned, 1> StructEVIndices;
 
  private:
    /// The operands of each instruction in each lane Operands[op_index][lane].
    /// Note: This helps avoid the replication of the code that performs the
    /// reordering of operands during buildTreeRec() and vectorizeTree().
    SmallVector<ValueList, 2> Operands;
 
    /// Copyable elements of the entry node.
    SmallPtrSet<const Value *, 4> CopyableElements;
 
    /// Intermediate instructions peeled from an associative chain (e.g. the
    /// inner add in add(add(v0,x),v1)). Not part of Scalars.
    SmallVector<Value *, 4> ReassocScalars;
 
    /// Sign of each flattened operand column of a reassociated add/sub
    /// chain, parallel to the operand columns: a negated column is
    /// subtracted from the positive total. Empty when no column is negated.
    SmallBitVector ReassocNegatedOps;
 
    /// MainOp and AltOp are recorded inside. S should be obtained from
    /// newTreeEntry.
    InstructionsState S = InstructionsState::invalid();
 
    /// Interleaving factor for interleaved loads Vectorize nodes.
    unsigned InterleaveFactor = 0;
 
    /// True if the node does not require scheduling.
    bool DoesNotNeedToSchedule = false;
 
    /// Set this bundle's \p OpIdx'th operand to \p OpVL.
    void setOperand(unsigned OpIdx, ArrayRef<Value *> OpVL) {
      if (Operands.size() < OpIdx + 1)
        Operands.resize(OpIdx + 1);
      assert(Operands[OpIdx].empty() && "Already resized?");
      assert(OpVL.size() <= Scalars.size() &&
             "Number of operands is greater than the number of scalars.");
      Operands[OpIdx].resize(OpVL.size());
      copy(OpVL, Operands[OpIdx].begin());
    }
 
    /// Maps values to their lanes in the node.
    mutable SmallDenseMap<Value *, unsigned> ValueToLane;
 
  public:
    /// Returns interleave factor for interleave nodes.
    unsigned getInterleaveFactor() const { return InterleaveFactor; }
    /// Sets interleaving factor for the interleaving nodes.
    void setInterleave(unsigned Factor) { InterleaveFactor = Factor; }
 
    /// Marks the node as one that does not require scheduling.
    void setDoesNotNeedToSchedule() { DoesNotNeedToSchedule = true; }
    /// Returns true if the node is marked as one that does not require
    /// scheduling.
    bool doesNotNeedToSchedule() const { return DoesNotNeedToSchedule; }
 
    /// Set this bundle's operands from \p Operands.
    void setOperands(ArrayRef<ValueList> Operands) {
      for (unsigned I : seq<unsigned>(Operands.size()))
        setOperand(I, Operands[I]);
    }
 
    /// Reorders operands of the node to the given mask \p Mask.
    void reorderOperands(ArrayRef<int> Mask) {
      for (ValueList &Operand : Operands)
        reorderScalars(Operand, Mask);
    }
 
    /// \returns the \p OpIdx operand of this TreeEntry.
    ValueList &getOperand(unsigned OpIdx) {
      assert(OpIdx < Operands.size() && "Off bounds");
      return Operands[OpIdx];
    }
 
    /// \returns the \p OpIdx operand of this TreeEntry.
    ArrayRef<Value *> getOperand(unsigned OpIdx) const {
      assert(OpIdx < Operands.size() && "Off bounds");
      return Operands[OpIdx];
    }
 
    /// \returns the number of operands.
    unsigned getNumOperands() const { return Operands.size(); }
 
    /// \return the single \p OpIdx operand.
    Value *getSingleOperand(unsigned OpIdx) const {
      assert(OpIdx < Operands.size() && "Off bounds");
      assert(!Operands[OpIdx].empty() && "No operand available");
      return Operands[OpIdx][0];
    }
 
    /// Some of the instructions in the list have alternate opcodes.
    bool isAltShuffle() const { return S.isAltShuffle(); }
 
    Instruction *getMatchingMainOpOrAltOp(Instruction *I) const {
      return S.getMatchingMainOpOrAltOp(I);
    }
 
    /// Chooses the correct key for scheduling data. If \p Op has the same (or
    /// alternate) opcode as \p OpValue, the key is \p Op. Otherwise the key is
    /// \p OpValue.
    Value *isOneOf(Value *Op) const {
      auto *I = dyn_cast<Instruction>(Op);
      if (I && getMatchingMainOpOrAltOp(I))
        return Op;
      return S.getMainOp();
    }
 
    void setOperations(const InstructionsState &S) {
      assert(S && "InstructionsState is invalid.");
      this->S = S;
    }
 
    Instruction *getMainOp() const { return S.getMainOp(); }
 
    Instruction *getAltOp() const { return S.getAltOp(); }
 
    /// The main/alternate opcodes for the list of instructions.
    unsigned getOpcode() const { return S.getOpcode(); }
 
    unsigned getAltOpcode() const { return S.getAltOpcode(); }
 
    bool hasState() const { return S.valid(); }
 
    /// Add \p V to the list of copyable elements.
    void addCopyableElement(Value *V) {
      assert(S.isCopyableElement(V) && "Not a copyable element.");
      CopyableElements.insert(V);
    }
 
    /// Returns true if \p V is a copyable element.
    bool isCopyableElement(Value *V) const {
      return CopyableElements.contains(V);
    }
 
    /// Checks if the value \p V is a transformed instruction, compatible either
    /// with main or alternate ops.
    bool isExpandedBinOp(Value *V) const {
      assert(hasState() && "InstructionsState is invalid.");
      if (isCopyableElement(V))
        return false;
      return S.isExpandedBinOp(V);
    }
 
    /// Checks if the operand at index \p Idx of instruction \p I is an expanded
    /// operand.
    bool isExpandedOperand(Instruction *I, unsigned Idx) const {
      assert(hasState() && "InstructionsState is invalid.");
      if (isCopyableElement(I))
        return false;
      if (!isExpandedBinOp(I))
        return false;
      return S.isExpandedOperand(I, Idx);
    }
 
    /// Returns true if any scalar in the list is a copyable element.
    bool hasCopyableElements() const { return !CopyableElements.empty(); }
 
    /// Adds \p V to the peeled reassociated scalars.
    void addReassocScalar(Value *V) { ReassocScalars.push_back(V); }
 
    /// True if operands were gathered from an associative chain.
    bool hasReassocScalars() const { return !ReassocScalars.empty(); }
 
    /// Returns peeled reassociated scalars.
    ArrayRef<Value *> getReassocScalars() const { return ReassocScalars; }
 
    /// Records the signs of the flattened operand columns.
    void setReassocNegatedOps(const SmallBitVector &NegatedOps) {
      assert(NegatedOps.size() == getNumOperands() &&
             "Signs must cover all operand columns.");
      ReassocNegatedOps = NegatedOps;
    }
 
    /// True if operand column \p Idx is subtracted rather than added.
    bool isReassocNegatedOp(unsigned Idx) const {
      return Idx < ReassocNegatedOps.size() && ReassocNegatedOps[Idx];
    }
 
    /// Returns the state of the operations.
    const InstructionsState &getOperations() const { return S; }
 
    /// When ReuseReorderShuffleIndices is empty it just returns position of \p
    /// V within vector of Scalars. Otherwise, try to remap on its reuse index.
    unsigned findLaneForValue(Value *V) const {
      auto Res = ValueToLane.try_emplace(V, getVectorFactor());
      if (!Res.second)
        return Res.first->second;
      unsigned &FoundLane = Res.first->getSecond();
      // Poison can take any lane, match it to the lane of the first non-poison
      // scalar.
      auto IsMatch = [V](Value *S) {
        return isa<PoisonValue>(V) ? !isa<PoisonValue>(S) : S == V;
      };
      for (auto *It = find_if(Scalars, IsMatch), *End = Scalars.end();
           It != End; std::advance(It, 1)) {
        if (!IsMatch(*It))
          continue;
        FoundLane = std::distance(Scalars.begin(), It);
        assert(FoundLane < Scalars.size() && "Couldn't find extract lane");
        if (!ReorderIndices.empty())
          FoundLane = ReorderIndices[FoundLane];
        assert(FoundLane < Scalars.size() && "Couldn't find extract lane");
        if (ReuseShuffleIndices.empty())
          break;
        if (auto *RIt = find(ReuseShuffleIndices, FoundLane);
            RIt != ReuseShuffleIndices.end()) {
          FoundLane = std::distance(ReuseShuffleIndices.begin(), RIt);
          break;
        }
      }
      assert(FoundLane < getVectorFactor() && "Unable to find given value.");
      return FoundLane;
    }
 
    /// Build a shuffle mask for graph entry which represents a merge of main
    /// and alternate operations.
    void
    buildAltOpShuffleMask(const function_ref<bool(Instruction *)> IsAltOp,
                          SmallVectorImpl<int> &Mask,
                          SmallVectorImpl<Value *> *OpScalars = nullptr,
                          SmallVectorImpl<Value *> *AltScalars = nullptr) const;
 
    /// Return true if this is a non-power-of-2 node.
    bool isNonPowOf2Vec() const {
      bool IsNonPowerOf2 = !has_single_bit(Scalars.size());
      return IsNonPowerOf2;
    }
 
    Value *getOrdered(unsigned Idx) const {
      if (ReorderIndices.empty())
        return Scalars[Idx];
      SmallVector<int> Mask;
      inversePermutation(ReorderIndices, Mask);
      return Scalars[Mask[Idx]];
    }
 
#ifndef NDEBUG
    /// Debug printer.
    LLVM_DUMP_METHOD void dump() const {
      dbgs() << Idx << ".\n";
      for (unsigned OpI = 0, OpE = Operands.size(); OpI != OpE; ++OpI) {
        dbgs() << "Operand " << OpI << ":\n";
        for (const Value *V : Operands[OpI])
          dbgs().indent(2) << *V << "\n";
      }
      dbgs() << "Scalars: \n";
      for (Value *V : Scalars) {
        dbgs().indent(2) << *V
                         << ((S && S.isExpandedBinOp(V)) ? " [[Expanded]]\n"
                                                         : "\n");
      }
      dbgs() << "State: ";
      if (S && hasCopyableElements())
        dbgs() << "[[Copyable]] ";
      switch (State) {
      case Vectorize:
        if (InterleaveFactor > 0) {
          dbgs() << "Vectorize with interleave factor " << InterleaveFactor
                 << "\n";
        } else {
          dbgs() << "Vectorize\n";
        }
        break;
      case ScatterVectorize:
        dbgs() << "ScatterVectorize\n";
        break;
      case StridedVectorize:
        dbgs() << "StridedVectorize\n";
        break;
      case ExpandVectorize:
        dbgs() << "ExpandVectorize\n";
        break;
      case CompressVectorize:
        dbgs() << "CompressVectorize\n";
        break;
      case BlendedLoadVectorize:
        dbgs() << "BlendedLoadVectorize\n";
        break;
      case NeedToGather:
        dbgs() << "NeedToGather\n";
        break;
      case CombinedVectorize:
        dbgs() << "CombinedVectorize\n";
        break;
      case SplitVectorize:
        dbgs() << "SplitVectorize\n";
        break;
      }
      if (S) {
        dbgs() << "MainOp: " << *S.getMainOp() << "\n";
        dbgs() << "AltOp: " << *S.getAltOp() << "\n";
      } else {
        dbgs() << "MainOp: NULL\n";
        dbgs() << "AltOp: NULL\n";
      }
      dbgs() << "VectorizedValue: ";
      if (VectorizedValue)
        dbgs() << *VectorizedValue << "\n";
      else
        dbgs() << "NULL\n";
      dbgs() << "ReuseShuffleIndices: ";
      if (ReuseShuffleIndices.empty())
        dbgs() << "Empty";
      else
        for (int ReuseIdx : ReuseShuffleIndices)
          dbgs() << ReuseIdx << ", ";
      dbgs() << "\n";
      dbgs() << "ReorderIndices: ";
      for (unsigned ReorderIdx : ReorderIndices)
        dbgs() << ReorderIdx << ", ";
      dbgs() << "\n";
      dbgs() << "UserTreeIndex: ";
      if (UserTreeIndex)
        dbgs() << UserTreeIndex;
      else
        dbgs() << "<invalid>";
      dbgs() << "\n";
      if (!StructEVIndices.empty()) {
        dbgs() << "StructEVIndices: ";
        interleaveComma(StructEVIndices, dbgs());
        dbgs() << "\n";
      }
      if (!CombinedEntriesWithIndices.empty()) {
        dbgs() << "Combined entries: ";
        interleaveComma(CombinedEntriesWithIndices, dbgs(), [&](const auto &P) {
          dbgs() << "Entry index " << P.first << " with offset " << P.second;
        });
        dbgs() << "\n";
      }
    }
#endif
  };
 
#ifndef NDEBUG
  void dumpTreeCosts(const TreeEntry *E, InstructionCost ReuseShuffleCost,
                     InstructionCost VecCost, InstructionCost ScalarCost,
                     StringRef Banner) const {
    dbgs() << "SLP: " << Banner << ":\n";
    E->dump();
    dbgs() << "SLP: Costs:\n";
    dbgs() << "SLP:     ReuseShuffleCost = " << ReuseShuffleCost << "\n";
    dbgs() << "SLP:     VectorCost = " << VecCost << "\n";
    dbgs() << "SLP:     ScalarCost = " << ScalarCost << "\n";
    dbgs() << "SLP:     ReuseShuffleCost + VecCost - ScalarCost = "
           << ReuseShuffleCost + VecCost - ScalarCost << "\n";
  }
#endif
 
  /// Create a new gather TreeEntry
  TreeEntry *newGatherTreeEntry(ArrayRef<Value *> VL,
                                const InstructionsState &S,
                                const EdgeInfo &UserTreeIdx,
                                ArrayRef<int> ReuseShuffleIndices = {}) {
    auto Invalid = ScheduleBundle::invalid();
    return newTreeEntry(VL, Invalid, S, UserTreeIdx, ReuseShuffleIndices);
  }
 
  /// Create a new VectorizableTree entry.
  TreeEntry *newTreeEntry(ArrayRef<Value *> VL, ScheduleBundle &Bundle,
                          const InstructionsState &S,
                          const EdgeInfo &UserTreeIdx,
                          ArrayRef<int> ReuseShuffleIndices = {},
                          ArrayRef<unsigned> ReorderIndices = {},
                          unsigned InterleaveFactor = 0) {
    TreeEntry::EntryState EntryState =
        Bundle ? TreeEntry::Vectorize : TreeEntry::NeedToGather;
    TreeEntry *E = newTreeEntry(VL, EntryState, Bundle, S, UserTreeIdx,
                                ReuseShuffleIndices, ReorderIndices);
    if (E && InterleaveFactor > 0)
      E->setInterleave(InterleaveFactor);
    return E;
  }
 
  TreeEntry *newTreeEntry(ArrayRef<Value *> VL,
                          TreeEntry::EntryState EntryState,
                          ScheduleBundle &Bundle, const InstructionsState &S,
                          const EdgeInfo &UserTreeIdx,
                          ArrayRef<int> ReuseShuffleIndices = {},
                          ArrayRef<unsigned> ReorderIndices = {}) {
    assert(((!Bundle && (EntryState == TreeEntry::NeedToGather ||
                         EntryState == TreeEntry::SplitVectorize)) ||
            (Bundle && EntryState != TreeEntry::NeedToGather &&
             EntryState != TreeEntry::SplitVectorize)) &&
           "Need to vectorize gather entry?");
    // Gathered loads still gathered? Do not create entry, use the original one.
    if (GatheredLoadsEntriesFirst.has_value() &&
        EntryState == TreeEntry::NeedToGather && S &&
        S.getOpcode() == Instruction::Load && UserTreeIdx.EdgeIdx == UINT_MAX &&
        !UserTreeIdx.UserTE)
      return nullptr;
    VectorizableTree.push_back(std::make_unique<TreeEntry>(VectorizableTree));
    TreeEntry *Last = VectorizableTree.back().get();
    Last->Idx = VectorizableTree.size() - 1;
    Last->State = EntryState;
    if (UserTreeIdx.UserTE)
      OperandsToTreeEntry.try_emplace(
          std::make_pair(UserTreeIdx.UserTE, UserTreeIdx.EdgeIdx), Last);
    Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(),
                                     ReuseShuffleIndices.end());
    if (ReorderIndices.empty()) {
      Last->Scalars.assign(VL.begin(), VL.end());
      if (S)
        Last->setOperations(S);
    } else {
      // Reorder scalars and build final mask.
      Last->Scalars.assign(VL.size(), nullptr);
      transform(ReorderIndices, Last->Scalars.begin(),
                [VL](unsigned Idx) -> Value * {
                  if (Idx >= VL.size())
                    return UndefValue::get(VL.front()->getType());
                  return VL[Idx];
                });
      InstructionsState S = getSameOpcode(Last->Scalars, *TLI);
      if (S)
        Last->setOperations(S);
      Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end());
    }
    if (EntryState == TreeEntry::SplitVectorize) {
      assert(S && "Split nodes must have operations.");
      Last->setOperations(S);
      SmallPtrSet<Value *, 4> Processed;
      for (Value *V : VL) {
        auto *I = dyn_cast<Instruction>(V);
        if (!I)
          continue;
        auto It = ScalarsInSplitNodes.find(V);
        if (It == ScalarsInSplitNodes.end()) {
          ScalarsInSplitNodes.try_emplace(V).first->getSecond().push_back(Last);
          (void)Processed.insert(V);
        } else if (Processed.insert(V).second) {
          assert(!is_contained(It->getSecond(), Last) &&
                 "Value already associated with the node.");
          It->getSecond().push_back(Last);
        }
      }
    } else if (!Last->isGather()) {
      if (isa<PHINode>(S.getMainOp()) ||
          isVectorLikeInstWithConstOps(S.getMainOp()) ||
          (!S.areInstructionsWithCopyableElements() &&
           doesNotNeedToSchedule(VL)) ||
          all_of(VL, [&](Value *V) { return S.isNonSchedulable(V); }))
        Last->setDoesNotNeedToSchedule();
      SmallPtrSet<Value *, 4> Processed;
      for (Value *V : VL) {
        if (isa<PoisonValue>(V))
          continue;
        if (S.isCopyableElement(V)) {
          Last->addCopyableElement(V);
          continue;
        }
        auto It = ScalarToTreeEntries.find(V);
        if (It == ScalarToTreeEntries.end()) {
          ScalarToTreeEntries.try_emplace(V).first->getSecond().push_back(Last);
          (void)Processed.insert(V);
        } else if (Processed.insert(V).second) {
          assert(!is_contained(It->getSecond(), Last) &&
                 "Value already associated with the node.");
          It->getSecond().push_back(Last);
        }
      }
      // Update the scheduler bundle to point to this TreeEntry.
      assert((!Bundle.getBundle().empty() || Last->doesNotNeedToSchedule()) &&
             "Bundle and VL out of sync");
      if (!Bundle.getBundle().empty()) {
#if !defined(NDEBUG) || defined(EXPENSIVE_CHECKS)
        auto *BundleMember = Bundle.getBundle().begin();
        SmallPtrSet<Value *, 4> Processed;
        for (Value *V : VL) {
          if (S.isNonSchedulable(V) || !Processed.insert(V).second)
            continue;
          ++BundleMember;
        }
        assert(BundleMember == Bundle.getBundle().end() &&
               "Bundle and VL out of sync");
#endif
        Bundle.setTreeEntry(Last);
      }
    } else {
      // Build a map for gathered scalars to the nodes where they are used.
      bool AllConstsOrCasts = true;
      for (Value *V : VL) {
        if (S && S.areInstructionsWithCopyableElements() &&
            S.isCopyableElement(V))
          Last->addCopyableElement(V);
        if (!isConstant(V)) {
          auto *I = dyn_cast<CastInst>(V);
          AllConstsOrCasts &= I && I->getType()->isIntegerTy();
          if (UserTreeIdx.EdgeIdx != UINT_MAX || !UserTreeIdx.UserTE ||
              !UserTreeIdx.UserTE->isGather())
            ValueToGatherNodes.try_emplace(V).first->getSecond().insert(Last);
        }
      }
      if (AllConstsOrCasts)
        CastMaxMinBWSizes =
            std::make_pair(std::numeric_limits<unsigned>::max(), 1);
      MustGather.insert_range(VL);
    }
 
    if (UserTreeIdx.UserTE)
      Last->UserTreeIndex = UserTreeIdx;
    return Last;
  }
 
  /// -- Vectorization State --
  /// Holds all of the tree entries.
  TreeEntry::VecTreeTy VectorizableTree;
 
#ifndef NDEBUG
  /// Debug printer.
  LLVM_DUMP_METHOD void dumpVectorizableTree() const {
    for (unsigned Id = 0, IdE = VectorizableTree.size(); Id != IdE; ++Id) {
      VectorizableTree[Id]->dump();
      if (TransformedToGatherNodes.contains(VectorizableTree[Id].get()))
        dbgs() << "[[TRANSFORMED TO GATHER]]";
      else if (DeletedNodes.contains(VectorizableTree[Id].get()))
        dbgs() << "[[DELETED NODE]]";
      dbgs() << "\n";
    }
  }
#endif
 
  /// Get list of vector entries, associated with the value \p V.
  ArrayRef<TreeEntry *> getTreeEntries(const Value *V) const {
    assert(V && "V cannot be nullptr.");
    auto It = ScalarToTreeEntries.find(V);
    if (It == ScalarToTreeEntries.end())
      return {};
    return It->getSecond();
  }
 
  /// Get list of split vector entries, associated with the value \p V.
  ArrayRef<TreeEntry *> getSplitTreeEntries(Value *V) const {
    assert(V && "V cannot be nullptr.");
    auto It = ScalarsInSplitNodes.find(V);
    if (It == ScalarsInSplitNodes.end())
      return {};
    return It->getSecond();
  }
 
  /// Returns first vector node for value \p V, matching values \p VL.
  TreeEntry *getSameValuesTreeEntry(Value *V, ArrayRef<Value *> VL,
                                    bool SameVF = false) const {
    assert(V && "V cannot be nullptr.");
    for (TreeEntry *TE : ScalarToTreeEntries.lookup(V))
      if ((!SameVF || TE->getVectorFactor() == VL.size()) && TE->isSame(VL))
        return TE;
    return nullptr;
  }
 
  /// Contains all the outputs of legality analysis for a list of values to
  /// vectorize.
  class ScalarsVectorizationLegality {
    InstructionsState S;
    bool IsLegal;
    bool TryToFindDuplicates;
    bool TrySplitVectorize;
 
  public:
    ScalarsVectorizationLegality(InstructionsState S, bool IsLegal,
                                 bool TryToFindDuplicates = true,
                                 bool TrySplitVectorize = false)
        : S(S), IsLegal(IsLegal), TryToFindDuplicates(TryToFindDuplicates),
          TrySplitVectorize(TrySplitVectorize) {
      assert((!IsLegal || (S.valid() && TryToFindDuplicates)) &&
             "Inconsistent state");
    }
    const InstructionsState &getInstructionsState() const { return S; };
    bool isLegal() const { return IsLegal; }
    bool tryToFindDuplicates() const { return TryToFindDuplicates; }
    bool trySplitVectorize() const { return TrySplitVectorize; }
  };
 
  /// Checks if the specified list of the instructions/values can be vectorized
  /// in general.
  ScalarsVectorizationLegality
  getScalarsVectorizationLegality(ArrayRef<Value *> VL, unsigned Depth,
                                  const EdgeInfo &UserTreeIdx) const;
 
  /// Checks if the specified list of the instructions/values can be vectorized
  /// and fills required data before actual scheduling of the instructions.
  TreeEntry::EntryState getScalarsVectorizationState(
      const InstructionsState &S, ArrayRef<Value *> VL,
      bool IsScatterVectorizeUserTE, OrdersType &CurrentOrder,
      SmallVectorImpl<Value *> &PointerOps, StridedPtrInfo &SPtrInfo,
      SmallVectorImpl<int> &ReuseShuffleIndices);
 
  /// Maps a specific scalar to its tree entry(ies).
  SmallDenseMap<Value *, SmallVector<TreeEntry *>> ScalarToTreeEntries;
 
  /// List of deleted non-profitable nodes.
  SmallPtrSet<const TreeEntry *, 8> DeletedNodes;
 
  /// List of nodes, transformed to gathered, with their conservative
  /// gather/buildvector cost estimation.
  SmallDenseMap<const TreeEntry *, InstructionCost> TransformedToGatherNodes;
 
  /// Maps the operand index and entry to the corresponding tree entry.
  SmallDenseMap<std::pair<const TreeEntry *, unsigned>, TreeEntry *>
      OperandsToTreeEntry;
 
  /// Scalars, used in split vectorize nodes.
  SmallDenseMap<Value *, SmallVector<TreeEntry *>> ScalarsInSplitNodes;
 
  /// Maps a value to the proposed vectorizable size.
  SmallDenseMap<Value *, unsigned> InstrElementSize;
 
  /// A list of scalars that we found that we need to keep as scalars.
  ValueSet MustGather;
 
  /// Maps each peeled reassociated scalar to owning entries. Keeps them
  /// treated as vectorized while an owner is live.
  SmallDenseMap<const Value *, SmallVector<const TreeEntry *>>
      ReassocScalarToTreeEntries;
 
  /// Peeled reassociated scalars that must survive erasure: claimed by a
  /// gather node, listed in some tree entry's scalars, or feeding another
  /// kept scalar.
  SmallPtrSet<const Value *, 8> KeptReassocScalars;
 
  /// A set of first non-schedulable values.
  ValueSet NonScheduledFirst;
 
  /// A map between the vectorized entries and the last instructions in the
  /// bundles. The bundles are built in use order, not in the def order of the
  /// instructions. So, we cannot rely directly on the last instruction in the
  /// bundle being the last instruction in the program order during
  /// vectorization process since the basic blocks are affected, need to
  /// pre-gather them before.
  SmallDenseMap<const TreeEntry *, WeakTrackingVH> EntryToLastInstruction;
 
  /// Keeps the mapping between the last instructions and their insertion
  /// points, which is an instruction-after-the-last-instruction.
  SmallDenseMap<const Instruction *, Instruction *> LastInstructionToPos;
 
  /// List of gather nodes, depending on other gather/vector nodes, which should
  /// be emitted after the vector instruction emission process to correctly
  /// handle order of the vector instructions and shuffles.
  SetVector<const TreeEntry *> PostponedGathers;
 
  using ValueToGatherNodesMap =
      DenseMap<Value *, SmallSetVector<const TreeEntry *, 4>>;
  ValueToGatherNodesMap ValueToGatherNodes;
 
  SmallDenseMap<TreeEntry *, StridedPtrInfo> TreeEntryToStridedPtrInfoMap;
 
  /// A list of the load entries (node indices), which can be vectorized using
  /// strided or masked gather approach, but attempted to be represented as
  /// contiguous loads.
  SetVector<unsigned> LoadEntriesToVectorize;
 
  /// true if graph nodes transforming mode is on.
  bool IsGraphTransformMode = false;
 
  /// The index of the first gathered load entry in the VectorizeTree.
  std::optional<unsigned> GatheredLoadsEntriesFirst;
 
  /// Root entries of the subtrees built for the splat gather nodes' unique
  /// scalars. They have no users in the tree and must be emitted explicitly
  /// before the root node.
  SmallVector<TreeEntry *> SplatGatheredScalarsRoots;
 
  /// Maps compress entries to their mask data for the final codegen.
  SmallDenseMap<const TreeEntry *,
                std::tuple<SmallVector<int>, VectorType *, unsigned, bool>>
      CompressEntryToData;
 
  /// The loop nest, used to check if only a single loop nest is vectorized, not
  /// multiple, to avoid side-effects from the loop-aware cost model.
  SmallVector<const Loop *> CurrentLoopNest;
 
  /// Per-depth SCEVs trip counts at every loop level where the tree builder has
  /// joined diverging sibling loops.
  SmallVector<const SCEV *> MergedLoopBTCs;
 
  /// Maps the loops to their loop nests.
  SmallDenseMap<const Loop *, SmallVector<const Loop *>> LoopToLoopNest;
 
  /// Per-loop cache of nest scale factors: the product of trip counts of the
  /// loop and all of its ancestors. Shared by getLoopNestScale() and (via it)
  /// by getScaleToLoopIterations() and getGatherNodeEffectiveScale().
  SmallDenseMap<const Loop *, uint64_t> LoopNestScaleCache;
 
  /// This POD struct describes one external user in the vectorized tree.
  struct ExternalUser {
    ExternalUser(Value *S, llvm::User *U, const TreeEntry &E, unsigned L)
        : Scalar(S), User(U), E(E), Lane(L) {}
 
    /// Which scalar in our function.
    Value *Scalar = nullptr;
 
    /// Which user that uses the scalar.
    llvm::User *User = nullptr;
 
    /// Vector node, the value is part of.
    const TreeEntry &E;
 
    /// Which lane does the scalar belong to.
    unsigned Lane;
  };
  using UserList = SmallVector<ExternalUser, 16>;
 
  /// Checks if two instructions may access the same memory.
  ///
  /// \p Loc1 is the location of \p Inst1. It is passed explicitly because it
  /// is invariant in the calling loop.
  bool isAliased(const MemoryLocation &Loc1, Instruction *Inst1,
                 Instruction *Inst2) {
    assert(Loc1.Ptr && isSimple(Inst1) && "Expected simple first instruction.");
    // First check if the result is already in the cache.
    AliasCacheKey Key = std::make_pair(Inst1, Inst2);
    auto Res = AliasCache.try_emplace(Key);
    if (!Res.second)
      return Res.first->second;
    bool Aliased = isModOrRefSet(BatchAA.getModRefInfo(Inst2, Loc1));
    // Store the result in the cache.
    Res.first->getSecond() = Aliased;
    return Aliased;
  }
 
  /// Returns true if the may-alias dependency between simple load/store
  /// instructions \p Inst1 and \p Inst2 could be disambiguated by a runtime
  /// alias check.
  bool isRuntimeCheckableAliasPair(Instruction *Inst1, Instruction *Inst2);
 
  /// Records the (distinct base object) pair behind the may-alias dependency
  /// of \p Inst1 and \p Inst2 as a runtime alias check guarding the region in
  /// block \p BB. Returns true if the pair was recorded.
  bool recordRuntimeAliasCheck(BasicBlock *BB, Instruction *Inst1,
                               Instruction *Inst2);
 
  /// Emits the collected runtime alias checks and versions the affected block,
  /// duplicating its body into a scalar fallback guarded by the checks.
  void versionBlocksForRuntimeChecks();
 
  /// Builds the i1 value that is true when any pair of checked base objects
  /// overlaps at runtime. The base address bounds are materialized from their
  /// SCEVs with \p Exp.
  Value *emitRuntimeAliasCheck(IRBuilderBase &Builder, SCEVExpander &Exp);
 
  /// Data to model and emit the runtime alias checks.
  struct RuntimeAliasCheckInfo {
    /// The block whose body is guarded by the checks. Exactly one block is
    /// supported per attempt.
    BasicBlock *BB = nullptr;
    /// Pairs of base objects that must be proven disjoint.
    SmallSetVector<std::pair<const Value *, const Value *>, 4> BasePairs;
    /// Accessed address range [Low, High) for each involved base object.
    SmallMapVector<const Value *, std::pair<const SCEV *, const SCEV *>, 4>
        Bounds;
 
    void clear() {
      BB = nullptr;
      BasePairs.clear();
      Bounds.clear();
    }
  };
 
  /// When true, scheduling drops may-alias memory dependencies between
  /// distinct, range-checkable base objects and records them as runtime alias
  /// checks instead.
  bool TryRuntimeAliasChecks = false;
 
  /// Runtime alias checks collected during the last optimistic buildTree().
  RuntimeAliasCheckInfo RTChecks;
 
  /// Base-object pairs already proven disjoint by the block's runtime alias
  /// check.
  SmallDenseMap<BasicBlock *,
                SmallDenseSet<std::pair<const Value *, const Value *>, 4>, 2>
      VersionedBlockCheckedPairs;
 
  /// Scalar fallback blocks.
  SmallPtrSet<BasicBlock *, 4> ScalarFallbackBlocks;
 
  /// Blocks for which a runtime-checks versioning attempt was made
  /// and did not produce a profitable versioning.
  SmallPtrSet<BasicBlock *, 8> FailedRuntimeChecksBlocks;
 
  /// Returns true if a may-alias dependency between the simple load/store
  /// instructions \p Inst1 and \p Inst2 in block \p BB is already covered by a
  /// runtime alias check emitted for \p BB by a previous versioning.
  bool isCoveredByExistingVersionCheck(BasicBlock *BB, Instruction *Inst1,
                                       Instruction *Inst2) const;
 
  /// True, if a may-alias dependency between distinct, range-checkable base
  /// objects is observed (whether or not it was dropped).
  bool HasRuntimeCheckableBlockers = false;
 
  /// True, if a kept may-alias dependency is not runtime-checkable (call or a
  /// non-simple memaccess).
  bool HasNonCheckableMemBlocker = false;
 
  /// Runtime checks are validated and bounded the collected checks.
  bool RTChecksFinalized = false;
 
  /// Set when a block was versioned with runtime alias checks, which changes
  /// the CFG. Used to drop CFG-analysis preservation for the run.
  bool CFGChanged = false;
 
  /// Guarded block body (non-PHI, non-terminator) in original source order.
  SmallVector<Instruction *> RTOrigBodyOrder;
 
  using AliasCacheKey = std::pair<Instruction *, Instruction *>;
 
  /// Cache for alias results.
  /// TODO: consider moving this to the AliasAnalysis itself.
  SmallDenseMap<AliasCacheKey, bool> AliasCache;
 
  // Cache for pointerMayBeCaptured calls inside AA.  This is preserved
  // globally through SLP because we don't perform any action which
  // invalidates capture results.
  BatchAAResults BatchAA;
 
  /// Temporary store for deleted instructions. Instructions will be deleted
  /// eventually when the BoUpSLP is destructed.  The deferral is required to
  /// ensure that there are no incorrect collisions in the AliasCache, which
  /// can happen if a new instruction is allocated at the same address as a
  /// previously deleted instruction.
  DenseSet<Instruction *> DeletedInstructions;
 
  /// Set of the instruction, being analyzed already for reductions.
  SmallPtrSet<Instruction *, 16> AnalyzedReductionsRoots;
 
  /// Set of hashes for the list of reduction values already being analyzed.
  DenseSet<size_t> AnalyzedReductionVals;
 
  /// Set of hashes for the bundles, rejected as non-vectorizable.
  SmallDenseSet<size_t, 8> AnalyzedBundles;
 
  /// Set of the values, which were a part of the analyzed vector nodes.
  SmallPtrSet<const Value *, 32> AnalyzedScalars;
 
  /// Cache of the number of parts for the types and the parts limit.
  mutable SmallDenseMap<std::tuple<Type *, Type *, unsigned>, unsigned>
      NumberOfPartsCache;
 
  /// Values, already been analyzed for mininmal bitwidth and found to be
  /// non-profitable.
  DenseSet<Value *> AnalyzedMinBWVals;
 
  /// A list of values that need to extracted out of the tree.
  /// This list holds pairs of (Internal Scalar : External User). External User
  /// can be nullptr, it means that this Internal Scalar will be used later,
  /// after vectorization.
  UserList ExternalUses;
 
  /// A list of GEPs which can be reaplced by scalar GEPs instead of
  /// extractelement instructions.
  SmallPtrSet<Value *, 4> ExternalUsesAsOriginalScalar;
 
  /// A list of scalar to be extracted without specific user necause of too many
  /// uses.
  SmallPtrSet<Value *, 4> ExternalUsesWithNonUsers;
 
  /// Replacements emitted for the external uses without users, consumed after
  /// the tree vectorization; must not be collected as dead operands of the
  /// erased scalars.
  SmallPtrSet<Value *, 4> ExternalUseReplacements;
 
  /// Values used only by @llvm.assume calls.
  SmallPtrSet<const Value *, 32> EphValues;
 
  /// Holds all of the instructions that we gathered, shuffle instructions and
  /// extractelements.
  SetVector<Instruction *> GatherShuffleExtractSeq;
 
  /// A list of blocks that we are going to CSE.
  DenseSet<BasicBlock *> CSEBlocks;
 
  /// List of hashes of vector of loads, which are known to be non vectorizable.
  DenseSet<size_t> ListOfKnonwnNonVectorizableLoads;
 
  /// Represents a scheduling entity, either ScheduleData, ScheduleCopyableData
  /// or ScheduleBundle. ScheduleData used to gather dependecies for a single
  /// instructions, while ScheduleBundle represents a batch of instructions,
  /// going to be groupped together. ScheduleCopyableData models extra user for
  /// "copyable" instructions.
  class ScheduleEntity {
    friend class ScheduleBundle;
    friend class ScheduleData;
    friend class ScheduleCopyableData;
 
  protected:
    enum class Kind { ScheduleData, ScheduleBundle, ScheduleCopyableData };
    Kind getKind() const { return K; }
    ScheduleEntity(Kind K) : K(K) {}
 
  private:
    /// Used for getting a "good" final ordering of instructions.
    int SchedulingPriority = 0;
    /// True if this instruction (or bundle) is scheduled (or considered as
    /// scheduled in the dry-run).
    bool IsScheduled = false;
    /// The kind of the ScheduleEntity.
    const Kind K = Kind::ScheduleData;
 
  public:
    ScheduleEntity() = delete;
    /// Gets/sets the scheduling priority.
    void setSchedulingPriority(int Priority) { SchedulingPriority = Priority; }
    int getSchedulingPriority() const { return SchedulingPriority; }
    bool isReady() const {
      if (const auto *SD = dyn_cast<ScheduleData>(this))
        return SD->isReady();
      if (const auto *CD = dyn_cast<ScheduleCopyableData>(this))
        return CD->isReady();
      return cast<ScheduleBundle>(this)->isReady();
    }
    /// Returns true if the dependency information has been calculated.
    /// Note that depenendency validity can vary between instructions within
    /// a single bundle.
    bool hasValidDependencies() const {
      if (const auto *SD = dyn_cast<ScheduleData>(this))
        return SD->hasValidDependencies();
      if (const auto *CD = dyn_cast<ScheduleCopyableData>(this))
        return CD->hasValidDependencies();
      return cast<ScheduleBundle>(this)->hasValidDependencies();
    }
    /// Gets the number of unscheduled dependencies.
    int getUnscheduledDeps() const {
      if (const auto *SD = dyn_cast<ScheduleData>(this))
        return SD->getUnscheduledDeps();
      if (const auto *CD = dyn_cast<ScheduleCopyableData>(this))
        return CD->getUnscheduledDeps();
      return cast<ScheduleBundle>(this)->unscheduledDepsInBundle();
    }
    /// Increments the number of unscheduled dependencies.
    int incrementUnscheduledDeps(int Incr) {
      if (auto *SD = dyn_cast<ScheduleData>(this))
        return SD->incrementUnscheduledDeps(Incr);
      return cast<ScheduleCopyableData>(this)->incrementUnscheduledDeps(Incr);
    }
    /// Gets the number of dependencies.
    int getDependencies() const {
      if (const auto *SD = dyn_cast<ScheduleData>(this))
        return SD->getDependencies();
      return cast<ScheduleCopyableData>(this)->getDependencies();
    }
    /// Gets the instruction.
    Instruction *getInst() const {
      if (const auto *SD = dyn_cast<ScheduleData>(this))
        return SD->getInst();
      return cast<ScheduleCopyableData>(this)->getInst();
    }
 
    /// Gets/sets if the bundle is scheduled.
    bool isScheduled() const { return IsScheduled; }
    void setScheduled(bool Scheduled) { IsScheduled = Scheduled; }
 
    static bool classof(const ScheduleEntity *) { return true; }
 
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    void dump(raw_ostream &OS) const {
      if (const auto *SD = dyn_cast<ScheduleData>(this))
        return SD->dump(OS);
      if (const auto *CD = dyn_cast<ScheduleCopyableData>(this))
        return CD->dump(OS);
      return cast<ScheduleBundle>(this)->dump(OS);
    }
 
    LLVM_DUMP_METHOD void dump() const {
      dump(dbgs());
      dbgs() << '\n';
    }
#endif // if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
  };
 
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
  friend inline raw_ostream &operator<<(raw_ostream &OS,
                                        const BoUpSLP::ScheduleEntity &SE) {
    SE.dump(OS);
    return OS;
  }
#endif
 
  /// Contains all scheduling relevant data for an instruction.
  /// A ScheduleData either represents a single instruction or a member of an
  /// instruction bundle (= a group of instructions which is combined into a
  /// vector instruction).
  class ScheduleData final : public ScheduleEntity {
  public:
    // The initial value for the dependency counters. It means that the
    // dependencies are not calculated yet.
    enum { InvalidDeps = -1 };
 
    ScheduleData() : ScheduleEntity(Kind::ScheduleData) {}
    static bool classof(const ScheduleEntity *Entity) {
      return Entity->getKind() == Kind::ScheduleData;
    }
 
    void init(int BlockSchedulingRegionID, Instruction *I) {
      NextLoadStore = nullptr;
      IsScheduled = false;
      SchedulingRegionID = BlockSchedulingRegionID;
      clearDependencies();
      Inst = I;
    }
 
    /// Verify basic self consistency properties
    void verify() {
      if (hasValidDependencies()) {
        assert(UnscheduledDeps <= Dependencies && "invariant");
      } else {
        assert(UnscheduledDeps == Dependencies && "invariant");
      }
 
      if (IsScheduled) {
        assert(hasValidDependencies() && UnscheduledDeps == 0 &&
               "unexpected scheduled state");
      }
    }
 
    /// Returns true if the dependency information has been calculated.
    /// Note that depenendency validity can vary between instructions within
    /// a single bundle.
    bool hasValidDependencies() const { return Dependencies != InvalidDeps; }
 
    /// Returns true if it is ready for scheduling, i.e. it has no more
    /// unscheduled depending instructions/bundles.
    bool isReady() const { return UnscheduledDeps == 0 && !IsScheduled; }
 
    /// Modifies the number of unscheduled dependencies for this instruction,
    /// and returns the number of remaining dependencies for the containing
    /// bundle.
    int incrementUnscheduledDeps(int Incr) {
      assert(hasValidDependencies() &&
             "increment of unscheduled deps would be meaningless");
      UnscheduledDeps += Incr;
      assert(UnscheduledDeps >= 0 &&
             "Expected valid number of unscheduled deps");
      return UnscheduledDeps;
    }
 
    /// Sets the number of unscheduled dependencies to the number of
    /// dependencies.
    void resetUnscheduledDeps() { UnscheduledDeps = Dependencies; }
 
    /// Clears all dependency information.
    void clearDependencies() {
      clearDirectDependencies();
      MemoryDependencies.clear();
      ControlDependencies.clear();
    }
 
    /// Clears all direct dependencies only, except for control and memory
    /// dependencies.
    /// Required for copyable elements to correctly handle control/memory deps
    /// and avoid extra reclaculation of such deps.
    void clearDirectDependencies() {
      Dependencies = InvalidDeps;
      resetUnscheduledDeps();
      IsScheduled = false;
    }
 
    /// Gets the number of unscheduled dependencies.
    int getUnscheduledDeps() const { return UnscheduledDeps; }
    /// Gets the number of dependencies.
    int getDependencies() const { return Dependencies; }
    /// Initializes the number of dependencies.
    void initDependencies() { Dependencies = 0; }
    /// Increments the number of dependencies.
    void incDependencies() { Dependencies++; }
 
    /// Gets scheduling region ID.
    int getSchedulingRegionID() const { return SchedulingRegionID; }
 
    /// Gets the instruction.
    Instruction *getInst() const { return Inst; }
 
    /// Gets the list of memory dependencies.
    ArrayRef<ScheduleData *> getMemoryDependencies() const {
      return MemoryDependencies;
    }
    /// Adds a memory dependency.
    void addMemoryDependency(ScheduleData *Dep) {
      MemoryDependencies.push_back(Dep);
    }
    /// Gets the list of control dependencies.
    ArrayRef<ScheduleData *> getControlDependencies() const {
      return ControlDependencies;
    }
    /// Adds a control dependency.
    void addControlDependency(ScheduleData *Dep) {
      ControlDependencies.push_back(Dep);
    }
    /// Gets/sets the next load/store instruction in the block.
    ScheduleData *getNextLoadStore() const { return NextLoadStore; }
    void setNextLoadStore(ScheduleData *Next) { NextLoadStore = Next; }
 
    void dump(raw_ostream &OS) const { OS << *Inst; }
 
    LLVM_DUMP_METHOD void dump() const {
      dump(dbgs());
      dbgs() << '\n';
    }
 
  private:
    Instruction *Inst = nullptr;
 
    /// Single linked list of all memory instructions (e.g. load, store, call)
    /// in the block - until the end of the scheduling region.
    ScheduleData *NextLoadStore = nullptr;
 
    /// The dependent memory instructions.
    /// This list is derived on demand in calculateDependencies().
    SmallVector<ScheduleData *> MemoryDependencies;
 
    /// List of instructions which this instruction could be control dependent
    /// on.  Allowing such nodes to be scheduled below this one could introduce
    /// a runtime fault which didn't exist in the original program.
    /// ex: this is a load or udiv following a readonly call which inf loops
    SmallVector<ScheduleData *> ControlDependencies;
 
    /// This ScheduleData is in the current scheduling region if this matches
    /// the current SchedulingRegionID of BlockScheduling.
    int SchedulingRegionID = 0;
 
    /// The number of dependencies. Constitutes of the number of users of the
    /// instruction plus the number of dependent memory instructions (if any).
    /// This value is calculated on demand.
    /// If InvalidDeps, the number of dependencies is not calculated yet.
    int Dependencies = InvalidDeps;
 
    /// The number of dependencies minus the number of dependencies of scheduled
    /// instructions. As soon as this is zero, the instruction/bundle gets ready
    /// for scheduling.
    /// Note that this is negative as long as Dependencies is not calculated.
    int UnscheduledDeps = InvalidDeps;
  };
 
#ifndef NDEBUG
  friend inline raw_ostream &operator<<(raw_ostream &OS,
                                        const BoUpSLP::ScheduleData &SD) {
    SD.dump(OS);
    return OS;
  }
#endif
 
  class ScheduleBundle final : public ScheduleEntity {
    /// The schedule data for the instructions in the bundle.
    SmallVector<ScheduleEntity *> Bundle;
    /// True if this bundle is valid.
    bool IsValid = true;
    /// The TreeEntry that this instruction corresponds to.
    TreeEntry *TE = nullptr;
    ScheduleBundle(bool IsValid)
        : ScheduleEntity(Kind::ScheduleBundle), IsValid(IsValid) {}
 
  public:
    ScheduleBundle() : ScheduleEntity(Kind::ScheduleBundle) {}
    static bool classof(const ScheduleEntity *Entity) {
      return Entity->getKind() == Kind::ScheduleBundle;
    }
 
    /// Verify basic self consistency properties
    void verify() const {
      for (const ScheduleEntity *SD : Bundle) {
        if (SD->hasValidDependencies()) {
          assert(SD->getUnscheduledDeps() <= SD->getDependencies() &&
                 "invariant");
        } else {
          assert(SD->getUnscheduledDeps() == SD->getDependencies() &&
                 "invariant");
        }
 
        if (isScheduled()) {
          assert(SD->hasValidDependencies() && SD->getUnscheduledDeps() == 0 &&
                 "unexpected scheduled state");
        }
      }
    }
 
    /// Returns the number of unscheduled dependencies in the bundle.
    int unscheduledDepsInBundle() const {
      assert(*this && "bundle must not be empty");
      int Sum = 0;
      for (const ScheduleEntity *BundleMember : Bundle) {
        if (BundleMember->getUnscheduledDeps() == ScheduleData::InvalidDeps)
          return ScheduleData::InvalidDeps;
        Sum += BundleMember->getUnscheduledDeps();
      }
      return Sum;
    }
 
    /// Returns true if the dependency information has been calculated.
    /// Note that depenendency validity can vary between instructions within
    /// a single bundle.
    bool hasValidDependencies() const {
      return all_of(Bundle, [](const ScheduleEntity *SD) {
        return SD->hasValidDependencies();
      });
    }
 
    /// Returns true if it is ready for scheduling, i.e. it has no more
    /// unscheduled depending instructions/bundles.
    bool isReady() const {
      assert(*this && "bundle must not be empty");
      return unscheduledDepsInBundle() == 0 && !isScheduled();
    }
 
    /// Returns the bundle of scheduling data, associated with the current
    /// instruction.
    ArrayRef<ScheduleEntity *> getBundle() { return Bundle; }
    ArrayRef<const ScheduleEntity *> getBundle() const { return Bundle; }
    /// Adds an instruction to the bundle.
    void add(ScheduleEntity *SD) { Bundle.push_back(SD); }
 
    /// Gets/sets the associated tree entry.
    void setTreeEntry(TreeEntry *TE) { this->TE = TE; }
    TreeEntry *getTreeEntry() const { return TE; }
 
    static ScheduleBundle invalid() { return {false}; }
 
    operator bool() const { return IsValid; }
 
#ifndef NDEBUG
    void dump(raw_ostream &OS) const {
      if (!*this) {
        OS << "[]";
        return;
      }
      OS << '[';
      interleaveComma(Bundle, OS, [&](const ScheduleEntity *SD) {
        if (isa<ScheduleCopyableData>(SD))
          OS << "<Copyable>";
        OS << *SD->getInst();
      });
      OS << ']';
    }
 
    LLVM_DUMP_METHOD void dump() const {
      dump(dbgs());
      dbgs() << '\n';
    }
#endif // NDEBUG
  };
 
#ifndef NDEBUG
  friend inline raw_ostream &operator<<(raw_ostream &OS,
                                        const BoUpSLP::ScheduleBundle &Bundle) {
    Bundle.dump(OS);
    return OS;
  }
#endif
 
  /// Contains all scheduling relevant data for the copyable instruction.
  /// It models the virtual instructions, supposed to replace the original
  /// instructions. E.g., if instruction %0 = load is a part of the bundle [%0,
  /// %1], where %1 = add, then the ScheduleCopyableData models virtual
  /// instruction %virt = add %0, 0.
  class ScheduleCopyableData final : public ScheduleEntity {
    /// The source schedule data for the instruction.
    Instruction *Inst = nullptr;
    /// The edge information for the instruction.
    const EdgeInfo EI;
    /// This ScheduleData is in the current scheduling region if this matches
    /// the current SchedulingRegionID of BlockScheduling.
    int SchedulingRegionID = 0;
    /// Bundle, this data is part of.
    ScheduleBundle &Bundle;
 
  public:
    ScheduleCopyableData(int BlockSchedulingRegionID, Instruction *I,
                         const EdgeInfo &EI, ScheduleBundle &Bundle)
        : ScheduleEntity(Kind::ScheduleCopyableData), Inst(I), EI(EI),
          SchedulingRegionID(BlockSchedulingRegionID), Bundle(Bundle) {}
    static bool classof(const ScheduleEntity *Entity) {
      return Entity->getKind() == Kind::ScheduleCopyableData;
    }
 
    /// Verify basic self consistency properties
    void verify() {
      if (hasValidDependencies()) {
        assert(UnscheduledDeps <= Dependencies && "invariant");
      } else {
        assert(UnscheduledDeps == Dependencies && "invariant");
      }
 
      if (IsScheduled) {
        assert(hasValidDependencies() && UnscheduledDeps == 0 &&
               "unexpected scheduled state");
      }
    }
 
    /// Returns true if the dependency information has been calculated.
    /// Note that depenendency validity can vary between instructions within
    /// a single bundle.
    bool hasValidDependencies() const {
      return Dependencies != ScheduleData::InvalidDeps;
    }
 
    /// Returns true if it is ready for scheduling, i.e. it has no more
    /// unscheduled depending instructions/bundles.
    bool isReady() const { return UnscheduledDeps == 0 && !IsScheduled; }
 
    /// Modifies the number of unscheduled dependencies for this instruction,
    /// and returns the number of remaining dependencies for the containing
    /// bundle.
    int incrementUnscheduledDeps(int Incr) {
      assert(hasValidDependencies() &&
             "increment of unscheduled deps would be meaningless");
      UnscheduledDeps += Incr;
      assert(UnscheduledDeps >= 0 && "invariant");
      return UnscheduledDeps;
    }
 
    /// Sets the number of unscheduled dependencies to the number of
    /// dependencies.
    void resetUnscheduledDeps() { UnscheduledDeps = Dependencies; }
 
    /// Gets the number of unscheduled dependencies.
    int getUnscheduledDeps() const { return UnscheduledDeps; }
    /// Gets the number of dependencies.
    int getDependencies() const { return Dependencies; }
    /// Initializes the number of dependencies.
    void initDependencies() { Dependencies = 0; }
    /// Increments the number of dependencies.
    void incDependencies() { Dependencies++; }
 
    /// Gets scheduling region ID.
    int getSchedulingRegionID() const { return SchedulingRegionID; }
 
    /// Gets the instruction.
    Instruction *getInst() const { return Inst; }
 
    /// Clears all dependency information.
    void clearDependencies() {
      Dependencies = ScheduleData::InvalidDeps;
      UnscheduledDeps = ScheduleData::InvalidDeps;
      IsScheduled = false;
    }
 
    /// Gets the edge information.
    const EdgeInfo &getEdgeInfo() const { return EI; }
 
    /// Gets the bundle.
    ScheduleBundle &getBundle() { return Bundle; }
    const ScheduleBundle &getBundle() const { return Bundle; }
 
#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
    void dump(raw_ostream &OS) const { OS << "[Copyable]" << *getInst(); }
 
    LLVM_DUMP_METHOD void dump() const {
      dump(dbgs());
      dbgs() << '\n';
    }
#endif // !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
 
  private:
    /// true, if it has valid dependency information. These nodes always have
    /// only single dependency.
    int Dependencies = ScheduleData::InvalidDeps;
 
    /// The number of dependencies minus the number of dependencies of scheduled
    /// instructions. As soon as this is zero, the instruction/bundle gets ready
    /// for scheduling.
    /// Note that this is negative as long as Dependencies is not calculated.
    int UnscheduledDeps = ScheduleData::InvalidDeps;
  };
 
#ifndef NDEBUG
  friend inline raw_ostream &
  operator<<(raw_ostream &OS, const BoUpSLP::ScheduleCopyableData &SD) {
    SD.dump(OS);
    return OS;
  }
#endif
 
  friend struct GraphTraits<BoUpSLP *>;
  friend struct DOTGraphTraits<BoUpSLP *>;
 
  /// Contains all scheduling data for a basic block.
  /// It does not schedules instructions, which are not memory read/write
  /// instructions and their operands are either constants, or arguments, or
  /// phis, or instructions from others blocks, or their users are phis or from
  /// the other blocks. The resulting vector instructions can be placed at the
  /// beginning of the basic block without scheduling (if operands does not need
  /// to be scheduled) or at the end of the block (if users are outside of the
  /// block). It allows to save some compile time and memory used by the
  /// compiler.
  /// ScheduleData is assigned for each instruction in between the boundaries of
  /// the tree entry, even for those, which are not part of the graph. It is
  /// required to correctly follow the dependencies between the instructions and
  /// their correct scheduling. The ScheduleData is not allocated for the
  /// instructions, which do not require scheduling, like phis, nodes with
  /// extractelements/insertelements only or nodes with instructions, with
  /// uses/operands outside of the block.
  struct BlockScheduling {
    BlockScheduling(BasicBlock *BB)
        : BB(BB), ChunkSize(BB->size()), ChunkPos(ChunkSize) {}
 
    void clear() {
      ScheduledBundles.clear();
      ScheduledBundlesList.clear();
      ScheduleCopyableDataMap.clear();
      ScheduleCopyableDataMapByInst.clear();
      ScheduleCopyableDataMapByInstUser.clear();
      ScheduleCopyableDataMapByUsers.clear();
      ReadyInsts.clear();
      RecalcCopyableOperandDeps.clear();
      IgnoredMemDeps.clear();
      ScheduleStart = nullptr;
      ScheduleEnd = nullptr;
      FirstLoadStoreInRegion = nullptr;
      LastLoadStoreInRegion = nullptr;
      RegionHasStackSave = false;
 
      // Reduce the maximum schedule region size by the size of the
      // previous scheduling run.
      ScheduleRegionSizeLimit -= ScheduleRegionSize;
      if (ScheduleRegionSizeLimit < MinScheduleRegionSize)
        ScheduleRegionSizeLimit = MinScheduleRegionSize;
      ScheduleRegionSize = 0;
 
      // Make a new scheduling region, i.e. all existing ScheduleData is not
      // in the new region yet.
      ++SchedulingRegionID;
    }
 
    ScheduleData *getScheduleData(Instruction *I) {
      if (!I)
        return nullptr;
      if (BB != I->getParent())
        // Avoid lookup if can't possibly be in map.
        return nullptr;
      ScheduleData *SD = ScheduleDataMap.lookup(I);
      if (SD && isInSchedulingRegion(*SD))
        return SD;
      return nullptr;
    }
 
    ScheduleData *getScheduleData(Value *V) {
      return getScheduleData(dyn_cast<Instruction>(V));
    }
 
    /// Returns the ScheduleCopyableData for the given edge (user tree entry and
    /// operand number) and value.
    ScheduleCopyableData *getScheduleCopyableData(const EdgeInfo &EI,
                                                  const Value *V) const {
      if (ScheduleCopyableDataMap.empty())
        return nullptr;
      auto It = ScheduleCopyableDataMap.find(std::make_pair(EI, V));
      if (It == ScheduleCopyableDataMap.end())
        return nullptr;
      ScheduleCopyableData *SD = It->getSecond().get();
      if (!isInSchedulingRegion(*SD))
        return nullptr;
      return SD;
    }
 
    /// Returns the ScheduleCopyableData for the given user \p User, operand
    /// number and operand \p V.
    SmallVector<ScheduleCopyableData *>
    getScheduleCopyableData(const Value *User, unsigned OperandIdx,
                            const Value *V) {
      if (ScheduleCopyableDataMapByInstUser.empty())
        return {};
      const auto It = ScheduleCopyableDataMapByInstUser.find(
          std::make_pair(std::make_pair(User, OperandIdx), V));
      if (It == ScheduleCopyableDataMapByInstUser.end())
        return {};
      SmallVector<ScheduleCopyableData *> Res;
      for (ScheduleCopyableData *SD : It->getSecond()) {
        if (isInSchedulingRegion(*SD))
          Res.push_back(SD);
      }
      return Res;
    }
 
    /// Returns true if all operands of the given instruction \p User are
    /// replaced by copyable data.
    /// \param User The user instruction.
    /// \param Op The operand, which might be replaced by the copyable data.
    /// \param SLP The SLP tree.
    /// \param NumOps The number of operands used. If the instruction uses the
    /// same operand several times, check for the first use, then the second,
    /// etc.
    bool areAllOperandsReplacedByCopyableData(Instruction *User,
                                              Instruction *Op, BoUpSLP &SLP,
                                              unsigned NumOps) const {
      assert(NumOps > 0 && "No operands");
      if (ScheduleCopyableDataMap.empty())
        return false;
      SmallDenseMap<TreeEntry *, unsigned> PotentiallyReorderedEntriesCount;
      ArrayRef<TreeEntry *> Entries = SLP.getTreeEntries(User);
      if (Entries.empty())
        return false;
      unsigned CurNumOps = 0;
      for (const Use &U : User->operands()) {
        if (U.get() != Op)
          continue;
        ++CurNumOps;
        // Check all tree entries, if they have operands replaced by copyable
        // data.
        for (TreeEntry *TE : Entries) {
          unsigned Inc = 0;
          bool IsNonSchedulableWithParentPhiNode =
              TE->doesNotNeedToSchedule() && TE->UserTreeIndex &&
              TE->UserTreeIndex.UserTE->hasState() &&
              TE->UserTreeIndex.UserTE->State != TreeEntry::SplitVectorize &&
              TE->UserTreeIndex.UserTE->getOpcode() == Instruction::PHI;
          // Count the number of unique phi nodes, which are the parent for
          // parent entry, and exit, if all the unique phis are processed.
          if (IsNonSchedulableWithParentPhiNode) {
            SmallPtrSet<Value *, 4> ParentsUniqueUsers;
            const TreeEntry *ParentTE = TE->UserTreeIndex.UserTE;
            for (Value *V : ParentTE->Scalars) {
              auto *PHI = dyn_cast<PHINode>(V);
              if (!PHI)
                continue;
              if (ParentsUniqueUsers.insert(PHI).second &&
                  is_contained(PHI->incoming_values(), User))
                ++Inc;
            }
          } else {
            Inc = count(TE->Scalars, User);
          }
 
          // Check if the user is commutative.
          // The commutatives are handled later, as their operands can be
          // reordered.
          // Same applies even for non-commutative cmps, because we can invert
          // their predicate potentially and, thus, reorder the operands.
          bool IsCommutativeUser =
              isCommutative(User) &&
              isCommutableOperand(User, User, U.getOperandNo());
          if (!IsCommutativeUser) {
            Instruction *MainOp = TE->getMatchingMainOpOrAltOp(User);
            IsCommutativeUser =
                isCommutative(MainOp, User) &&
                isCommutableOperand(MainOp, User, U.getOperandNo());
          }
          // The commutative user with the same operands can be safely
          // considered as non-commutative, operands reordering does not change
          // the semantics. Same for cmps with the same operands: inverting
          // the predicate does not change the operand columns in this case.
          assert(
              (!IsCommutativeUser ||
               (((isCommutative(User) && isCommutableOperand(User, User, 0) &&
                  isCommutableOperand(User, User, 1)) ||
                 (isCommutative(TE->getMatchingMainOpOrAltOp(User), User) &&
                  isCommutableOperand(TE->getMatchingMainOpOrAltOp(User), User,
                                      0) &&
                  isCommutableOperand(TE->getMatchingMainOpOrAltOp(User), User,
                                      1))))) &&
              "Expected commutative user with 2 first commutable operands");
          bool IsCommutativeWithSameOps =
              IsCommutativeUser && User->getOperand(0) == User->getOperand(1);
          if ((!IsCommutativeUser || IsCommutativeWithSameOps) &&
              (!isa<CmpInst>(User) ||
               User->getOperand(0) == User->getOperand(1))) {
            if (CurNumOps != NumOps)
              continue;
            // A reassociated node flattens the operand chain, so the operand
            // may be placed in any operand column rather than at the
            // instruction's operand number.
            if (TE->hasReassocScalars()) {
              bool ReplacedByCopyable = false;
              for (auto It = find(TE->Scalars, User); It != TE->Scalars.end();
                   It = find(make_range(std::next(It), TE->Scalars.end()),
                             User)) {
                int Lane = std::distance(TE->Scalars.begin(), It);
                for (unsigned OpIdx : seq<unsigned>(TE->getNumOperands()))
                  ReplacedByCopyable |=
                      TE->getOperand(OpIdx)[Lane] == Op &&
                      getScheduleCopyableData(EdgeInfo(TE, OpIdx), Op);
              }
              if (ReplacedByCopyable)
                continue;
              return false;
            }
            EdgeInfo EI(TE, U.getOperandNo());
            if (getScheduleCopyableData(EI, Op))
              continue;
            return false;
          }
          // Only count the occurrence matching this call's NumOps.
          if (CurNumOps != NumOps)
            continue;
          PotentiallyReorderedEntriesCount.try_emplace(TE, 0)
              .first->getSecond() += Inc;
        }
      }
      if (PotentiallyReorderedEntriesCount.empty())
        return true;
      // Check the commutative/cmp entries.
      for (auto &P : PotentiallyReorderedEntriesCount) {
        SmallPtrSet<Value *, 4> ParentsUniqueUsers;
        bool IsNonSchedulableWithParentPhiNode =
            P.first->doesNotNeedToSchedule() && P.first->UserTreeIndex &&
            P.first->UserTreeIndex.UserTE->hasState() &&
            P.first->UserTreeIndex.UserTE->State != TreeEntry::SplitVectorize &&
            P.first->UserTreeIndex.UserTE->getOpcode() == Instruction::PHI;
        auto *It = find(P.first->Scalars, User);
        do {
          assert(It != P.first->Scalars.end() &&
                 "User is not in the tree entry");
          int Lane = std::distance(P.first->Scalars.begin(), It);
          assert(Lane >= 0 && "Lane is not found");
          if (isa<StoreInst, InsertValueInst>(User) &&
              !P.first->ReorderIndices.empty())
            Lane = P.first->ReorderIndices[Lane];
          assert(Lane < static_cast<int>(P.first->Scalars.size()) &&
                 "Couldn't find extract lane");
          // Count the number of unique phi nodes, which are the parent for
          // parent entry, and exit, if all the unique phis are processed.
          if (IsNonSchedulableWithParentPhiNode) {
            const TreeEntry *ParentTE = P.first->UserTreeIndex.UserTE;
            Value *User = ParentTE->Scalars[Lane];
            if (!ParentsUniqueUsers.insert(User).second) {
              It =
                  find(make_range(std::next(It), P.first->Scalars.end()), User);
              continue;
            }
          }
          // Flattened nodes may place an operand in any column; scan all of
          // them so copyable scheduling does not double-count.
          for (unsigned OpIdx :
               seq<unsigned>(P.first->hasReassocScalars()
                                 ? P.first->getNumOperands()
                                 : getNumberOfPotentiallyCommutativeOps(
                                       P.first->getMainOp()))) {
            if (P.first->getOperand(OpIdx)[Lane] == Op &&
                getScheduleCopyableData(EdgeInfo(P.first, OpIdx), Op))
              --P.getSecond();
          }
          // If parent node is schedulable, it will be handled correctly.
          It = find(make_range(std::next(It), P.first->Scalars.end()), User);
        } while (It != P.first->Scalars.end());
      }
      return all_of(PotentiallyReorderedEntriesCount,
                    [&](const std::pair<const TreeEntry *, unsigned> &P) {
                      return P.second == NumOps - 1;
                    });
    }
 
    SmallVector<ScheduleCopyableData *>
    getScheduleCopyableData(const Instruction *I) const {
      if (ScheduleCopyableDataMapByInst.empty())
        return {};
      const auto It = ScheduleCopyableDataMapByInst.find(I);
      if (It == ScheduleCopyableDataMapByInst.end())
        return {};
      SmallVector<ScheduleCopyableData *> Res;
      for (ScheduleCopyableData *SD : It->getSecond()) {
        if (isInSchedulingRegion(*SD))
          Res.push_back(SD);
      }
      return Res;
    }
 
    SmallVector<ScheduleCopyableData *>
    getScheduleCopyableDataUsers(const Instruction *User) const {
      if (ScheduleCopyableDataMapByUsers.empty())
        return {};
      const auto It = ScheduleCopyableDataMapByUsers.find(User);
      if (It == ScheduleCopyableDataMapByUsers.end())
        return {};
      SmallVector<ScheduleCopyableData *> Res;
      for (ScheduleCopyableData *SD : It->getSecond()) {
        if (isInSchedulingRegion(*SD))
          Res.push_back(SD);
      }
      return Res;
    }
 
    /// Reordering \p TE permutes its operand columns and may move an operand
    /// between the edges covered and not covered by copyable scheduling
    /// data, making the computed dependency counts stale. Mark the schedule
    /// data of \p TE's copyable-modeled operands for recalculation at the
    /// next bundle scheduling.
    void markCopyableDepsForRecalc(const TreeEntry &TE) {
      for (unsigned OpIdx : seq<unsigned>(TE.getNumOperands()))
        for (Value *V : TE.getOperand(OpIdx))
          if (auto *I = dyn_cast<Instruction>(V))
            if (ScheduleData *SD = getScheduleData(I);
                SD && !getScheduleCopyableData(I).empty())
              RecalcCopyableOperandDeps.insert(SD);
    }
 
    ScheduleCopyableData &addScheduleCopyableData(const EdgeInfo &EI,
                                                  Instruction *I,
                                                  int SchedulingRegionID,
                                                  ScheduleBundle &Bundle) {
      assert(!getScheduleCopyableData(EI, I) && "already in the map");
      ScheduleCopyableData *CD =
          ScheduleCopyableDataMap
              .try_emplace(std::make_pair(EI, I),
                           std::make_unique<ScheduleCopyableData>(
                               SchedulingRegionID, I, EI, Bundle))
              .first->getSecond()
              .get();
      ScheduleCopyableDataMapByInst[I].push_back(CD);
      if (EI.UserTE) {
        ArrayRef<Value *> Op = EI.UserTE->getOperand(EI.EdgeIdx);
        const auto *It = find(Op, I);
        assert(It != Op.end() && "Lane not set");
        SmallPtrSet<Instruction *, 4> Visited;
        do {
          int Lane = std::distance(Op.begin(), It);
          assert(Lane >= 0 && "Lane not set");
          if (isa<StoreInst, InsertValueInst>(EI.UserTE->Scalars[Lane]) &&
              !EI.UserTE->ReorderIndices.empty())
            Lane = EI.UserTE->ReorderIndices[Lane];
          assert(Lane < static_cast<int>(EI.UserTE->Scalars.size()) &&
                 "Couldn't find extract lane");
          auto *In = cast<Instruction>(EI.UserTE->Scalars[Lane]);
          if (!Visited.insert(In).second) {
            It = find(make_range(std::next(It), Op.end()), I);
            continue;
          }
          ScheduleCopyableDataMapByInstUser
              .try_emplace(std::make_pair(std::make_pair(In, EI.EdgeIdx), I))
              .first->getSecond()
              .push_back(CD);
          ScheduleCopyableDataMapByUsers.try_emplace(I)
              .first->getSecond()
              .insert(CD);
          // Remove extra deps for users, becoming non-immediate users of the
          // instruction. It may happen, if the chain of same copyable elements
          // appears in the tree.
          if (In == I) {
            EdgeInfo UserEI = EI.UserTE->UserTreeIndex;
            if (ScheduleCopyableData *UserCD =
                    getScheduleCopyableData(UserEI, In))
              ScheduleCopyableDataMapByUsers[I].remove(UserCD);
          }
          It = find(make_range(std::next(It), Op.end()), I);
        } while (It != Op.end());
      } else {
        ScheduleCopyableDataMapByUsers.try_emplace(I).first->getSecond().insert(
            CD);
      }
      return *CD;
    }
 
    ArrayRef<ScheduleBundle *> getScheduleBundles(Value *V) const {
      auto *I = dyn_cast<Instruction>(V);
      if (!I)
        return {};
      auto It = ScheduledBundles.find(I);
      if (It == ScheduledBundles.end())
        return {};
      return It->getSecond();
    }
 
    /// Returns true if the entity is in the scheduling region.
    bool isInSchedulingRegion(const ScheduleEntity &SD) const {
      if (const auto *Data = dyn_cast<ScheduleData>(&SD))
        return Data->getSchedulingRegionID() == SchedulingRegionID;
      if (const auto *CD = dyn_cast<ScheduleCopyableData>(&SD))
        return CD->getSchedulingRegionID() == SchedulingRegionID;
      return all_of(cast<ScheduleBundle>(SD).getBundle(),
                    [&](const ScheduleEntity *BundleMember) {
                      return isInSchedulingRegion(*BundleMember);
                    });
    }
 
    /// Marks an instruction as scheduled and puts all dependent ready
    /// instructions into the ready-list.
    template <typename ReadyListType>
    void schedule(const BoUpSLP &R, const InstructionsState &S,
                  const EdgeInfo &EI, ScheduleEntity *Data,
                  ReadyListType &ReadyList) {
      auto ProcessBundleMember = [&](ScheduleEntity *BundleMember,
                                     ArrayRef<ScheduleBundle *> Bundles) {
        // Handle the def-use chain dependencies.
 
        // Decrement the unscheduled counter and insert to ready list if ready.
        auto DecrUnsched = [&](auto *Data, bool IsControl = false) {
          if ((IsControl || Data->hasValidDependencies()) &&
              Data->incrementUnscheduledDeps(-1) == 0) {
            // There are no more unscheduled dependencies after
            // decrementing, so we can put the dependent instruction
            // into the ready list.
            SmallVector<ScheduleBundle *, 1> CopyableBundle;
            ArrayRef<ScheduleBundle *> Bundles;
            if (auto *CD = dyn_cast<ScheduleCopyableData>(Data)) {
              CopyableBundle.push_back(&CD->getBundle());
              Bundles = CopyableBundle;
            } else {
              Bundles = getScheduleBundles(Data->getInst());
            }
            if (!Bundles.empty()) {
              for (ScheduleBundle *Bundle : Bundles) {
                if (Bundle->unscheduledDepsInBundle() == 0) {
                  assert(!Bundle->isScheduled() &&
                         "already scheduled bundle gets ready");
                  ReadyList.insert(Bundle);
                  LLVM_DEBUG(dbgs()
                             << "SLP:    gets ready: " << *Bundle << "\n");
                }
              }
              return;
            }
            assert(!Data->isScheduled() &&
                   "already scheduled bundle gets ready");
            assert(!isa<ScheduleCopyableData>(Data) &&
                   "Expected non-copyable data");
            ReadyList.insert(Data);
            LLVM_DEBUG(dbgs() << "SLP:    gets ready: " << *Data << "\n");
          }
        };
 
        auto DecrUnschedForInst = [&](Instruction *User, unsigned OpIdx,
                                      Instruction *I) {
          if (!ScheduleCopyableDataMap.empty()) {
            SmallVector<ScheduleCopyableData *> CopyableData =
                getScheduleCopyableData(User, OpIdx, I);
            bool ReleasedAsCopyable = false;
            for (ScheduleCopyableData *CD : CopyableData) {
              // Copyable elements modeled on a copyable user lane depend on
              // the user's copyable scheduling data, not on the user itself,
              // and are released when that copyable data is scheduled. The
              // user's own schedule data still carries the def-use dependency
              // in this case, so it must be released below.
              if (CD->getEdgeInfo().UserTE->isCopyableElement(User))
                continue;
              DecrUnsched(CD, /*IsControl=*/false);
              ReleasedAsCopyable = true;
            }
            if (ReleasedAsCopyable)
              return;
          }
          if (ScheduleData *OpSD = getScheduleData(I))
            DecrUnsched(OpSD, /*IsControl=*/false);
        };
 
        // If BundleMember is a vector bundle, its operands may have been
        // reordered during buildTree(). We therefore need to get its operands
        // through the TreeEntry.
        if (!Bundles.empty()) {
          auto *In = BundleMember->getInst();
          // Count uses of each instruction operand.
          SmallDenseMap<const Instruction *, unsigned> OperandsUses;
          unsigned TotalOpCount = 0;
          if (isa<ScheduleCopyableData>(BundleMember)) {
            // Copyable data is used only once (uses itself).
            TotalOpCount = OperandsUses[In] = 1;
          } else {
            for (const Use &U : In->operands()) {
              if (auto *I = dyn_cast<Instruction>(U.get())) {
                auto Res = OperandsUses.try_emplace(I, 0);
                unsigned ExtraDeps = 1;
                // Count all expanded operands in the binops.
                for (ScheduleBundle *Bundle : Bundles) {
                  if (const TreeEntry *TE = Bundle->getTreeEntry()) {
                    if (TE->isExpandedBinOp(In))
                      ++ExtraDeps;
                  } else if (S.isExpandedBinOp(In)) {
                    ++ExtraDeps;
                  }
                }
                Res.first->getSecond() += ExtraDeps;
                TotalOpCount += ExtraDeps;
              }
            }
          }
          // Tracks whether the bundle member instruction itself shows up in
          // some operand column of its node (only copyable elements modeled
          // through their own operands, like absorbed fmuls, do not).
          bool FoundInOpColumns = false;
          // Decrement the unscheduled counter and insert to ready list if
          // ready.
          auto DecrUnschedForInst =
              [&](Instruction *I, TreeEntry *UserTE, unsigned OpIdx,
                  SmallDenseSet<std::pair<const ScheduleEntity *, unsigned>>
                      &Checked,
                  bool IsExpandedOperand = false,
                  bool CopyableDepsOnly = false) {
                if (!ScheduleCopyableDataMap.empty()) {
                  const EdgeInfo EI = {UserTE, OpIdx};
                  if (ScheduleCopyableData *CD =
                          getScheduleCopyableData(EI, I)) {
                    if (!Checked.insert(std::make_pair(CD, OpIdx)).second)
                      return;
                    DecrUnsched(CD, /*IsControl=*/false);
                    return;
                  }
                }
                if (CopyableDepsOnly)
                  return;
                auto It = OperandsUses.find(I);
                if (It == OperandsUses.end()) {
                  // Column value may be a peeled intermediate, not a direct
                  // operand of In; its deps are released when it is scheduled.
                  LLVM_DEBUG(dbgs() << "SLP:   operand " << *I
                                    << " not modeled as a direct operand of "
                                    << *In << ", skipping.\n");
                  return;
                }
                if (It->second > 0) {
                  if (ScheduleData *OpSD = getScheduleData(I)) {
                    if (!IsExpandedOperand &&
                        !Checked.insert(std::make_pair(OpSD, OpIdx)).second)
                      return;
                    --It->getSecond();
                    assert(TotalOpCount > 0 && "No more operands to decrement");
                    --TotalOpCount;
                    DecrUnsched(OpSD, /*IsControl=*/false);
                  } else {
                    --It->getSecond();
                    assert(TotalOpCount > 0 && "No more operands to decrement");
                    --TotalOpCount;
                  }
                }
              };
 
          SmallDenseSet<std::pair<const ScheduleEntity *, unsigned>> Checked;
          for (ScheduleBundle *Bundle : Bundles) {
            if (ScheduleCopyableDataMap.empty() && TotalOpCount == 0)
              break;
            SmallPtrSet<Value *, 4> ParentsUniqueUsers;
            // Need to search for the lane since the tree entry can be
            // reordered.
            auto *It = find(Bundle->getTreeEntry()->Scalars, In);
            bool IsNonSchedulableWithParentPhiNode =
                Bundle->getTreeEntry()->doesNotNeedToSchedule() &&
                Bundle->getTreeEntry()->UserTreeIndex &&
                Bundle->getTreeEntry()->UserTreeIndex.UserTE->hasState() &&
                Bundle->getTreeEntry()->UserTreeIndex.UserTE->State !=
                    TreeEntry::SplitVectorize &&
                Bundle->getTreeEntry()->UserTreeIndex.UserTE->getOpcode() ==
                    Instruction::PHI;
            do {
              int Lane =
                  std::distance(Bundle->getTreeEntry()->Scalars.begin(), It);
              assert(Lane >= 0 && "Lane not set");
              if (isa<StoreInst, InsertValueInst>(In) &&
                  !Bundle->getTreeEntry()->ReorderIndices.empty())
                Lane = Bundle->getTreeEntry()->ReorderIndices[Lane];
              assert(Lane < static_cast<int>(
                                Bundle->getTreeEntry()->Scalars.size()) &&
                     "Couldn't find extract lane");
 
              // Since vectorization tree is being built recursively this
              // assertion ensures that the tree entry has all operands set
              // before reaching this code. Couple of exceptions known at the
              // moment are extracts where their second (immediate) operand is
              // not added. Since immediates do not affect scheduler behavior
              // this is considered okay.
              assert(
                  In &&
                  (isa<ExtractValueInst, ExtractElementInst, CallBase>(In) ||
                   In->getNumOperands() ==
                       Bundle->getTreeEntry()->getNumOperands() ||
                   (isa<ZExtInst>(In) && Bundle->getTreeEntry()->getOpcode() ==
                                             Instruction::Select) ||
                   Bundle->getTreeEntry()->isCopyableElement(In) ||
                   Bundle->getTreeEntry()->hasReassocScalars()) &&
                  "Missed TreeEntry operands?");
 
              // Count the number of unique phi nodes, which are the parent
              // entry, and handle the non-copyable deps only on the first lane
              // for each such phi. Copyable deps are counted per operand column
              // lane and are released on every lane.
              bool CopyableDepsOnly =
                  IsNonSchedulableWithParentPhiNode &&
                  !ParentsUniqueUsers
                       .insert(Bundle->getTreeEntry()
                                   ->UserTreeIndex.UserTE->Scalars[Lane])
                       .second;
 
              // A blended-load operand node is the synthetic blend mask, not an
              // IR operand of the load. Use the real pointer operand for
              // scheduling so the def-use counters stay balanced; the mask is
              // available earlier through the pointer's select.
              bool IsBlended = Bundle->getTreeEntry()->State ==
                               TreeEntry::BlendedLoadVectorize;
              for (unsigned OpIdx :
                   seq<unsigned>(Bundle->getTreeEntry()->getNumOperands()))
                if (auto *I = dyn_cast<Instruction>(
                        IsBlended ? In->getOperand(OpIdx)
                                  : Bundle->getTreeEntry()->getOperand(
                                        OpIdx)[Lane])) {
                  FoundInOpColumns |= (I == In) && !CopyableDepsOnly;
                  LLVM_DEBUG(dbgs() << "SLP:   check for readiness (def): "
                                    << *I << "\n");
                  DecrUnschedForInst(
                      I, Bundle->getTreeEntry(), OpIdx, Checked,
                      Bundle->getTreeEntry()->isExpandedOperand(In, OpIdx),
                      /*CopyableDepsOnly=*/CopyableDepsOnly);
                }
              // If parent node is schedulable, it will be handled correctly.
              if (Bundle->getTreeEntry()->isCopyableElement(In))
                break;
              It = std::find(std::next(It),
                             Bundle->getTreeEntry()->Scalars.end(), In);
            } while (It != Bundle->getTreeEntry()->Scalars.end());
          }
          // A copyable element absorbed into its user modeling (e.g. a
          // copyable fmul turned into fmuladd(a, b, -0.0)) does not appear in
          // the operand columns of its own node, so the scan above never
          // releases the schedule data of the copyable instruction itself.
          // Release it here to keep the unscheduled-deps counters balanced,
          // consuming its self-use count so the reassociated-operand release
          // below cannot release the same schedule data twice.
          if (isa<ScheduleCopyableData>(BundleMember) && !FoundInOpColumns) {
            auto UseIt = OperandsUses.find(In);
            if (UseIt != OperandsUses.end() && UseIt->second > 0) {
              --UseIt->getSecond();
              --TotalOpCount;
            }
            if (ScheduleData *OpSD = getScheduleData(In))
              DecrUnsched(OpSD, /*IsControl=*/false);
          }
          // Vector intrinsics may keep some arguments scalar (e.g. the
          // exponent of llvm.powi). Such scalar arguments are not modeled as
          // tree-entry operands, so the per-lane loop above never releases the
          // dependency that calculateDependencies() registered for the
          // definition feeding such an argument. Release it here to keep the
          // unscheduled-deps counters balanced; otherwise the operand's bundle
          // may never become ready and scheduling would assert.
          if (TotalOpCount > 0) {
            if (auto *CI = dyn_cast<CallInst>(In)) {
              Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, R.TLI);
              for (unsigned ArgIdx : seq<unsigned>(CI->arg_size())) {
                if (!isVectorIntrinsicWithScalarOpAtArg(ID, ArgIdx, R.TTI))
                  continue;
                auto *OpI = dyn_cast<Instruction>(CI->getArgOperand(ArgIdx));
                if (!OpI)
                  continue;
                auto UseIt = OperandsUses.find(OpI);
                if (UseIt == OperandsUses.end() || UseIt->second == 0)
                  continue;
                --UseIt->getSecond();
                --TotalOpCount;
                if (ScheduleData *OpSD = getScheduleData(OpI)) {
                  LLVM_DEBUG(dbgs()
                             << "SLP:   check for readiness (scalar arg): "
                             << *OpI << "\n");
                  DecrUnsched(OpSD, /*IsControl=*/false);
                }
              }
            }
            // Peeled intermediates stay as direct operands but drop out of
            // operand columns; release their scheduling deps here.
            for (const ScheduleBundle *Bundle : Bundles) {
              if (TotalOpCount == 0)
                break;
              TreeEntry *TE = Bundle->getTreeEntry();
              if (!TE->hasReassocScalars())
                continue;
              for (Value *V : TE->getReassocScalars()) {
                auto *OpI = dyn_cast<Instruction>(V);
                if (!OpI)
                  continue;
                auto UseIt = OperandsUses.find(OpI);
                if (UseIt == OperandsUses.end() || UseIt->second == 0)
                  continue;
                LLVM_DEBUG(dbgs() << "SLP:   check for readiness "
                                     "(reassociated operand): "
                                  << *OpI << "\n");
                // Copyable deps may live on per-edge ScheduleCopyableData.
                bool ReleasedAsCopyable = false;
                if (!ScheduleCopyableDataMap.empty()) {
                  for (const Use &U : In->operands()) {
                    if (U.get() != OpI)
                      continue;
                    for (ScheduleCopyableData *CD :
                         getScheduleCopyableData(In, U.getOperandNo(), OpI)) {
                      // Deps of reassoc scalars modeled as copyable tree
                      // operands are released by the operand scan above;
                      // release each remaining dep only once.
                      if (Checked.insert(std::make_pair(CD, U.getOperandNo()))
                              .second)
                        DecrUnsched(CD, /*IsControl=*/false);
                    }
                  }
                  // The dep is released through copyable data only if this
                  // very entry models the scalar as a copyable operand on one
                  // of its edges, mirroring the dependency calculation;
                  // copyable data on some other entry's edge does not cover
                  // the dep registered for this entry.
                  for (auto It = find(TE->Scalars, In);
                       It != TE->Scalars.end() && !ReleasedAsCopyable;
                       It = find(make_range(std::next(It), TE->Scalars.end()),
                                 In)) {
                    int Lane = std::distance(TE->Scalars.begin(), It);
                    for (unsigned OpIdx : seq<unsigned>(TE->getNumOperands()))
                      ReleasedAsCopyable |=
                          TE->getOperand(OpIdx)[Lane] == OpI &&
                          getScheduleCopyableData(EdgeInfo(TE, OpIdx), OpI);
                  }
                }
                if (!ReleasedAsCopyable) {
                  if (ScheduleData *OpSD = getScheduleData(OpI))
                    for (unsigned I = 0, E = UseIt->second; I != E; ++I)
                      DecrUnsched(OpSD, /*IsControl=*/false);
                }
                TotalOpCount -= UseIt->second;
                UseIt->second = 0;
              }
            }
          }
        } else {
          // If BundleMember is a stand-alone instruction, no operand reordering
          // has taken place, so we directly access its operands.
          for (Use &U : BundleMember->getInst()->operands()) {
            if (auto *I = dyn_cast<Instruction>(U.get())) {
              LLVM_DEBUG(dbgs()
                         << "SLP:   check for readiness (def): " << *I << "\n");
              DecrUnschedForInst(BundleMember->getInst(), U.getOperandNo(), I);
            }
          }
        }
        // Handle the memory dependencies.
        auto *SD = dyn_cast<ScheduleData>(BundleMember);
        if (!SD)
          return;
        SmallPtrSet<const ScheduleData *, 4> VisitedMemory;
        for (ScheduleData *MemoryDep : SD->getMemoryDependencies()) {
          if (!VisitedMemory.insert(MemoryDep).second)
            continue;
          // There are no more unscheduled dependencies after decrementing,
          // so we can put the dependent instruction into the ready list.
          LLVM_DEBUG(dbgs() << "SLP:   check for readiness (mem): "
                            << *MemoryDep << "\n");
          DecrUnsched(MemoryDep);
        }
        // Handle the control dependencies.
        SmallPtrSet<const ScheduleData *, 4> VisitedControl;
        for (ScheduleData *Dep : SD->getControlDependencies()) {
          if (!VisitedControl.insert(Dep).second)
            continue;
          // There are no more unscheduled dependencies after decrementing,
          // so we can put the dependent instruction into the ready list.
          LLVM_DEBUG(dbgs()
                     << "SLP:   check for readiness (ctrl): " << *Dep << "\n");
          DecrUnsched(Dep, /*IsControl=*/true);
        }
      };
      if (auto *SD = dyn_cast<ScheduleData>(Data)) {
        SD->setScheduled(/*Scheduled=*/true);
        LLVM_DEBUG(dbgs() << "SLP:   schedule " << *SD << "\n");
        SmallVector<std::unique_ptr<ScheduleBundle>> PseudoBundles;
        SmallVector<ScheduleBundle *> Bundles;
        Instruction *In = SD->getInst();
        ArrayRef<TreeEntry *> Entries = R.getTreeEntries(In);
        if (!Entries.empty()) {
          for (TreeEntry *TE : Entries) {
            if (!isa<ExtractValueInst, ExtractElementInst, CallBase>(In) &&
                In->getNumOperands() != TE->getNumOperands() &&
                !TE->hasReassocScalars())
              continue;
            auto &BundlePtr =
                PseudoBundles.emplace_back(std::make_unique<ScheduleBundle>());
            BundlePtr->setTreeEntry(TE);
            BundlePtr->add(SD);
            Bundles.push_back(BundlePtr.get());
          }
        }
        ProcessBundleMember(SD, Bundles);
      } else {
        ScheduleBundle &Bundle = *cast<ScheduleBundle>(Data);
        Bundle.setScheduled(/*Scheduled=*/true);
        LLVM_DEBUG(dbgs() << "SLP:   schedule " << Bundle << "\n");
        auto AreAllBundlesScheduled =
            [&](const ScheduleEntity *SD,
                ArrayRef<ScheduleBundle *> SDBundles) {
              if (isa<ScheduleCopyableData>(SD))
                return true;
              return !SDBundles.empty() &&
                     all_of(SDBundles, [&](const ScheduleBundle *SDBundle) {
                       return SDBundle->isScheduled();
                     });
            };
        for (ScheduleEntity *SD : Bundle.getBundle()) {
          ArrayRef<ScheduleBundle *> SDBundles;
          if (!isa<ScheduleCopyableData>(SD))
            SDBundles = getScheduleBundles(SD->getInst());
          if (!AreAllBundlesScheduled(SD, SDBundles))
            continue;
          SD->setScheduled(/*Scheduled=*/true);
          Instruction *In = SD->getInst();
          // The instruction may also belong to tree entries that do not need
          // scheduling (e.g. all their values are used outside the block), so
          // no schedule bundle is registered for them. Such an entry can still
          // model one of this instruction's operands as a copyable element, or
          // model the instruction itself as an expanded binop, registered on
          // that non-scheduled parent edge. That dependency would never be
          // decremented when the instruction is scheduled through a different
          // bundle, leaving the operand's bundle permanently unscheduled and
          // tripping the unscheduled-deps assertion. Add pseudo-bundles for
          // these missing tree entries, so their operand dependencies are
          // decremented here as well. Real operand dependencies are protected
          // against double counting by the per-operand use counter.
          if (isa<ScheduleCopyableData>(SD) ||
              (ScheduleCopyableDataMap.empty() &&
               none_of(R.getTreeEntries(In), [&](const TreeEntry *TE) {
                 return TE->isExpandedBinOp(In);
               }))) {
            ProcessBundleMember(SD, isa<ScheduleCopyableData>(SD) ? &Bundle
                                                                  : SDBundles);
            continue;
          }
          SmallVector<std::unique_ptr<ScheduleBundle>> PseudoBundles;
          SmallVector<ScheduleBundle *> AllBundles(SDBundles.begin(),
                                                   SDBundles.end());
          for (TreeEntry *TE : R.getTreeEntries(In)) {
            if (TE->isCopyableElement(In))
              continue;
            if (!isa<ExtractValueInst, ExtractElementInst, CallBase>(In) &&
                In->getNumOperands() != TE->getNumOperands() &&
                !TE->hasReassocScalars())
              continue;
            if (any_of(SDBundles, [&](const ScheduleBundle *SDBundle) {
                  return SDBundle->getTreeEntry() == TE;
                }))
              continue;
            ScheduleBundle &PseudoBundle =
                *PseudoBundles.emplace_back(std::make_unique<ScheduleBundle>());
            PseudoBundle.setTreeEntry(TE);
            PseudoBundle.add(SD);
            AllBundles.push_back(&PseudoBundle);
          }
          ProcessBundleMember(SD, AllBundles);
        }
      }
    }
 
    /// Verify basic self consistency properties of the data structure.
    void verify() {
      if (!ScheduleStart)
        return;
 
      assert(ScheduleStart->getParent() == ScheduleEnd->getParent() &&
             ScheduleStart->comesBefore(ScheduleEnd) &&
             "Not a valid scheduling region?");
 
      for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
        ArrayRef<ScheduleBundle *> Bundles = getScheduleBundles(I);
        if (!Bundles.empty()) {
          for (ScheduleBundle *Bundle : Bundles) {
            assert(isInSchedulingRegion(*Bundle) &&
                   "primary schedule data not in window?");
            Bundle->verify();
          }
          continue;
        }
        auto *SD = getScheduleData(I);
        if (!SD)
          continue;
        assert(isInSchedulingRegion(*SD) &&
               "primary schedule data not in window?");
        SD->verify();
      }
 
      assert(all_of(ReadyInsts,
                    [](const ScheduleEntity *Bundle) {
                      return Bundle->isReady();
                    }) &&
             "item in ready list not ready?");
    }
 
    /// Put all instructions into the ReadyList which are ready for scheduling.
    template <typename ReadyListType>
    void initialFillReadyList(ReadyListType &ReadyList) {
      SmallPtrSet<ScheduleBundle *, 16> Visited;
      for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
        ScheduleData *SD = getScheduleData(I);
        if (SD && SD->hasValidDependencies() && SD->isReady()) {
          if (ArrayRef<ScheduleBundle *> Bundles = getScheduleBundles(I);
              !Bundles.empty()) {
            for (ScheduleBundle *Bundle : Bundles) {
              if (!Visited.insert(Bundle).second)
                continue;
              if (Bundle->hasValidDependencies() && Bundle->isReady()) {
                ReadyList.insert(Bundle);
                LLVM_DEBUG(dbgs() << "SLP:    initially in ready list: "
                                  << *Bundle << "\n");
              }
            }
            continue;
          }
          ReadyList.insert(SD);
          LLVM_DEBUG(dbgs()
                     << "SLP:    initially in ready list: " << *SD << "\n");
        }
      }
    }
 
    /// Build a bundle from the ScheduleData nodes corresponding to the
    /// scalar instruction for each lane.
    /// \param VL The list of scalar instructions.
    /// \param S The state of the instructions.
    /// \param EI The edge in the SLP graph or the user node/operand number.
    ScheduleBundle &buildBundle(ArrayRef<Value *> VL,
                                const InstructionsState &S, const EdgeInfo &EI);
 
    /// Checks if a bundle of instructions can be scheduled, i.e. has no
    /// cyclic dependencies. This is only a dry-run, no instructions are
    /// actually moved at this stage.
    /// \returns the scheduling bundle. The returned Optional value is not
    /// std::nullopt if \p VL is allowed to be scheduled.
    std::optional<ScheduleBundle *>
    tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP,
                      const InstructionsState &S, const EdgeInfo &EI);
 
    /// Allocates schedule data chunk.
    ScheduleData *allocateScheduleDataChunks();
 
    /// Extends the scheduling region so that V is inside the region.
    /// \returns true if the region size is within the limit.
    bool extendSchedulingRegion(Value *V, const InstructionsState &S);
 
    /// Initialize the ScheduleData structures for new instructions in the
    /// scheduling region.
    void initScheduleData(Instruction *FromI, Instruction *ToI,
                          ScheduleData *PrevLoadStore,
                          ScheduleData *NextLoadStore);
 
    /// Updates the dependency information of a bundle and of all instructions/
    /// bundles which depend on the original bundle.
    void calculateDependencies(ScheduleBundle &Bundle, bool InsertInReadyList,
                               BoUpSLP *SLP,
                               const SmallPtrSetImpl<Value *> &ExpandedOps,
                               ArrayRef<ScheduleData *> ControlDeps = {});
 
    /// Sets all instruction in the scheduling region to un-scheduled.
    void resetSchedule();
 
    BasicBlock *BB;
 
    /// Simple memory allocation for ScheduleData.
    SmallVector<std::unique_ptr<ScheduleData[]>> ScheduleDataChunks;
 
    /// The size of a ScheduleData array in ScheduleDataChunks.
    int ChunkSize;
 
    /// The allocator position in the current chunk, which is the last entry
    /// of ScheduleDataChunks.
    int ChunkPos;
 
    /// Attaches ScheduleData to Instruction.
    /// Note that the mapping survives during all vectorization iterations, i.e.
    /// ScheduleData structures are recycled.
    SmallDenseMap<Instruction *, ScheduleData *> ScheduleDataMap;
 
    /// Attaches ScheduleCopyableData to EdgeInfo (UserTreeEntry + operand
    /// number) and the operand instruction, represented as copyable element.
    SmallDenseMap<std::pair<EdgeInfo, const Value *>,
                  std::unique_ptr<ScheduleCopyableData>>
        ScheduleCopyableDataMap;
 
    /// Represents mapping between instruction and all related
    /// ScheduleCopyableData (for all uses in the tree, represenedt as copyable
    /// element). The SLP tree may contain several representations of the same
    /// instruction.
    SmallDenseMap<const Instruction *, SmallVector<ScheduleCopyableData *>>
        ScheduleCopyableDataMapByInst;
 
    /// Represents mapping between user value and operand number, the operand
    /// value and all related ScheduleCopyableData. The relation is 1:n, because
    /// the same user may refernce the same operand in different tree entries
    /// and the operand may be modelled by the different copyable data element.
    SmallDenseMap<std::pair<std::pair<const Value *, unsigned>, const Value *>,
                  SmallVector<ScheduleCopyableData *>>
        ScheduleCopyableDataMapByInstUser;
 
    /// Represents mapping between instruction and all related
    /// ScheduleCopyableData. It represents the mapping between the actual
    /// instruction and the last copyable data element in the chain. E.g., if
    /// the graph models the following instructions:
    /// %0 = non-add instruction ...
    /// ...
    /// %4 = add %3, 1
    /// %5 = add %4, 1
    /// %6 = insertelement poison, %0, 0
    /// %7 = insertelement %6, %5, 1
    /// And the graph is modeled as:
    /// [%5, %0] -> [%4, copyable %0 <0> ] -> [%3, copyable %0 <1> ]
    ///          -> [1, 0]                 -> [%1, 0]
    ///
    /// this map will map %0 only to the copyable element <1>, which is the last
    /// user (direct user of the actual instruction). <0> uses <1>, so <1> will
    /// keep the map to <0>, not the %0.
    SmallDenseMap<const Instruction *,
                  SmallSetVector<ScheduleCopyableData *, 4>>
        ScheduleCopyableDataMapByUsers;
 
    /// Attaches ScheduleBundle to Instruction.
    SmallDenseMap<Instruction *, SmallVector<ScheduleBundle *>>
        ScheduledBundles;
    /// The list of ScheduleBundles.
    SmallVector<std::unique_ptr<ScheduleBundle>> ScheduledBundlesList;
 
    /// The ready-list for scheduling (only used for the dry-run).
    SetVector<ScheduleEntity *> ReadyInsts;
 
    /// The first instruction of the scheduling region.
    Instruction *ScheduleStart = nullptr;
 
    /// The first instruction _after_ the scheduling region.
    Instruction *ScheduleEnd = nullptr;
 
    /// The first memory accessing instruction in the scheduling region
    /// (can be null).
    ScheduleData *FirstLoadStoreInRegion = nullptr;
 
    /// The last memory accessing instruction in the scheduling region
    /// (can be null).
    ScheduleData *LastLoadStoreInRegion = nullptr;
 
    /// Is there an llvm.stacksave or llvm.stackrestore in the scheduling
    /// region?  Used to optimize the dependence calculation for the
    /// common case where there isn't.
    bool RegionHasStackSave = false;
 
    /// The current size of the scheduling region.
    int ScheduleRegionSize = 0;
 
    /// The maximum size allowed for the scheduling region.
    int ScheduleRegionSizeLimit = ScheduleRegionSizeBudget;
 
    /// Operands that are modeled as copyable elements in a previously built
    /// vectorized node and that are used directly by another,
    /// not-yet-registered node sharing a schedulable instruction with it. Their
    /// direct dependencies must be recomputed at the next bundle scheduling,
    /// when the new node is already registered in the tree, so that the direct
    /// use is accounted for. If the new node is the last scheduled bundle and
    /// no further scheduling consumes this list, the leftover entries are
    /// dropped on the next region reset and the dependencies are recomputed
    /// against the full tree in scheduleBlock instead. A set is used to avoid
    /// recomputing the same operand more than once.
    SmallSetVector<ScheduleData *, 8> RecalcCopyableOperandDeps;
 
    /// Ordered pairs (Src, Dst) of memory instructions whose may-alias
    /// dependency has been dropped in favor of a runtime alias check.
    SmallDenseSet<std::pair<Instruction *, Instruction *>, 8> IgnoredMemDeps;
 
    /// The ID of the scheduling region. For a new vectorization iteration this
    /// is incremented which "removes" all ScheduleData from the region.
    /// Make sure that the initial SchedulingRegionID is greater than the
    /// initial SchedulingRegionID in ScheduleData (which is 0).
    int SchedulingRegionID = 1;
  };
 
  /// Attaches the BlockScheduling structures to basic blocks.
  MapVector<BasicBlock *, std::unique_ptr<BlockScheduling>> BlocksSchedules;
 
  /// Performs the "real" scheduling. Done before vectorization is actually
  /// performed in a basic block.
  void scheduleBlock(const BoUpSLP &R, BlockScheduling *BS);
 
  /// List of users to ignore during scheduling and that don't need extracting.
  const SmallDenseSet<Value *> *UserIgnoreList = nullptr;
 
  /// Narrowed reduction chain instructions, dropped together with the
  /// reduction. Subset of UserIgnoreList.
  SmallPtrSet<Value *, 4> NarrowedChainInsts;
 
  /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of
  /// sorted SmallVectors of unsigned.
  struct OrdersTypeDenseMapInfo {
    static unsigned getHashValue(const OrdersType &V) {
      return static_cast<unsigned>(hash_combine_range(V));
    }
 
    static bool isEqual(const OrdersType &LHS, const OrdersType &RHS) {
      return LHS == RHS;
    }
  };
 
  // Analysis and block reference.
  Function *F;
  ScalarEvolution *SE;
  TargetTransformInfo *TTI;
  TargetLibraryInfo *TLI;
  LoopInfo *LI;
  DominatorTree *DT;
  AssumptionCache *AC;
  DemandedBits *DB;
  const DataLayout *DL;
  OptimizationRemarkEmitter *ORE;
  /// Cached cost-model mode for this function.
  /// If -Os/-Oz, use CodeSize. Otherwise use RecipThroughput.
  const TargetTransformInfo::TargetCostKind CostKind;
 
  unsigned MaxVecRegSize; // This is set by TTI or overridden by cl::opt.
  unsigned MinVecRegSize; // Set by cl::opt (default: 128).
 
  /// Instruction builder to construct the vectorized tree.
  IRBuilder<TargetFolder> Builder;
 
  /// A map of scalar integer values to the smallest bit width with which they
  /// can legally be represented. The values map to (width, signed) pairs,
  /// where "width" indicates the minimum bit width and "signed" is True if the
  /// value must be signed-extended, rather than zero-extended, back to its
  /// original width.
  DenseMap<const TreeEntry *, std::pair<uint64_t, bool>> MinBWs;
 
  /// Final size of the reduced vector, if the current graph represents the
  /// input for the reduction and it was possible to narrow the size of the
  /// reduction.
  unsigned ReductionBitWidth = 0;
 
  /// Canonical graph size before the transformations.
  unsigned BaseGraphSize = 1;
 
  /// If the tree contains any zext/sext/trunc nodes, contains max-min pair of
  /// type sizes, used in the tree.
  std::optional<std::pair<unsigned, unsigned>> CastMaxMinBWSizes;
 
  /// Indices of the vectorized nodes, which supposed to be the roots of the new
  /// bitwidth analysis attempt, like trunc, IToFP or ICmp.
  DenseSet<unsigned> ExtraBitWidthNodes;
};
 
template <> struct llvm::DenseMapInfo<BoUpSLP::EdgeInfo> {
  using FirstInfo = DenseMapInfo<BoUpSLP::TreeEntry *>;
  using SecondInfo = DenseMapInfo<unsigned>;
  static unsigned getHashValue(const BoUpSLP::EdgeInfo &Val) {
    return detail::combineHashValue(FirstInfo::getHashValue(Val.UserTE),
                                    SecondInfo::getHashValue(Val.EdgeIdx));
  }
 
  static bool isEqual(const BoUpSLP::EdgeInfo &LHS,
                      const BoUpSLP::EdgeInfo &RHS) {
    return LHS == RHS;
  }
};
 
template <> struct llvm::GraphTraits<BoUpSLP *> {
  using TreeEntry = BoUpSLP::TreeEntry;
 
  /// NodeRef has to be a pointer per the GraphWriter.
  using NodeRef = TreeEntry *;
 
  using ContainerTy = BoUpSLP::TreeEntry::VecTreeTy;
 
  /// Add the VectorizableTree to the index iterator to be able to return
  /// TreeEntry pointers.
  struct ChildIteratorType
      : public iterator_adaptor_base<
            ChildIteratorType, SmallVector<BoUpSLP::EdgeInfo, 1>::iterator> {
    ContainerTy &VectorizableTree;
 
    ChildIteratorType(SmallVector<BoUpSLP::EdgeInfo, 1>::iterator W,
                      ContainerTy &VT)
        : ChildIteratorType::iterator_adaptor_base(W), VectorizableTree(VT) {}
 
    NodeRef operator*() { return I->UserTE; }
  };
 
  static NodeRef getEntryNode(BoUpSLP &R) { return &R.getRootNode(); }
 
  static ChildIteratorType child_begin(NodeRef N) {
    return {&N->UserTreeIndex, N->Container};
  }
 
  static ChildIteratorType child_end(NodeRef N) {
    return {&N->UserTreeIndex + 1, N->Container};
  }
 
  /// For the node iterator we just need to turn the TreeEntry iterator into a
  /// TreeEntry* iterator so that it dereferences to NodeRef.
  class nodes_iterator {
    using ItTy = ContainerTy::iterator;
    ItTy It;
 
  public:
    nodes_iterator(const ItTy &It2) : It(It2) {}
    NodeRef operator*() { return It->get(); }
    nodes_iterator operator++() {
      ++It;
      return *this;
    }
    bool operator!=(const nodes_iterator &N2) const { return N2.It != It; }
  };
 
  static nodes_iterator nodes_begin(BoUpSLP *R) {
    return nodes_iterator(R->VectorizableTree.begin());
  }
 
  static nodes_iterator nodes_end(BoUpSLP *R) {
    return nodes_iterator(R->VectorizableTree.end());
  }
 
  static unsigned size(BoUpSLP *R) { return R->VectorizableTree.size(); }
};
 
template <>
struct llvm::DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits {
  using TreeEntry = BoUpSLP::TreeEntry;
 
  DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) {}
 
  std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) {
    std::string Str;
    raw_string_ostream OS(Str);
    OS << Entry->Idx << ".\n";
    if (isSplat(Entry->Scalars))
      OS << "<splat> ";
    for (auto *V : Entry->Scalars) {
      OS << *V;
      if (llvm::any_of(R->ExternalUses, [&](const BoUpSLP::ExternalUser &EU) {
            return EU.Scalar == V;
          }))
        OS << " <extract>";
      OS << "\n";
    }
    return Str;
  }
 
  static std::string getNodeAttributes(const TreeEntry *Entry,
                                       const BoUpSLP *) {
    if (Entry->isGather())
      return "color=red";
    if (Entry->State == TreeEntry::ScatterVectorize ||
        Entry->State == TreeEntry::StridedVectorize ||
        Entry->State == TreeEntry::ExpandVectorize ||
        Entry->State == TreeEntry::CompressVectorize ||
        Entry->State == TreeEntry::BlendedLoadVectorize)
      return "color=blue";
    return "";
  }
};
 
BoUpSLP::~BoUpSLP() {
  SmallVector<WeakTrackingVH> DeadInsts;
  for (auto *I : DeletedInstructions) {
    if (!I->getParent()) {
      // Temporarily insert instruction back to erase them from parent and
      // memory later.
      if (isa<PHINode>(I))
        // Phi nodes must be the very first instructions in the block.
        I->insertBefore(F->getEntryBlock(),
                        F->getEntryBlock().getFirstNonPHIIt());
      else
        I->insertBefore(F->getEntryBlock().getTerminator()->getIterator());
      continue;
    }
    for (Use &U : I->operands()) {
      auto *Op = dyn_cast<Instruction>(U.get());
      if (Op && !DeletedInstructions.count(Op) && Op->hasOneUser() &&
          wouldInstructionBeTriviallyDead(Op, TLI))
        DeadInsts.emplace_back(Op);
    }
    I->dropAllReferences();
  }
  for (auto *I : DeletedInstructions) {
    assert(I->use_empty() &&
           "trying to erase instruction with users.");
    I->eraseFromParent();
  }
 
  // Cleanup any dead scalar code feeding the vectorized instructions
  RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI);
 
#ifdef EXPENSIVE_CHECKS
  // If we could guarantee that this call is not extremely slow, we could
  // remove the ifdef limitation (see PR47712).
  assert(!verifyFunction(*F, &dbgs()));
#endif
}
 
/// Reorders the given \p Reuses mask according to the given \p Mask. \p Reuses
/// contains original mask for the scalars reused in the node. Procedure
/// transform this mask in accordance with the given \p Mask.
static void reorderReuses(SmallVectorImpl<int> &Reuses, ArrayRef<int> Mask) {
  assert(!Mask.empty() && Reuses.size() == Mask.size() &&
         "Expected non-empty mask.");
  SmallVector<int> Prev(Reuses.begin(), Reuses.end());
  Prev.swap(Reuses);
  for (unsigned I = 0, E = Prev.size(); I < E; ++I)
    if (Mask[I] != PoisonMaskElem)
      Reuses[Mask[I]] = Prev[I];
}
 
/// Reorders the given \p Order according to the given \p Mask. \p Order - is
/// the original order of the scalars. Procedure transforms the provided order
/// in accordance with the given \p Mask. If the resulting \p Order is just an
/// identity order, \p Order is cleared.
static void reorderOrder(SmallVectorImpl<unsigned> &Order, ArrayRef<int> Mask,
                         bool BottomOrder = false) {
  assert(!Mask.empty() && "Expected non-empty mask.");
  unsigned Sz = Mask.size();
  if (BottomOrder) {
    SmallVector<unsigned> PrevOrder;
    if (Order.empty()) {
      PrevOrder.resize(Sz);
      std::iota(PrevOrder.begin(), PrevOrder.end(), 0);
    } else {
      PrevOrder.swap(Order);
    }
    Order.assign(Sz, Sz);
    for (unsigned I = 0; I < Sz; ++I)
      if (Mask[I] != PoisonMaskElem)
        Order[I] = PrevOrder[Mask[I]];
    if (all_of(enumerate(Order), [&](const auto &Data) {
          return Data.value() == Sz || Data.index() == Data.value();
        })) {
      Order.clear();
      return;
    }
    fixupOrderingIndices(Order);
    return;
  }
  SmallVector<int> MaskOrder;
  if (Order.empty()) {
    MaskOrder.resize(Sz);
    std::iota(MaskOrder.begin(), MaskOrder.end(), 0);
  } else {
    inversePermutation(Order, MaskOrder);
  }
  reorderReuses(MaskOrder, Mask);
  if (ShuffleVectorInst::isIdentityMask(MaskOrder, Sz)) {
    Order.clear();
    return;
  }
  Order.assign(Sz, Sz);
  for (unsigned I = 0; I < Sz; ++I)
    if (MaskOrder[I] != PoisonMaskElem)
      Order[MaskOrder[I]] = I;
  fixupOrderingIndices(Order);
}
 
std::optional<BoUpSLP::OrdersType>
BoUpSLP::findReusedOrderedScalars(const BoUpSLP::TreeEntry &TE,
                                  bool TopToBottom, bool IgnoreReorder) {
  assert(TE.isGather() && "Expected gather node only.");
  // Try to find subvector extract/insert patterns and reorder only such
  // patterns.
  SmallVector<Value *> GatheredScalars(TE.Scalars.begin(), TE.Scalars.end());
  Type *ScalarTy = GatheredScalars.front()->getType();
  size_t NumScalars = GatheredScalars.size();
  if (!isValidElementType(ScalarTy, SLPReVec))
    return std::nullopt;
  auto *VecTy = getWidenedType(ScalarTy, NumScalars);
  unsigned NumParts = getNumberOfParts(VecTy, ScalarTy, NumScalars);
  SmallVector<int> ExtractMask;
  SmallVector<int> Mask;
  SmallVector<SmallVector<const TreeEntry *>> Entries;
  SmallVector<std::optional<TargetTransformInfo::ShuffleKind>> ExtractShuffles =
      tryToGatherExtractElements(GatheredScalars, ExtractMask, NumParts);
  SmallVector<std::optional<TargetTransformInfo::ShuffleKind>> GatherShuffles =
      isGatherShuffledEntry(&TE, GatheredScalars, Mask, Entries, NumParts,
                            /*ForOrder=*/true);
  // No shuffled operands - ignore.
  if (GatherShuffles.empty() && ExtractShuffles.empty())
    return std::nullopt;
  OrdersType CurrentOrder(NumScalars, NumScalars);
  if (GatherShuffles.size() == 1 &&
      *GatherShuffles.front() == TTI::SK_PermuteSingleSrc &&
      Entries.front().front()->isSame(TE.Scalars)) {
    // If the full matched node in whole tree rotation - no need to consider the
    // matching order, rotating the whole tree.
    if (TopToBottom)
      return std::nullopt;
    // No need to keep the order for the same user node.
    if (Entries.front().front()->UserTreeIndex.UserTE ==
        TE.UserTreeIndex.UserTE)
      return std::nullopt;
    // No need to keep the order for the matched root node, if it can be freely
    // reordered.
    if (!IgnoreReorder && Entries.front().front()->Idx == 0)
      return std::nullopt;
    // If shuffling 2 elements only and the matching node has reverse reuses -
    // no need to count order, both work fine.
    if (!Entries.front().front()->ReuseShuffleIndices.empty() &&
        TE.getVectorFactor() == 2 && Mask.size() == 2 &&
        any_of(enumerate(Entries.front().front()->ReuseShuffleIndices),
               [](const auto &P) {
                 return P.value() % 2 != static_cast<int>(P.index()) % 2;
               }))
      return std::nullopt;
 
    // Perfect match in the graph, will reuse the previously vectorized
    // node. Cost is 0.
    std::iota(CurrentOrder.begin(), CurrentOrder.end(), 0);
    return CurrentOrder;
  }
  auto IsSplatMask = [](ArrayRef<int> Mask) {
    int SingleElt = PoisonMaskElem;
    return all_of(Mask, [&](int I) {
      if (SingleElt == PoisonMaskElem && I != PoisonMaskElem)
        SingleElt = I;
      return I == PoisonMaskElem || I == SingleElt;
    });
  };
  // Exclusive broadcast mask - ignore.
  if ((ExtractShuffles.empty() && IsSplatMask(Mask) &&
       (Entries.size() != 1 ||
        Entries.front().front()->ReorderIndices.empty())) ||
      (GatherShuffles.empty() && IsSplatMask(ExtractMask)))
    return std::nullopt;
  SmallBitVector ShuffledSubMasks(NumParts);
  auto TransformMaskToOrder = [&](MutableArrayRef<unsigned> CurrentOrder,
                                  ArrayRef<int> Mask, int PartSz, int NumParts,
                                  function_ref<unsigned(unsigned)> GetVF) {
    for (int I : seq<int>(NumParts)) {
      if (ShuffledSubMasks.test(I))
        continue;
      const int VF = GetVF(I);
      if (VF == 0)
        continue;
      unsigned Limit = getNumElems(CurrentOrder.size(), PartSz, I);
      MutableArrayRef<unsigned> Slice = CurrentOrder.slice(I * PartSz, Limit);
      // Shuffle of at least 2 vectors - ignore.
      if (any_of(Slice, not_equal_to(NumScalars))) {
        llvm::fill(Slice, NumScalars);
        ShuffledSubMasks.set(I);
        continue;
      }
      // Try to include as much elements from the mask as possible.
      int FirstMin = INT_MAX;
      int SecondVecFound = false;
      for (int K : seq<int>(Limit)) {
        int Idx = Mask[I * PartSz + K];
        if (Idx == PoisonMaskElem) {
          Value *V = GatheredScalars[I * PartSz + K];
          if (isConstant(V) && !isa<PoisonValue>(V)) {
            SecondVecFound = true;
            break;
          }
          continue;
        }
        if (Idx < VF) {
          if (FirstMin > Idx)
            FirstMin = Idx;
        } else {
          SecondVecFound = true;
          break;
        }
      }
      FirstMin = (FirstMin / PartSz) * PartSz;
      // Shuffle of at least 2 vectors - ignore.
      if (SecondVecFound) {
        llvm::fill(Slice, NumScalars);
        ShuffledSubMasks.set(I);
        continue;
      }
      for (int K : seq<int>(Limit)) {
        int Idx = Mask[I * PartSz + K];
        if (Idx == PoisonMaskElem)
          continue;
        Idx -= FirstMin;
        if (Idx >= PartSz) {
          // Cross-part / second-vector reference: this slice cannot be
          // ordered as a single first-vector permutation, give up.
          SecondVecFound = true;
          break;
        }
        // For the last partial slice, Limit < PartSz and Idx in [Limit,
        // PartSz) addresses the unused padded tail (no scalar at that
        // position). Skip the write but keep ordering the remaining K's.
        if (static_cast<unsigned>(I * PartSz + Idx) >= CurrentOrder.size())
          continue;
        if (CurrentOrder[I * PartSz + Idx] >
                static_cast<unsigned>(I * PartSz + K) &&
            CurrentOrder[I * PartSz + Idx] !=
                static_cast<unsigned>(I * PartSz + Idx))
          CurrentOrder[I * PartSz + Idx] = I * PartSz + K;
      }
      // Shuffle of at least 2 vectors - ignore.
      if (SecondVecFound) {
        llvm::fill(Slice, NumScalars);
        ShuffledSubMasks.set(I);
        continue;
      }
    }
  };
  int PartSz = getPartNumElems(NumScalars, NumParts);
  if (!ExtractShuffles.empty())
    TransformMaskToOrder(
        CurrentOrder, ExtractMask, PartSz, NumParts, [&](unsigned I) {
          if (I >= ExtractShuffles.size() || !ExtractShuffles[I])
            return 0U;
          unsigned VF = 0;
          unsigned Sz = getNumElems(TE.getVectorFactor(), PartSz, I);
          for (unsigned Idx : seq<unsigned>(Sz)) {
            int K = I * PartSz + Idx;
            if (static_cast<unsigned>(K) >= ExtractMask.size())
              break;
            if (ExtractMask[K] == PoisonMaskElem)
              continue;
            if (!TE.ReuseShuffleIndices.empty())
              K = TE.ReuseShuffleIndices[K];
            if (K == PoisonMaskElem)
              continue;
            if (!TE.ReorderIndices.empty())
              K = std::distance(TE.ReorderIndices.begin(),
                                find(TE.ReorderIndices, K));
            auto *EI = dyn_cast<ExtractElementInst>(TE.Scalars[K]);
            if (!EI)
              continue;
            VF = std::max(VF, EI->getVectorOperandType()
                                  ->getElementCount()
                                  .getKnownMinValue());
          }
          return VF;
        });
  // Check special corner case - single shuffle of the same entry.
  if (GatherShuffles.size() == 1 && NumParts != 1) {
    if (ShuffledSubMasks.any())
      return std::nullopt;
    PartSz = NumScalars;
    NumParts = 1;
  }
  if (!Entries.empty())
    TransformMaskToOrder(CurrentOrder, Mask, PartSz, NumParts, [&](unsigned I) {
      if (I >= GatherShuffles.size() || !GatherShuffles[I])
        return 0U;
      return std::max(Entries[I].front()->getVectorFactor(),
                      Entries[I].back()->getVectorFactor());
    });
  unsigned NumUndefs = count(CurrentOrder, NumScalars);
  if (ShuffledSubMasks.all() || (NumScalars > 2 && NumUndefs >= NumScalars / 2))
    return std::nullopt;
  return std::move(CurrentOrder);
}
 
static bool arePointersCompatible(Value *Ptr1, Value *Ptr2,
                                  const TargetLibraryInfo &TLI,
                                  bool CompareOpcodes = true) {
  if (getUnderlyingObject(Ptr1, RecursionMaxDepth) !=
      getUnderlyingObject(Ptr2, RecursionMaxDepth))
    return false;
  auto *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
  auto *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
  return (!GEP1 || GEP1->getNumOperands() == 2) &&
         (!GEP2 || GEP2->getNumOperands() == 2) &&
         (((!GEP1 || isConstant(GEP1->getOperand(1))) &&
           (!GEP2 || isConstant(GEP2->getOperand(1)))) ||
          !CompareOpcodes ||
          (GEP1 && GEP2 &&
           getSameOpcode({GEP1->getOperand(1), GEP2->getOperand(1)}, TLI)));
}
 
/// Calculates minimal alignment as a common alignment.
template <typename T>
static Align computeCommonAlignment(ArrayRef<Value *> VL) {
  Align CommonAlignment = cast<T>(VL.consume_front())->getAlign();
  for (Value *V : VL)
    CommonAlignment = std::min(CommonAlignment, cast<T>(V)->getAlign());
  return CommonAlignment;
}
 
/// Check if \p Order represents reverse order.
static bool isReverseOrder(ArrayRef<unsigned> Order) {
  assert(!Order.empty() &&
         "Order is empty. Please check it before using isReverseOrder.");
  unsigned Sz = Order.size();
  return all_of(enumerate(Order), [&](const auto &Pair) {
    return Pair.value() == Sz || Sz - Pair.index() - 1 == Pair.value();
  });
}
 
/// Checks if the provided list of pointers \p Pointers represents the strided
/// pointers for type ElemTy. If they are not, nullptr is returned.
/// Otherwise, SCEV* of the stride value is returned.
/// If `PointerOps` can be rearanged into the following sequence:
/// ```
/// %x + c_0 * stride,
/// %x + c_1 * stride,
/// %x + c_2 * stride
/// ...
/// ```
/// where each `c_i` is constant. The SCEV of the `stride` will be returned.
static const SCEV *calculateRtStride(ArrayRef<Value *> PointerOps, Type *ElemTy,
                                     const DataLayout &DL, ScalarEvolution &SE,
                                     SmallVectorImpl<unsigned> &SortedIndices) {
  SmallVector<const SCEV *> SCEVs;
  const SCEV *PtrSCEVLowest = nullptr;
  const SCEV *PtrSCEVHighest = nullptr;
  // Find lower/upper pointers from the PointerOps (i.e. with lowest and highest
  // addresses).
  for (Value *Ptr : PointerOps) {
    const SCEV *PtrSCEV = SE.getSCEV(Ptr);
    if (!PtrSCEV)
      return nullptr;
    SCEVs.push_back(PtrSCEV);
    if (!PtrSCEVLowest && !PtrSCEVHighest) {
      PtrSCEVLowest = PtrSCEVHighest = PtrSCEV;
      continue;
    }
    const SCEV *Diff = SE.getMinusSCEV(PtrSCEV, PtrSCEVLowest);
    if (isa<SCEVCouldNotCompute>(Diff))
      return nullptr;
    if (Diff->isNonConstantNegative()) {
      PtrSCEVLowest = PtrSCEV;
      continue;
    }
    const SCEV *Diff1 = SE.getMinusSCEV(PtrSCEVHighest, PtrSCEV);
    if (isa<SCEVCouldNotCompute>(Diff1))
      return nullptr;
    if (Diff1->isNonConstantNegative()) {
      PtrSCEVHighest = PtrSCEV;
      continue;
    }
  }
  // Dist = PtrSCEVHighest - PtrSCEVLowest;
  const SCEV *Dist = SE.getMinusSCEV(PtrSCEVHighest, PtrSCEVLowest);
  if (isa<SCEVCouldNotCompute>(Dist))
    return nullptr;
  int Size = DL.getTypeStoreSize(ElemTy);
  auto TryGetStride = [&](const SCEV *Dist,
                          const SCEV *Multiplier) -> const SCEV * {
    if (const auto *M = dyn_cast<SCEVMulExpr>(Dist)) {
      if (M->getOperand(0) == Multiplier)
        return M->getOperand(1);
      if (M->getOperand(1) == Multiplier)
        return M->getOperand(0);
      return nullptr;
    }
    if (Multiplier == Dist)
      return SE.getConstant(Dist->getType(), 1);
    return SE.getUDivExactExpr(Dist, Multiplier);
  };
  // Stride_in_elements = Dist / element_size * (num_elems - 1).
  const SCEV *Stride = nullptr;
  if (Size != 1 || SCEVs.size() > 1) {
    const SCEV *Sz = SE.getConstant(Dist->getType(), Size * (SCEVs.size() - 1));
    Stride = TryGetStride(Dist, Sz);
    if (!Stride)
      return nullptr;
  }
  if (!Stride || isa<SCEVConstant>(Stride))
    return nullptr;
  // Iterate through all pointers and check if all distances are
  // unique multiple of Stride.
  using DistOrdPair = std::pair<int64_t, int>;
  auto Compare = llvm::less_first();
  std::set<DistOrdPair, decltype(Compare)> Offsets(Compare);
  bool IsConsecutive = true;
  for (const auto [Idx, PtrSCEV] : enumerate(SCEVs)) {
    unsigned Dist = 0;
    if (PtrSCEV != PtrSCEVLowest) {
      const SCEV *Diff = SE.getMinusSCEV(PtrSCEV, PtrSCEVLowest);
      const SCEV *Coeff = TryGetStride(Diff, Stride);
      if (!Coeff)
        return nullptr;
      const auto *SC = dyn_cast<SCEVConstant>(Coeff);
      if (!SC || isa<SCEVCouldNotCompute>(SC))
        return nullptr;
      if (!SE.getMinusSCEV(PtrSCEV, SE.getAddExpr(PtrSCEVLowest,
                                                  SE.getMulExpr(Stride, SC)))
               ->isZero())
        return nullptr;
      Dist = SC->getAPInt().getZExtValue();
    }
    // If the strides are not the same or repeated, we can't vectorize.
    if ((Dist / Size) * Size != Dist || (Dist / Size) >= SCEVs.size())
      return nullptr;
    auto Res = Offsets.emplace(Dist, Idx);
    if (!Res.second)
      return nullptr;
    // Consecutive order if the inserted element is the last one.
    IsConsecutive = IsConsecutive && std::next(Res.first) == Offsets.end();
  }
  SortedIndices.clear();
  if (!IsConsecutive) {
    // Fill SortedIndices array only if it is non-consecutive.
    SortedIndices.resize(PointerOps.size());
    for (const auto [Idx, Pair] : enumerate(Offsets))
      SortedIndices[Idx] = Pair.second;
  }
  return Stride;
}
 
/// Creates subvector insert. Generates shuffle using \p Generator or
/// using default shuffle.
static Value *createInsertVector(
    IRBuilderBase &Builder, Value *Vec, Value *V, unsigned Index,
    function_ref<Value *(Value *, Value *, ArrayRef<int>)> Generator = {}) {
  if (isa<PoisonValue>(Vec) && isa<PoisonValue>(V))
    return Vec;
  const unsigned SubVecVF = getNumElements(V->getType());
  // Create shuffle, insertvector requires that index is multiple of
  // the subvector length.
  const unsigned VecVF = getNumElements(Vec->getType());
  SmallVector<int> Mask(VecVF, PoisonMaskElem);
  if (isa<PoisonValue>(Vec)) {
    auto *Begin = std::next(Mask.begin(), Index);
    std::iota(Begin, std::next(Begin, SubVecVF), 0);
    Vec = Builder.CreateShuffleVector(V, Mask);
    return Vec;
  }
  std::iota(Mask.begin(), Mask.end(), 0);
  std::iota(std::next(Mask.begin(), Index),
            std::next(Mask.begin(), Index + SubVecVF), VecVF);
  if (Generator)
    return Generator(Vec, V, Mask);
  // 1. Resize V to the size of Vec.
  SmallVector<int> ResizeMask(VecVF, PoisonMaskElem);
  std::iota(ResizeMask.begin(), std::next(ResizeMask.begin(), SubVecVF), 0);
  V = Builder.CreateShuffleVector(V, ResizeMask);
  // 2. Insert V into Vec.
  return Builder.CreateShuffleVector(Vec, V, Mask);
}
 
/// Generates subvector extract using \p Generator or using default shuffle.
static Value *createExtractVector(IRBuilderBase &Builder, Value *Vec,
                                  unsigned SubVecVF, unsigned Index) {
  SmallVector<int> Mask(SubVecVF, PoisonMaskElem);
  std::iota(Mask.begin(), Mask.end(), Index);
  return Builder.CreateShuffleVector(Vec, Mask);
}
 
/// Builds compress-like mask for shuffles for the given \p PointerOps, ordered
/// with \p Order.
/// \return true if the mask represents strided access, false - otherwise.
static bool buildCompressMask(ArrayRef<Value *> PointerOps,
                              ArrayRef<unsigned> Order, Type *ScalarTy,
                              const DataLayout &DL, ScalarEvolution &SE,
                              SmallVectorImpl<int> &CompressMask) {
  const unsigned Sz = PointerOps.size();
  CompressMask.assign(Sz, PoisonMaskElem);
  // The first element always set.
  CompressMask[0] = 0;
  // Check if the mask represents strided access.
  std::optional<unsigned> Stride = 0;
  Value *Ptr0 = Order.empty() ? PointerOps.front() : PointerOps[Order.front()];
  for (unsigned I : seq<unsigned>(1, Sz)) {
    Value *Ptr = Order.empty() ? PointerOps[I] : PointerOps[Order[I]];
    std::optional<int64_t> OptPos =
        getPointersDiff(ScalarTy, Ptr0, ScalarTy, Ptr, DL, SE);
    if (!OptPos || OptPos > std::numeric_limits<unsigned>::max())
      return false;
    unsigned Pos = static_cast<unsigned>(*OptPos);
    CompressMask[I] = Pos;
    if (!Stride)
      continue;
    if (*Stride == 0) {
      *Stride = Pos;
      continue;
    }
    if (Pos != *Stride * I)
      Stride.reset();
  }
  return Stride.has_value();
}
 
/// Checks if the \p VL can be transformed to a (masked)load + compress or
/// (masked) interleaved load.
static bool isMaskedLoadCompress(
    ArrayRef<Value *> VL, ArrayRef<Value *> PointerOps,
    ArrayRef<unsigned> Order, const TargetTransformInfo &TTI,
    const DataLayout &DL, ScalarEvolution &SE, AssumptionCache &AC,
    const DominatorTree &DT, const TargetLibraryInfo &TLI,
    const TTI::TargetCostKind CostKind,
    const function_ref<bool(Value *)> AreAllUsersVectorized, bool &IsMasked,
    unsigned &InterleaveFactor, SmallVectorImpl<int> &CompressMask,
    VectorType *&LoadVecTy) {
  InterleaveFactor = 0;
  Type *ScalarTy = VL.front()->getType();
  const size_t Sz = VL.size();
  auto *VecTy = cast<VectorType>(getWidenedType(ScalarTy, Sz));
  SmallVector<int> Mask;
  if (!Order.empty())
    inversePermutation(Order, Mask);
  // Check external uses.
  for (const auto [I, V] : enumerate(VL)) {
    if (AreAllUsersVectorized(V))
      continue;
    InstructionCost ExtractCost =
        TTI.getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind,
                               Mask.empty() ? I : Mask[I]);
    InstructionCost ScalarCost =
        TTI.getInstructionCost(cast<Instruction>(V), CostKind);
    if (ExtractCost <= ScalarCost)
      return false;
  }
  Value *Ptr0;
  Value *PtrN;
  if (Order.empty()) {
    Ptr0 = PointerOps.front();
    PtrN = PointerOps.back();
  } else {
    Ptr0 = PointerOps[Order.front()];
    PtrN = PointerOps[Order.back()];
  }
  std::optional<int64_t> Diff =
      getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, DL, SE);
  if (!Diff)
    return false;
  const size_t MaxRegSize =
      TTI.getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
          .getFixedValue();
  // Check for very large distances between elements.
  if (*Diff / Sz >= MaxRegSize / 8)
    return false;
  LoadVecTy = cast<FixedVectorType>(getWidenedType(ScalarTy, *Diff + 1));
  auto *LI = cast<LoadInst>(Order.empty() ? VL.front() : VL[Order.front()]);
  Align CommonAlignment = LI->getAlign();
  SimplifyQuery SQ(
      DL, &TLI, &DT, &AC,
      cast<LoadInst>(Order.empty() ? VL.back() : VL[Order.back()]));
  IsMasked = !isSafeToLoadUnconditionally(Ptr0, LoadVecTy, CommonAlignment, SQ);
  if (IsMasked && !TTI.isLegalMaskedLoad(LoadVecTy, CommonAlignment,
                                         LI->getPointerAddressSpace()))
    return false;
  // TODO: perform the analysis of each scalar load for better
  // safe-load-unconditionally analysis.
  bool IsStrided =
      buildCompressMask(PointerOps, Order, ScalarTy, DL, SE, CompressMask);
  assert(CompressMask.size() >= 2 && "At least two elements are required");
  SmallVector<Value *> OrderedPointerOps(PointerOps);
  if (!Order.empty())
    reorderScalars(OrderedPointerOps, Mask);
  auto [ScalarGEPCost, VectorGEPCost] =
      getGEPCosts(TTI, OrderedPointerOps, OrderedPointerOps.front(),
                  Instruction::Load, CostKind, ScalarTy, LoadVecTy);
  // The cost of scalar loads.
  InstructionCost ScalarLoadsCost =
      accumulate(VL, InstructionCost(),
                 [&](InstructionCost C, Value *V) {
                   return C + TTI.getInstructionCost(cast<Instruction>(V),
                                                     CostKind);
                 }) +
      ScalarGEPCost;
  APInt DemandedElts = APInt::getAllOnes(Sz);
  InstructionCost GatherCost =
      getScalarizationOverhead(TTI, SLPReVec, ScalarTy, VecTy, DemandedElts,
                               /*Insert=*/true,
                               /*Extract=*/false, CostKind) +
      ScalarLoadsCost;
  InstructionCost LoadCost = 0;
  if (IsMasked) {
    LoadCost = TTI.getMemIntrinsicInstrCost(
        MemIntrinsicCostAttributes(Intrinsic::masked_load, LoadVecTy,
                                   CommonAlignment,
                                   LI->getPointerAddressSpace()),
        CostKind);
  } else {
    LoadCost =
        TTI.getMemoryOpCost(Instruction::Load, LoadVecTy, CommonAlignment,
                            LI->getPointerAddressSpace(), CostKind);
  }
  if (IsStrided && !IsMasked && Order.empty()) {
    // Check for potential segmented(interleaved) loads.
    VectorType *AlignedLoadVecTy = cast<VectorType>(getWidenedType(
        ScalarTy,
        getFullVectorNumberOfElements(TTI, ScalarTy, *Diff + 1, SLPReVec)));
    SimplifyQuery SQ(DL, &TLI, &DT, &AC, cast<LoadInst>(VL.back()));
    if (!isSafeToLoadUnconditionally(Ptr0, AlignedLoadVecTy, CommonAlignment,
                                     SQ))
      AlignedLoadVecTy = LoadVecTy;
    if (TTI.isLegalInterleavedAccessType(AlignedLoadVecTy, CompressMask[1],
                                         CommonAlignment,
                                         LI->getPointerAddressSpace())) {
      InstructionCost InterleavedCost =
          VectorGEPCost + TTI.getInterleavedMemoryOpCost(
                              Instruction::Load, AlignedLoadVecTy,
                              CompressMask[1], {}, CommonAlignment,
                              LI->getPointerAddressSpace(), CostKind, IsMasked);
      if (InterleavedCost < GatherCost) {
        InterleaveFactor = CompressMask[1];
        LoadVecTy = AlignedLoadVecTy;
        return true;
      }
    }
  }
  // Estimating the compression shuffle cost below can be extremely expensive
  // for a very wide LoadVecTy, which is split into a large number of vector
  // registers (see processShuffleMasks). The shuffle cost is always
  // non-negative, so if the load cost alone already reaches the gather cost the
  // masked-load-compress cannot be profitable. Bail out before the costly
  // shuffle cost estimation in that case.
  if (VectorGEPCost + LoadCost >= GatherCost)
    return false;
  InstructionCost CompressCost = getShuffleCost(
      TTI, TTI::SK_PermuteSingleSrc, LoadVecTy, CostKind, CompressMask);
  if (!Order.empty()) {
    SmallVector<int> NewMask(Sz, PoisonMaskElem);
    for (unsigned I : seq<unsigned>(Sz)) {
      NewMask[I] = CompressMask[Mask[I]];
    }
    CompressMask.swap(NewMask);
  }
  InstructionCost TotalVecCost = VectorGEPCost + LoadCost + CompressCost;
  return TotalVecCost < GatherCost;
}
 
/// Checks if the \p VL can be transformed to a (masked)load + compress or
/// (masked) interleaved load.
static bool
isMaskedLoadCompress(ArrayRef<Value *> VL, ArrayRef<Value *> PointerOps,
                     ArrayRef<unsigned> Order, const TargetTransformInfo &TTI,
                     const DataLayout &DL, ScalarEvolution &SE,
                     AssumptionCache &AC, const DominatorTree &DT,
                     const TargetLibraryInfo &TLI,
                     const TTI::TargetCostKind CostKind,
                     const function_ref<bool(Value *)> AreAllUsersVectorized) {
  bool IsMasked;
  unsigned InterleaveFactor;
  SmallVector<int> CompressMask;
  VectorType *LoadVecTy;
  return isMaskedLoadCompress(VL, PointerOps, Order, TTI, DL, SE, AC, DT, TLI,
                              CostKind, AreAllUsersVectorized, IsMasked,
                              InterleaveFactor, CompressMask, LoadVecTy);
}
 
/// Checks if the stores \p VL with pointers \p PointerOps can be lowered as a
/// single masked store. On success \p StoreVecTy is the widened store type and
/// \p ReuseShuffleIndices is the expand mask that places each stored value at
/// its element offset from the base (poison in the gaps).
static bool isMaskedStoreCompress(
    ArrayRef<Value *> VL, ArrayRef<Value *> PointerOps,
    ArrayRef<unsigned> Order, const TargetTransformInfo &TTI,
    const DataLayout &DL, ScalarEvolution &SE, Align CommonAlignment,
    SmallVectorImpl<int> &ReuseShuffleIndices, FixedVectorType *&StoreVecTy) {
  Type *ScalarTy = cast<StoreInst>(VL.front())->getValueOperand()->getType();
  const size_t Sz = VL.size();
  // Only simple scalar element types are supported.
  if (Sz < 2 || (!ScalarTy->isIntOrPtrTy() && !ScalarTy->isFloatingPointTy()))
    return false;
  Value *Ptr0 = Order.empty() ? PointerOps.front() : PointerOps[Order.front()];
  Value *PtrN = Order.empty() ? PointerOps.back() : PointerOps[Order.back()];
  std::optional<int64_t> Diff =
      getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, DL, SE);
  if (!Diff || *Diff <= 0)
    return false;
  // Avoid widened vectors with very large gaps between the stored elements.
  const unsigned MaxRegSize =
      TTI.getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
          .getFixedValue();
  const unsigned ScalarBits = DL.getTypeSizeInBits(ScalarTy).getFixedValue();
  if (ScalarBits == 0 ||
      static_cast<uint64_t>(*Diff) / Sz >= MaxRegSize / ScalarBits)
    return false;
  StoreVecTy = cast<FixedVectorType>(getWidenedType(ScalarTy, *Diff + 1));
  unsigned AS = cast<StoreInst>(VL.front())->getPointerAddressSpace();
  if (!TTI.isLegalMaskedStore(StoreVecTy, CommonAlignment, AS,
                              TTI::ConstantMask))
    return false;
  // Build the expand mask: store I (in address-sorted order) is placed at its
  // element offset from the base, other widened lanes are poison.
  ReuseShuffleIndices.assign(*Diff + 1, PoisonMaskElem);
  int64_t Prev = -1;
  for (unsigned I : seq<unsigned>(Sz)) {
    Value *Ptr = Order.empty() ? PointerOps[I] : PointerOps[Order[I]];
    std::optional<int64_t> Off =
        getPointersDiff(ScalarTy, Ptr0, ScalarTy, Ptr, DL, SE);
    if (!Off || *Off <= Prev || *Off > *Diff)
      return false;
    ReuseShuffleIndices[*Off] = static_cast<int>(I);
    Prev = *Off;
  }
  return true;
}
 
/// Checks if strided loads can be generated out of \p VL loads with pointers \p
/// PointerOps:
/// 1. Target with strided load support is detected.
/// 2. The number of loads is greater than MinProfitableStridedLoads, or the
/// potential stride <= MaxProfitableStride and the potential stride is
/// power-of-2 (to avoid perf regressions for the very small number of loads)
/// and max distance > number of loads, or potential stride is -1.
/// 3. The loads are ordered, or number of unordered loads <=
/// MaxProfitableUnorderedLoads, or loads are in reversed order. (this check is
/// to avoid extra costs for very expensive shuffles).
/// 4. Any pointer operand is an instruction with the users outside of the
/// current graph (for masked gathers extra extractelement instructions
/// might be required).
bool BoUpSLP::isStridedLoad(ArrayRef<Value *> PointerOps, Type *ScalarTy,
                            Align Alignment, const int64_t Diff,
                            const size_t Sz) const {
  if (Diff % (Sz - 1) != 0)
    return false;
 
  // Try to generate strided load node.
  auto IsAnyPointerUsedOutGraph = any_of(PointerOps, [&](Value *V) {
    return isa<Instruction>(V) && any_of(V->users(), [&](User *U) {
             return !isVectorized(U) && !MustGather.contains(U);
           });
  });
 
  const uint64_t AbsoluteDiff = std::abs(Diff);
  auto *VecTy = getWidenedType(ScalarTy, Sz);
  if (IsAnyPointerUsedOutGraph ||
      (AbsoluteDiff > Sz &&
       (Sz > MinProfitableStridedLoads ||
        (AbsoluteDiff <= MaxProfitableStride * Sz && AbsoluteDiff % Sz == 0 &&
         has_single_bit(AbsoluteDiff / Sz)))) ||
      Diff == -(static_cast<int64_t>(Sz) - 1)) {
    int64_t Stride = Diff / static_cast<int64_t>(Sz - 1);
    if (Diff != Stride * static_cast<int64_t>(Sz - 1))
      return false;
    if (!TTI->isLegalStridedLoadStore(VecTy, Alignment))
      return false;
    return true;
  }
  return false;
}
 
bool BoUpSLP::analyzeConstantStrideCandidate(
    const ArrayRef<Value *> PointerOps, Type *ScalarTy, Align Alignment,
    const SmallVectorImpl<unsigned> &SortedIndices, const int64_t Diff,
    Value *Ptr0, StridedPtrInfo &SPtrInfo) const {
  const size_t Sz = PointerOps.size();
  SmallVector<int64_t> SortedOffsetsFromBase(Sz);
  // Go through `PointerOps` in sorted order and record offsets from
  // PointerOps[0]. We use PointerOps[0] rather than Ptr0 because
  // sortPtrAccesses only validates getPointersDiff for pairs relative to
  // PointerOps[0]. This is safe since only offset differences are used below.
  for (unsigned I : seq<unsigned>(Sz)) {
    Value *Ptr =
        SortedIndices.empty() ? PointerOps[I] : PointerOps[SortedIndices[I]];
    std::optional<int64_t> Offset =
        getPointersDiff(ScalarTy, PointerOps[0], ScalarTy, Ptr, *DL, *SE);
    assert(Offset && "sortPtrAccesses should have validated this pointer");
    SortedOffsetsFromBase[I] = *Offset;
  }
 
  // The code below checks that `SortedOffsetsFromBase` looks as follows:
  // ```
  // [
  //   (e_{0, 0}, e_{0, 1}, ..., e_{0, GroupSize - 1}), // first group
  //   (e_{1, 0}, e_{1, 1}, ..., e_{1, GroupSize - 1}), // secon group
  //   ...
  //   (e_{NumGroups - 1, 0}, e_{NumGroups - 1, 1}, ..., e_{NumGroups - 1,
  //   GroupSize - 1}), // last group
  // ]
  // ```
  // The distance between consecutive elements within each group should all be
  // the same `StrideWithinGroup`. The distance between the first elements of
  // consecutive groups should all be the same `StrideBetweenGroups`.
 
  int64_t StrideWithinGroup =
      SortedOffsetsFromBase[1] - SortedOffsetsFromBase[0];
  // Determine size of the first group. Later we will check that all other
  // groups have the same size.
  auto IsEndOfGroupIndex = [=, &SortedOffsetsFromBase](unsigned Idx) {
    return SortedOffsetsFromBase[Idx] - SortedOffsetsFromBase[Idx - 1] !=
           StrideWithinGroup;
  };
  auto Indices = seq<unsigned>(1, Sz);
  auto FoundIt = llvm::find_if(Indices, IsEndOfGroupIndex);
  unsigned GroupSize = FoundIt != Indices.end() ? *FoundIt : Sz;
 
  unsigned VecSz = Sz;
  Type *NewScalarTy = ScalarTy;
 
  // Quick detour: at this point we can say what the type of strided load would
  // be if all the checks pass. Check if this type is legal for the target.
  bool NeedsWidening = Sz != GroupSize;
  const uint64_t UnitBitWidth = DL->getTypeSizeInBits(ScalarTy).getFixedValue();
  if (NeedsWidening) {
    if (Sz % GroupSize != 0)
      return false;
 
    if (StrideWithinGroup != 1)
      return false;
    VecSz = Sz / GroupSize;
    NewScalarTy = Type::getIntNTy(SE->getContext(), UnitBitWidth * GroupSize);
  } else if (ScalarTy->isVectorTy()) {
    NewScalarTy = Type::getIntNTy(SE->getContext(), UnitBitWidth);
  }
 
  if (!isStridedLoad(PointerOps, NewScalarTy, Alignment, Diff, VecSz))
    return false;
 
  int64_t StrideIntVal = StrideWithinGroup;
  if (NeedsWidening) {
    // Continue with checking the "shape" of `SortedOffsetsFromBase`.
    // Check that the strides between groups are all the same.
    unsigned CurrentGroupStartIdx = GroupSize;
    int64_t StrideBetweenGroups =
        SortedOffsetsFromBase[GroupSize] - SortedOffsetsFromBase[0];
    StrideIntVal = StrideBetweenGroups;
    for (; CurrentGroupStartIdx < Sz; CurrentGroupStartIdx += GroupSize) {
      if (SortedOffsetsFromBase[CurrentGroupStartIdx] -
              SortedOffsetsFromBase[CurrentGroupStartIdx - GroupSize] !=
          StrideBetweenGroups)
        return false;
    }
 
    auto CheckGroup = [=](const unsigned StartIdx) -> bool {
      auto Indices = seq<unsigned>(StartIdx + 1, Sz);
      auto FoundIt = llvm::find_if(Indices, IsEndOfGroupIndex);
      unsigned GroupEndIdx = FoundIt != Indices.end() ? *FoundIt : Sz;
      return GroupEndIdx - StartIdx == GroupSize;
    };
    for (unsigned I = 0; I < Sz; I += GroupSize) {
      if (!CheckGroup(I))
        return false;
    }
  }
 
  Type *StrideTy = DL->getIndexType(Ptr0->getType());
  SPtrInfo.StrideVal = ConstantInt::getSigned(StrideTy, StrideIntVal);
  SPtrInfo.Ty = cast<FixedVectorType>(getWidenedType(NewScalarTy, VecSz));
  return true;
}
 
bool BoUpSLP::analyzeRtStrideCandidate(ArrayRef<Value *> PointerOps,
                                       Type *BaseTy, Align CommonAlignment,
                                       SmallVectorImpl<unsigned> &SortedIndices,
                                       StridedPtrInfo &SPtrInfo,
                                       bool IsLoad) const {
  const unsigned Sz = PointerOps.size();
  const unsigned MinProfitableStridedOps =
      IsLoad ? MinProfitableStridedLoads : MinProfitableStridedStores;
  if (Sz * getNumElements(BaseTy) < MinProfitableStridedOps)
    return false;
 
  // The element type of the strided access is fixed by the number of distinct
  // constant offsets the pointers fall into, which has to divide their number.
  auto GetStridedTy = [&](unsigned NumOffsets) -> FixedVectorType * {
    Type *NewScalarTy = BaseTy;
    if (NumOffsets > 1 || BaseTy->isVectorTy())
      NewScalarTy = Type::getIntNTy(
          SE->getContext(),
          DL->getTypeSizeInBits(BaseTy).getFixedValue() * NumOffsets);
    return dyn_cast<FixedVectorType>(
        getWidenedType(NewScalarTy, Sz / NumOffsets));
  };
  auto IsLegalStridedTy = [&](FixedVectorType *StridedTy) {
    return StridedTy && TTI->isTypeLegal(StridedTy) &&
           TTI->isLegalStridedLoadStore(StridedTy, CommonAlignment);
  };
  // Determining the offsets below builds SCEVs for every pointer, which is
  // expensive. Bail out if the target does not support any strided accesses at
  // all.
  if (none_of(seq<unsigned>(1, Sz), [&](unsigned NumOffsets) {
        return Sz % NumOffsets == 0 &&
               IsLegalStridedTy(GetStridedTy(NumOffsets));
      }))
    return false;
 
  // If each value in `PointerOps` is of the form `%x + Offset` where `Offset`
  // is constant, we partition `PointerOps` sequence into subsequences of
  // pointers with the same offset. For each offset we record values from
  // `PointerOps` and their indicies in `PointerOps`.
  SmallDenseMap<int64_t, std::pair<SmallVector<Value *>, SmallVector<unsigned>>>
      OffsetToPointerOpIdxMap;
  // Track to make sure that only VecSz different stride multiples are consumed
  // Prevents cases such as:
  // 1, x + 0, x + 1, 2x + 0 from being recognized as legal RT strided as there
  // are 2 "0" and 2 "1" offsets and a stride of "x" between both offsets
  SmallDenseSet<const SCEV *> StrideMultiples;
  for (auto [Idx, Ptr] : enumerate(PointerOps)) {
    const SCEV *PtrSCEV = SE->getSCEV(Ptr);
    if (!PtrSCEV)
      return false;
 
    const auto *Add = dyn_cast<SCEVAddExpr>(PtrSCEV);
    int64_t Offset = 0;
    const SCEV *StrideMultiple = PtrSCEV;
    if (Add) {
      // `Offset` is non-zero.
      for (int I : seq<int>(Add->getNumOperands())) {
        const auto *SC = dyn_cast<SCEVConstant>(Add->getOperand(I));
        if (!SC)
          continue;
        Offset = SC->getAPInt().getSExtValue();
        if (Offset >= std::numeric_limits<int64_t>::max() - 1) {
          Offset = 0;
          continue;
        }
        StrideMultiple = SE->getMinusSCEV(StrideMultiple, SC);
        break;
      }
    }
    OffsetToPointerOpIdxMap[Offset].first.push_back(Ptr);
    OffsetToPointerOpIdxMap[Offset].second.push_back(Idx);
    StrideMultiples.insert(StrideMultiple);
  }
  unsigned NumOffsets = OffsetToPointerOpIdxMap.size();
 
  // Quick detour: at this point we can say what the type of strided load would
  // be if all the checks pass. Check if this type is legal for the target.
  unsigned VecSz = Sz;
  if (NumOffsets > 1) {
    if (Sz % NumOffsets != 0)
      return false;
    VecSz = Sz / NumOffsets;
  }
 
  if (StrideMultiples.size() != VecSz)
    return false;
 
  auto *StridedLoadTy = GetStridedTy(NumOffsets);
  if (!IsLegalStridedTy(StridedLoadTy))
    return false;
 
  // Check if the offsets are contiguous and that each group has the required
  // size.
  SmallVector<int64_t> SortedOffsetsV(NumOffsets);
  for (auto [Idx, MapPair] : enumerate(OffsetToPointerOpIdxMap)) {
    if (MapPair.second.first.size() != VecSz)
      return false;
    SortedOffsetsV[Idx] = MapPair.first;
  }
  sort(SortedOffsetsV);
 
  if (NumOffsets > 1) {
    int64_t BaseBytes = DL->getTypeStoreSize(BaseTy);
    for (int I : seq<int>(1, SortedOffsetsV.size())) {
      if (SortedOffsetsV[I] - SortedOffsetsV[I - 1] != BaseBytes)
        return false;
    }
  }
 
  // Introduce some notation for the explanations below. Let `PointerOps_j`
  // denote  the subsequence of `PointerOps` with offsets equal to
  // `SortedOffsetsV[j]`. Let `SortedIndices_j` be a such that the sequence
  // ```
  // PointerOps_j[SortedIndices_j[0]],
  // PointerOps_j[SortedIndices_j[1]],
  // PointerOps_j[SortedIndices_j[2]],
  // ...
  // ```
  // is sorted. Also, let `IndicesInAllPointerOps_j` be the vector
  // of indices of the subsequence `PointerOps_j` in all of `PointerOps`,
  // i.e `PointerOps_j[i] = PointerOps[IndicesInAllPointerOps_j[i]]`.
  // The entire sorted `PointerOps` looks like this:
  // ```
  // PointerOps_0[SortedIndices_0[0]] = PointerOps[IndicesInAllPointerOps_0[0]],
  // PointerOps_1[SortedIndices_1[0]] = PointerOps[IndicesInAllPointerOps_1[0]],
  // PointerOps_2[SortedIndices_2[0]] = PointerOps[IndicesInAllPointerOps_2[0]],
  // ...
  // PointerOps_(NumOffsets - 1)[SortedIndices_(NumOffsets - 1)[0]] =
  // PointerOps[IndicesInAllPointerOps_(NumOffsets - 1)[0]],
  //
  // PointerOps_0[SortedIndices_0[1]] = PointerOps[IndicesInAllPointerOps_0[1]],
  // PointerOps_1[SortedIndices_1[1]] = PointerOps[IndicesInAllPointerOps_1[1]],
  // PointerOps_2[SortedIndices_2[1]] = PointerOps[IndicesInAllPointerOps_2[1]],
  // ...
  // PointerOps_(NumOffsets - 1)[SortedIndices_(NumOffsets - 1)[1]] =
  // PointerOps[IndicesInAllPointerOps_(NumOffsets - 1)[1]],
  //
  // PointerOps_0[SortedIndices_0[2]] = PointerOps[IndicesInAllPointerOps_0[2]],
  // PointerOps_1[SortedIndices_1[2]] = PointerOps[IndicesInAllPointerOps_1[2]],
  // PointerOps_2[SortedIndices_2[2]] = PointerOps[IndicesInAllPointerOps_2[2]],
  // ...
  // PointerOps_(NumOffsets - 1)[SortedIndices_(NumOffsets - 1)[2]] =
  // PointerOps[IndicesInAllPointerOps_(NumOffsets - 1)[2]],
  // ...
  // ...
  // ...
  // PointerOps_0[SortedIndices_0[VecSz - 1]] =
  // PointerOps[IndicesInAllPointerOps_0[VecSz - 1]],
  // PointerOps_1[SortedIndices_1[VecSz - 1]] =
  // PointerOps[IndicesInAllPointerOps_1[VecSz - 1]],
  // PointerOps_2[SortedIndices_2[VecSz - 1]] =
  // PointerOps[IndicesInAllPointerOps_2[VecSz - 1]],
  // ...
  // PointerOps_(NumOffsets - 1)[SortedIndices_(NumOffsets - 1)[VecSz - 1]] =
  // PointerOps[IndicesInAllPointerOps_(NumOffsets - 1)[VecSz - 1]],
  // ```
  // In order to be able to generate a strided load, for each `PointerOps_j`
  // check that the distance between adjacent pointers are all equal to the same
  // value (stride).
  //
  // As we do that, also calculate SortedIndices. Since we should not modify
  // `SortedIndices` unless we know that all the checks succeed, record the
  // indicies into `SortedIndicesDraft`.
  SmallVector<unsigned> SortedIndicesDraft(Sz);
 
  // Given sorted indices for a particular offset (as calculated by
  // calculateRtStride), update the `SortedIndicesDraft` for all of PointerOps.
  // Let `Offset` be `SortedOffsetsV[OffsetNum]`.
  // \param `OffsetNum` the index of `Offset` in `SortedOffsetsV`.
  // \param `IndicesInAllPointerOps` vector of indices of the
  // subsequence `PointerOps_OffsetNum` in `PointerOps`, i.e. using the above
  // notation `IndicesInAllPointerOps = IndicesInAllPointerOps_OffsetNum`.
  // \param `SortedIndicesForOffset = SortedIndices_OffsetNum`
  auto UpdateSortedIndices =
      [&](SmallVectorImpl<unsigned> &SortedIndicesForOffset,
          ArrayRef<unsigned> IndicesInAllPointerOps, const int64_t OffsetNum) {
        if (SortedIndicesForOffset.empty()) {
          SortedIndicesForOffset.resize(IndicesInAllPointerOps.size());
          std::iota(SortedIndicesForOffset.begin(),
                    SortedIndicesForOffset.end(), 0);
        }
        for (const auto [Num, Idx] : enumerate(SortedIndicesForOffset)) {
          SortedIndicesDraft[Num * NumOffsets + OffsetNum] =
              IndicesInAllPointerOps[Idx];
        }
      };
 
  int64_t LowestOffset = SortedOffsetsV[0];
  ArrayRef<Value *> PointerOps0 = OffsetToPointerOpIdxMap[LowestOffset].first;
 
  SmallVector<unsigned> SortedIndicesForOffset0;
  const SCEV *Stride0 =
      calculateRtStride(PointerOps0, BaseTy, *DL, *SE, SortedIndicesForOffset0);
  if (!Stride0)
    return false;
 
  ArrayRef<unsigned> IndicesInAllPointerOps0 =
      OffsetToPointerOpIdxMap[LowestOffset].second;
  UpdateSortedIndices(SortedIndicesForOffset0, IndicesInAllPointerOps0, 0);
 
  // Now that we know what the common stride and coefficients has to be check
  // the remaining `PointerOps_j`.
  SmallVector<unsigned> SortedIndicesForOffset;
  for (int J : seq<int>(1, NumOffsets)) {
    SortedIndicesForOffset.clear();
 
    int64_t Offset = SortedOffsetsV[J];
    ArrayRef<Value *> PointerOpsForOffset =
        OffsetToPointerOpIdxMap[Offset].first;
    ArrayRef<unsigned> IndicesInAllPointerOps =
        OffsetToPointerOpIdxMap[Offset].second;
    const SCEV *StrideWithinGroup = calculateRtStride(
        PointerOpsForOffset, BaseTy, *DL, *SE, SortedIndicesForOffset);
 
    if (!StrideWithinGroup || StrideWithinGroup != Stride0)
      return false;
 
    UpdateSortedIndices(SortedIndicesForOffset, IndicesInAllPointerOps, J);
  }
 
  SortedIndices.clear();
  SortedIndices = std::move(SortedIndicesDraft);
  SPtrInfo.StrideSCEV = Stride0;
  SPtrInfo.Ty = StridedLoadTy;
  return true;
}
 
BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
    ArrayRef<Value *> VL, const Value *VL0, SmallVectorImpl<unsigned> &Order,
    SmallVectorImpl<Value *> &PointerOps, StridedPtrInfo &SPtrInfo,
    unsigned *BestVF, bool TryRecursiveCheck) const {
  // Check that a vectorized load would load the same memory as a scalar
  // load. For example, we don't want to vectorize loads that are smaller
  // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM
  // treats loading/storing it as an i8 struct. If we vectorize loads/stores
  // from such a struct, we read/write packed bits disagreeing with the
  // unvectorized version.
  if (BestVF)
    *BestVF = 0;
  if (areKnownNonVectorizableLoads(VL))
    return LoadsState::Gather;
  Type *ScalarTy = VL0->getType();
 
  if (DL->getTypeSizeInBits(ScalarTy) != DL->getTypeAllocSizeInBits(ScalarTy))
    return LoadsState::Gather;
 
  // Make sure all loads in the bundle are simple - we can't vectorize
  // atomic or volatile loads.
  PointerOps.clear();
  const size_t Sz = VL.size();
  PointerOps.resize(Sz);
  auto *POIter = PointerOps.begin();
  for (Value *V : VL) {
    auto *L = dyn_cast<LoadInst>(V);
    if (!L || !L->isSimple())
      return LoadsState::Gather;
    *POIter = L->getPointerOperand();
    ++POIter;
  }
 
  Order.clear();
  // Check the order of pointer operands or that all pointers are the same.
  bool IsSorted = sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, Order);
 
  auto *VecTy = dyn_cast<VectorType>(getWidenedType(ScalarTy, Sz));
  if (!VecTy)
    return LoadsState::Gather;
  Align CommonAlignment = computeCommonAlignment<LoadInst>(VL);
  // Cache masked gather legality - both the !IsSorted path below and the
  // post-branch check use the same VecTy/CommonAlignment, and the underlying
  // TTI calls are virtual.
  std::optional<bool> MaskedGatherLegal;
  auto IsMaskedGatherLegal = [&] {
    if (!MaskedGatherLegal)
      MaskedGatherLegal =
          TTI->isLegalMaskedGather(VecTy, CommonAlignment) &&
          !TTI->forceScalarizeMaskedGather(VecTy, CommonAlignment);
    return *MaskedGatherLegal;
  };
  if (!IsSorted) {
    // Check for a group of loads, each selecting its address (directly, or
    // via a constant-offset GEP) between the same two candidate base
    // pointers - the shape if-converted, fully-unrolled loop bodies of the
    // form `x = cond ? A[i] : B[i]` take. If found, model it as two masked
    // loads (one per candidate) blended by the (vectorized) condition,
    // rather than falling back to a gather of the individual scalar loads.
    Value *TrueBase = nullptr;
    Value *FalseBase = nullptr;
    SmallVector<Value *> Conditions;
    if (isSelectedBaseLoad(ScalarTy, PointerOps, *DL, TrueBase, FalseBase,
                           Conditions) &&
        TTI->isLegalMaskedLoad(VecTy, CommonAlignment,
                               cast<LoadInst>(VL0)->getPointerAddressSpace()))
      return LoadsState::BlendedLoadVectorize;
 
    if (analyzeRtStrideCandidate(PointerOps, ScalarTy, CommonAlignment, Order,
                                 SPtrInfo, /*isLoad=*/true))
      return LoadsState::StridedVectorize;
 
    if (!IsMaskedGatherLegal())
      return LoadsState::Gather;
 
    if (!all_of(PointerOps, [&](Value *P) {
          return arePointersCompatible(P, PointerOps.front(), *TLI);
        }))
      return LoadsState::Gather;
 
  } else {
    Value *Ptr0;
    Value *PtrN;
    if (Order.empty()) {
      Ptr0 = PointerOps.front();
      PtrN = PointerOps.back();
    } else {
      Ptr0 = PointerOps[Order.front()];
      PtrN = PointerOps[Order.back()];
    }
    // sortPtrAccesses validates getPointersDiff for all pointers relative to
    // PointerOps[0], so compute the span using PointerOps[0] as intermediate:
    //   Diff = offset(PtrN) - offset(Ptr0) relative to PointerOps[0]
    std::optional<int64_t> Diff0 =
        getPointersDiff(ScalarTy, PointerOps[0], ScalarTy, Ptr0, *DL, *SE);
    std::optional<int64_t> DiffN =
        getPointersDiff(ScalarTy, PointerOps[0], ScalarTy, PtrN, *DL, *SE);
    assert(Diff0 && DiffN &&
           "sortPtrAccesses should have validated these pointers");
    int64_t Diff = *DiffN - *Diff0;
    // Check that the sorted loads are consecutive.
    if (static_cast<uint64_t>(Diff) == Sz - 1)
      return LoadsState::Vectorize;
    if (isMaskedLoadCompress(VL, PointerOps, Order, *TTI, *DL, *SE, *AC, *DT,
                             *TLI, CostKind, [&](Value *V) {
                               return areAllUsersVectorized(
                                   cast<Instruction>(V), UserIgnoreList);
                             }))
      return LoadsState::CompressVectorize;
    Align Alignment =
        cast<LoadInst>(Order.empty() ? VL.front() : VL[Order.front()])
            ->getAlign();
    if (analyzeConstantStrideCandidate(PointerOps, ScalarTy, Alignment, Order,
                                       Diff, Ptr0, SPtrInfo))
      return LoadsState::StridedVectorize;
  }
  if (!IsMaskedGatherLegal())
    return LoadsState::Gather;
  // Correctly identify compare the cost of loads + shuffles rather than
  // strided/masked gather loads. Returns true if vectorized + shuffles
  // representation is better than just gather.
  auto CheckForShuffledLoads = [&, &TTI = *TTI](Align CommonAlignment,
                                                unsigned *BestVF,
                                                bool ProfitableGatherPointers) {
    if (BestVF)
      *BestVF = 0;
    // Compare masked gather cost and loads + insert subvector costs.
    auto [ScalarGEPCost, VectorGEPCost] =
        getGEPCosts(TTI, PointerOps, PointerOps.front(), Instruction::Load,
                    CostKind, ScalarTy, VecTy);
    // Estimate the cost of masked gather GEP. If not a splat, roughly
    // estimate as a buildvector, otherwise estimate as splat.
    APInt DemandedElts = APInt::getAllOnes(Sz);
    Type *PtrScalarTy = PointerOps.front()->getType()->getScalarType();
    auto *PtrVecTy = cast<VectorType>(getWidenedType(PtrScalarTy, Sz));
    // Cache the underlying object of PointerOps.front() - it is invariant
    // across the per-V comparisons below and getUnderlyingObject walks
    // GEP/cast chains.
    const Value *FrontUO = getUnderlyingObject(PointerOps.front());
    if (static_cast<unsigned>(count_if(
            PointerOps, IsaPred<GetElementPtrInst>)) < PointerOps.size() - 1 ||
        any_of(PointerOps,
               [&](Value *V) { return getUnderlyingObject(V) != FrontUO; }))
      VectorGEPCost += getScalarizationOverhead(
          TTI, SLPReVec, PtrScalarTy, PtrVecTy, DemandedElts, /*Insert=*/true,
          /*Extract=*/false, CostKind);
    else
      VectorGEPCost +=
          getScalarizationOverhead(
              TTI, SLPReVec, PtrScalarTy, PtrVecTy, APInt::getOneBitSet(Sz, 0),
              /*Insert=*/true, /*Extract=*/false, CostKind) +
          getShuffleCost(TTI, TTI::SK_Broadcast, PtrVecTy, CostKind);
    // The cost of scalar loads.
    InstructionCost ScalarLoadsCost =
        accumulate(VL, InstructionCost(),
                   [&](InstructionCost C, Value *V) {
                     return C + TTI.getInstructionCost(cast<Instruction>(V),
                                                       CostKind);
                   }) +
        ScalarGEPCost;
    // The cost of masked gather.
    InstructionCost MaskedGatherCost =
        TTI.getMemIntrinsicInstrCost(
            MemIntrinsicCostAttributes(Intrinsic::masked_gather, VecTy,
                                       cast<LoadInst>(VL0)->getPointerOperand(),
                                       /*VariableMask=*/false, CommonAlignment),
            CostKind) +
        (ProfitableGatherPointers ? 0 : VectorGEPCost);
    InstructionCost GatherCost =
        getScalarizationOverhead(TTI, SLPReVec, ScalarTy, VecTy, DemandedElts,
                                 /*Insert=*/true,
                                 /*Extract=*/false, CostKind) +
        ScalarLoadsCost;
    // The list of loads is small or perform partial check already - directly
    // compare masked gather cost and gather cost.
    constexpr unsigned ListLimit = 4;
    if (!TryRecursiveCheck || VL.size() < ListLimit)
      return MaskedGatherCost - GatherCost >= -SLPCostThreshold;
 
    unsigned Sz = DL->getTypeSizeInBits(ScalarTy);
    unsigned MinVF = getMinVF(2 * Sz);
    DemandedElts.clearAllBits();
    // Iterate through possible vectorization factors and check if vectorized +
    // shuffles is better than just gather.
    for (unsigned VF = getFloorFullVectorNumberOfElements(
             TTI, ScalarTy, VL.size() - 1, SLPReVec);
         VF >= MinVF; VF = getFloorFullVectorNumberOfElements(
                          TTI, ScalarTy, VF - 1, SLPReVec)) {
      SmallVector<std::pair<unsigned, LoadsState>> States;
      for (unsigned Cnt = 0, End = VL.size(); Cnt < End; Cnt += VF) {
        const unsigned SliceVF = std::min(VF, End - Cnt);
        ArrayRef<Value *> Slice = VL.slice(Cnt, SliceVF);
        SmallVector<unsigned> Order;
        SmallVector<Value *> PointerOps;
        LoadsState LS = canVectorizeLoads(Slice, Slice.front(), Order,
                                          PointerOps, SPtrInfo, BestVF,
                                          /*TryRecursiveCheck=*/false);
        // Check that the sorted loads are consecutive.
        if (LS == LoadsState::Gather) {
          if (BestVF) {
            DemandedElts.setAllBits();
            break;
          }
          DemandedElts.setBits(Cnt, Cnt + SliceVF);
          continue;
        }
        // If need the reorder - consider as high-cost masked gather for now.
        if ((LS == LoadsState::Vectorize ||
             LS == LoadsState::StridedVectorize ||
             LS == LoadsState::CompressVectorize) &&
            !Order.empty() && !isReverseOrder(Order))
          LS = LoadsState::ScatterVectorize;
        States.emplace_back(Cnt, LS);
      }
      if (DemandedElts.isAllOnes())
        // All loads gathered - try smaller VF.
        continue;
      // Can be vectorized later as a serie of loads/insertelements.
      InstructionCost VecLdCost = 0;
      if (!DemandedElts.isZero()) {
        VecLdCost = getScalarizationOverhead(TTI, SLPReVec, ScalarTy, VecTy,
                                             DemandedElts,
                                             /*Insert=*/true,
                                             /*Extract=*/false, CostKind) +
                    ScalarGEPCost;
        for (unsigned Idx : seq<unsigned>(VL.size()))
          if (DemandedElts[Idx])
            VecLdCost +=
                TTI.getInstructionCost(cast<Instruction>(VL[Idx]), CostKind);
      }
      for (const auto &[SliceStart, LS] : States) {
        const unsigned SliceVF = std::min<unsigned>(VF, VL.size() - SliceStart);
        auto *SubVecTy = cast<VectorType>(getWidenedType(ScalarTy, SliceVF));
        auto *LI0 = cast<LoadInst>(VL[SliceStart]);
        InstructionCost VectorGEPCost =
            (LS == LoadsState::ScatterVectorize && ProfitableGatherPointers)
                ? 0
                : getGEPCosts(TTI,
                              ArrayRef(PointerOps).slice(SliceStart, SliceVF),
                              LI0->getPointerOperand(), Instruction::Load,
                              CostKind, ScalarTy, SubVecTy)
                      .second;
        if (LS == LoadsState::ScatterVectorize) {
          if (static_cast<unsigned>(
                  count_if(PointerOps, IsaPred<GetElementPtrInst>)) <
                  PointerOps.size() - 1 ||
              any_of(PointerOps, [&](Value *V) {
                return getUnderlyingObject(V) != FrontUO;
              }))
            VectorGEPCost += getScalarizationOverhead(
                TTI, SLPReVec, ScalarTy, SubVecTy, APInt::getAllOnes(SliceVF),
                /*Insert=*/true, /*Extract=*/false, CostKind);
          else
            VectorGEPCost +=
                getScalarizationOverhead(TTI, SLPReVec, ScalarTy, SubVecTy,
                                         APInt::getOneBitSet(SliceVF, 0),
                                         /*Insert=*/true, /*Extract=*/false,
                                         CostKind) +
                getShuffleCost(TTI, TTI::SK_Broadcast, SubVecTy, CostKind);
        }
        switch (LS) {
        case LoadsState::Vectorize:
          VecLdCost +=
              TTI.getMemoryOpCost(Instruction::Load, SubVecTy, LI0->getAlign(),
                                  LI0->getPointerAddressSpace(), CostKind,
                                  TTI::OperandValueInfo()) +
              VectorGEPCost;
          break;
        case LoadsState::StridedVectorize:
          VecLdCost += TTI.getMemIntrinsicInstrCost(
                           MemIntrinsicCostAttributes(
                               Intrinsic::experimental_vp_strided_load,
                               SubVecTy, LI0->getPointerOperand(),
                               /*VariableMask=*/false, CommonAlignment),
                           CostKind) +
                       VectorGEPCost;
          break;
        case LoadsState::CompressVectorize:
          VecLdCost +=
              TTI.getMemIntrinsicInstrCost(
                  MemIntrinsicCostAttributes(Intrinsic::masked_load, SubVecTy,
                                             CommonAlignment,
                                             LI0->getPointerAddressSpace()),
                  CostKind) +
              getShuffleCost(TTI, TTI::SK_PermuteSingleSrc, SubVecTy, CostKind);
          break;
        case LoadsState::ScatterVectorize:
          VecLdCost += TTI.getMemIntrinsicInstrCost(
                           MemIntrinsicCostAttributes(
                               Intrinsic::masked_gather, SubVecTy,
                               LI0->getPointerOperand(),
                               /*VariableMask=*/false, CommonAlignment),
                           CostKind) +
                       VectorGEPCost;
          break;
        case LoadsState::BlendedLoadVectorize:
          // Two masked loads (one per candidate base) plus a select; no address
          // vector is materialized, so VectorGEPCost is skipped.
          VecLdCost +=
              getBlendedLoadCost(TTI, SubVecTy, CommonAlignment,
                                 LI0->getPointerAddressSpace(), CostKind);
          break;
        case LoadsState::Gather:
          llvm_unreachable("Gathers are not added to States");
        }
        SmallVector<int> ShuffleMask(VL.size());
        const unsigned SliceIdx = SliceStart / VF;
        for (int Idx : seq<int>(VL.size()))
          ShuffleMask[Idx] = Idx / VF == SliceIdx ? VL.size() + Idx % VF : Idx;
        if (SliceStart > 0)
          VecLdCost +=
              getShuffleCost(TTI, TTI::SK_InsertSubvector, VecTy, CostKind,
                             ShuffleMask, SliceStart, SubVecTy);
      }
      // If masked gather cost is higher - better to vectorize, so
      // consider it as a gather node. It will be better estimated
      // later.
      if (MaskedGatherCost >= VecLdCost &&
          VecLdCost - GatherCost < -SLPCostThreshold) {
        if (BestVF)
          *BestVF = VF;
        return true;
      }
    }
    return MaskedGatherCost - GatherCost >= -SLPCostThreshold;
  };
  // TODO: need to improve analysis of the pointers, if not all of them are
  // GEPs or have > 2 operands, we end up with a gather node, which just
  // increases the cost.
  Loop *L = LI->getLoopFor(cast<LoadInst>(VL0)->getParent());
  bool ProfitableGatherPointers =
      L && Sz > 2 && static_cast<unsigned>(count_if(PointerOps, [L](Value *V) {
                       return L->isLoopInvariant(V);
                     })) <= Sz / 2;
  if (ProfitableGatherPointers || all_of(PointerOps, [](Value *P) {
        auto *GEP = dyn_cast<GetElementPtrInst>(P);
        return (!GEP && doesNotNeedToBeScheduled(P)) ||
               (GEP && GEP->getNumOperands() == 2 &&
                isa<Constant, Instruction>(GEP->getOperand(1)));
      })) {
    // Check if potential masked gather can be represented as series
    // of loads + insertsubvectors.
    // If masked gather cost is higher - better to vectorize, so
    // consider it as a gather node. It will be better estimated
    // later.
    if (!TryRecursiveCheck || !CheckForShuffledLoads(CommonAlignment, BestVF,
                                                     ProfitableGatherPointers))
      return LoadsState::ScatterVectorize;
  }
 
  return LoadsState::Gather;
}
 
static bool clusterSortPtrAccesses(ArrayRef<Value *> VL,
                                   ArrayRef<BasicBlock *> BBs, Type *ElemTy,
                                   const DataLayout &DL, ScalarEvolution &SE,
                                   SmallVectorImpl<unsigned> &SortedIndices) {
  assert(
      all_of(VL, [](const Value *V) { return V->getType()->isPointerTy(); }) &&
      "Expected list of pointer operands.");
  // Map from bases to a vector of (Ptr, Offset, OrigIdx), which we insert each
  // Ptr into, sort and return the sorted indices with values next to one
  // another.
  SmallMapVector<
      std::pair<BasicBlock *, Value *>,
      SmallVector<SmallVector<std::tuple<Value *, int64_t, unsigned>>>, 8>
      Bases;
  Bases
      .try_emplace(std::make_pair(
          BBs.front(), getUnderlyingObject(VL.front(), RecursionMaxDepth)))
      .first->second.emplace_back().emplace_back(VL.front(), 0U, 0U);
 
  SortedIndices.clear();
  for (auto [Cnt, Ptr] : enumerate(VL.drop_front())) {
    auto Key = std::make_pair(BBs[Cnt + 1],
                              getUnderlyingObject(Ptr, RecursionMaxDepth));
    bool Found = any_of(Bases.try_emplace(Key).first->second,
                        [&, &Cnt = Cnt, &Ptr = Ptr](auto &Base) {
                          std::optional<int64_t> Diff =
                              getPointersDiff(ElemTy, std::get<0>(Base.front()),
                                              ElemTy, Ptr, DL, SE,
                                              /*StrictCheck=*/true);
                          if (!Diff)
                            return false;
 
                          Base.emplace_back(Ptr, *Diff, Cnt + 1);
                          return true;
                        });
 
    if (!Found) {
      // If we haven't found enough to usefully cluster, return early.
      if (Bases.size() > VL.size() / 2 - 1)
        return false;
 
      // Not found already - add a new Base
      Bases.find(Key)->second.emplace_back().emplace_back(Ptr, 0, Cnt + 1);
    }
  }
 
  if (Bases.size() == VL.size())
    return false;
 
  if (Bases.size() == 1 && (Bases.front().second.size() == 1 ||
                            Bases.front().second.size() == VL.size()))
    return false;
 
  // For each of the bases sort the pointers by Offset and check if any of the
  // base become consecutively allocated.
  auto ComparePointers = [](Value *Ptr1, Value *Ptr2) {
    SmallPtrSet<Value *, 13> FirstPointers;
    SmallPtrSet<Value *, 13> SecondPointers;
    Value *P1 = Ptr1;
    Value *P2 = Ptr2;
    unsigned Depth = 0;
    while (!FirstPointers.contains(P2) && !SecondPointers.contains(P1)) {
      if (P1 == P2 || Depth > RecursionMaxDepth)
        return false;
      FirstPointers.insert(P1);
      SecondPointers.insert(P2);
      P1 = getUnderlyingObject(P1, /*MaxLookup=*/1);
      P2 = getUnderlyingObject(P2, /*MaxLookup=*/1);
      ++Depth;
    }
    assert((FirstPointers.contains(P2) || SecondPointers.contains(P1)) &&
           "Unable to find matching root.");
    return FirstPointers.contains(P2) && !SecondPointers.contains(P1);
  };
  for (auto &Base : Bases) {
    for (auto &Vec : Base.second) {
      if (Vec.size() > 1) {
        stable_sort(Vec, llvm::less_second());
        int64_t InitialOffset = std::get<1>(Vec[0]);
        bool AnyConsecutive =
            all_of(enumerate(Vec), [InitialOffset](const auto &P) {
              return std::get<1>(P.value()) ==
                     int64_t(P.index()) + InitialOffset;
            });
        // Fill SortedIndices array only if it looks worth-while to sort the
        // ptrs.
        if (!AnyConsecutive)
          return false;
      }
    }
    stable_sort(Base.second, [&](const auto &V1, const auto &V2) {
      return ComparePointers(std::get<0>(V1.front()), std::get<0>(V2.front()));
    });
  }
 
  for (auto &T : Bases)
    for (const auto &Vec : T.second)
      for (const auto &P : Vec)
        SortedIndices.push_back(std::get<2>(P));
 
  assert(SortedIndices.size() == VL.size() &&
         "Expected SortedIndices to be the size of VL");
  return true;
}
 
std::optional<BoUpSLP::OrdersType>
BoUpSLP::findPartiallyOrderedLoads(const BoUpSLP::TreeEntry &TE) {
  assert(TE.isGather() && "Expected gather node only.");
  Type *ScalarTy = TE.Scalars[0]->getType();
 
  SmallVector<Value *> Ptrs;
  Ptrs.reserve(TE.Scalars.size());
  SmallVector<BasicBlock *> BBs;
  BBs.reserve(TE.Scalars.size());
  for (Value *V : TE.Scalars) {
    auto *L = dyn_cast<LoadInst>(V);
    if (!L || !L->isSimple())
      return std::nullopt;
    Ptrs.push_back(L->getPointerOperand());
    BBs.push_back(L->getParent());
  }
 
  BoUpSLP::OrdersType Order;
  if (!LoadEntriesToVectorize.contains(TE.Idx) &&
      clusterSortPtrAccesses(Ptrs, BBs, ScalarTy, *DL, *SE, Order))
    return std::move(Order);
  return std::nullopt;
}
 
/// Check if two insertelement instructions are from the same buildvector.
static bool areTwoInsertFromSameBuildVector(
    InsertElementInst *VU, InsertElementInst *V,
    function_ref<Value *(InsertElementInst *)> GetBaseOperand) {
  // Instructions must be from the same basic blocks.
  if (VU->getParent() != V->getParent())
    return false;
  // Checks if 2 insertelements are from the same buildvector.
  if (VU->getType() != V->getType())
    return false;
  // Multiple used inserts are separate nodes.
  if (!VU->hasOneUse() && !V->hasOneUse())
    return false;
  auto *IE1 = VU;
  auto *IE2 = V;
  std::optional<unsigned> Idx1 = getElementIndex(IE1);
  std::optional<unsigned> Idx2 = getElementIndex(IE2);
  if (Idx1 == std::nullopt || Idx2 == std::nullopt)
    return false;
  // Go through the vector operand of insertelement instructions trying to find
  // either VU as the original vector for IE2 or V as the original vector for
  // IE1.
  SmallBitVector ReusedIdx(VU->getType()->getElementCount().getKnownMinValue());
  bool IsReusedIdx = false;
  do {
    if (IE2 == VU && !IE1)
      return VU->hasOneUse();
    if (IE1 == V && !IE2)
      return V->hasOneUse();
    if (IE1 && IE1 != V) {
      unsigned Idx1 = getElementIndex(IE1).value_or(*Idx2);
      IsReusedIdx |= ReusedIdx.test(Idx1);
      ReusedIdx.set(Idx1);
      if ((IE1 != VU && !IE1->hasOneUse()) || IsReusedIdx)
        IE1 = nullptr;
      else
        IE1 = dyn_cast_or_null<InsertElementInst>(GetBaseOperand(IE1));
    }
    if (IE2 && IE2 != VU) {
      unsigned Idx2 = getElementIndex(IE2).value_or(*Idx1);
      IsReusedIdx |= ReusedIdx.test(Idx2);
      ReusedIdx.set(Idx2);
      if ((IE2 != V && !IE2->hasOneUse()) || IsReusedIdx)
        IE2 = nullptr;
      else
        IE2 = dyn_cast_or_null<InsertElementInst>(GetBaseOperand(IE2));
    }
  } while (!IsReusedIdx && (IE1 || IE2));
  return false;
}
 
std::optional<BoUpSLP::OrdersType>
BoUpSLP::getReorderingData(const TreeEntry &TE, bool TopToBottom,
                           bool IgnoreReorder) {
  // No need to reorder if need to shuffle reuses, still need to shuffle the
  // node.
  if (!TE.ReuseShuffleIndices.empty()) {
    if (isSplat(TE.Scalars))
      return std::nullopt;
    // Check if reuse shuffle indices can be improved by reordering.
    // For this, check that reuse mask is "clustered", i.e. each scalar values
    // is used once in each submask of size <number_of_scalars>.
    // Example: 4 scalar values.
    // ReuseShuffleIndices mask: 0, 1, 2, 3, 3, 2, 0, 1 - clustered.
    //                           0, 1, 2, 3, 3, 3, 1, 0 - not clustered, because
    //                           element 3 is used twice in the second submask.
    unsigned Sz = TE.Scalars.size();
    if (TE.isGather()) {
      if (std::optional<OrdersType> CurrentOrder =
              findReusedOrderedScalars(TE, TopToBottom, IgnoreReorder)) {
        SmallVector<int> Mask;
        fixupOrderingIndices(*CurrentOrder);
        inversePermutation(*CurrentOrder, Mask);
        addMask(Mask, TE.ReuseShuffleIndices);
        OrdersType Res(TE.getVectorFactor(), TE.getVectorFactor());
        unsigned Sz = TE.Scalars.size();
        for (int K = 0, E = TE.getVectorFactor() / Sz; K < E; ++K) {
          for (auto [I, Idx] : enumerate(ArrayRef(Mask).slice(K * Sz, Sz)))
            if (Idx != PoisonMaskElem)
              Res[Idx + K * Sz] = I + K * Sz;
        }
        return std::move(Res);
      }
    }
    if (Sz == 2 && TE.getVectorFactor() == 4 &&
        ::getNumberOfParts(
            *TTI,
            getWidenedType(getValueType(TE.Scalars.front(), SLPReVec),
                           2 * TE.getVectorFactor()),
            getValueType(TE.Scalars.front(), SLPReVec)) == 1)
      return std::nullopt;
    if (TE.ReuseShuffleIndices.size() % Sz != 0)
      return std::nullopt;
    if (!ShuffleVectorInst::isOneUseSingleSourceMask(TE.ReuseShuffleIndices,
                                                     Sz)) {
      SmallVector<int> ReorderMask(Sz, PoisonMaskElem);
      if (TE.ReorderIndices.empty())
        std::iota(ReorderMask.begin(), ReorderMask.end(), 0);
      else
        inversePermutation(TE.ReorderIndices, ReorderMask);
      addMask(ReorderMask, TE.ReuseShuffleIndices);
      unsigned VF = ReorderMask.size();
      OrdersType ResOrder(VF, VF);
      unsigned NumParts = divideCeil(VF, Sz);
      SmallBitVector UsedVals(NumParts);
      for (unsigned I = 0; I < VF; I += Sz) {
        int Val = PoisonMaskElem;
        unsigned UndefCnt = 0;
        unsigned Limit = std::min(Sz, VF - I);
        if (any_of(ArrayRef(ReorderMask).slice(I, Limit),
                   [&](int Idx) {
                     if (Val == PoisonMaskElem && Idx != PoisonMaskElem)
                       Val = Idx;
                     if (Idx == PoisonMaskElem)
                       ++UndefCnt;
                     return Idx != PoisonMaskElem && Idx != Val;
                   }) ||
            Val >= static_cast<int>(NumParts) || Val == PoisonMaskElem ||
            UsedVals.test(Val) || UndefCnt > Sz / 2)
          return std::nullopt;
        UsedVals.set(Val);
        for (unsigned K = 0; K < NumParts; ++K) {
          unsigned Idx = Val + Sz * K;
          if (Idx < VF && I + K < VF)
            ResOrder[Idx] = I + K;
        }
      }
      return std::move(ResOrder);
    }
    unsigned VF = TE.getVectorFactor();
    // Try build correct order for extractelement instructions.
    SmallVector<int> ReusedMask(TE.ReuseShuffleIndices.begin(),
                                TE.ReuseShuffleIndices.end());
    if (TE.hasState() && TE.getOpcode() == Instruction::ExtractElement &&
        !TE.hasCopyableElements() && all_of(TE.Scalars, [Sz](Value *V) {
          if (isa<PoisonValue>(V))
            return true;
          std::optional<unsigned> Idx = getExtractIndex(cast<Instruction>(V));
          return Idx && *Idx < Sz;
        })) {
      assert(!TE.isAltShuffle() && "Alternate instructions are only supported "
                                   "by BinaryOperator and CastInst.");
      SmallVector<int> ReorderMask(Sz, PoisonMaskElem);
      if (TE.ReorderIndices.empty())
        std::iota(ReorderMask.begin(), ReorderMask.end(), 0);
      else
        inversePermutation(TE.ReorderIndices, ReorderMask);
      for (unsigned I = 0; I < VF; ++I) {
        int &Idx = ReusedMask[I];
        if (Idx == PoisonMaskElem)
          continue;
        Value *V = TE.Scalars[ReorderMask[Idx]];
        std::optional<unsigned> EI = getExtractIndex(cast<Instruction>(V));
        Idx = std::distance(ReorderMask.begin(), find(ReorderMask, *EI));
      }
    }
    // Build the order of the VF size, need to reorder reuses shuffles, they are
    // always of VF size.
    OrdersType ResOrder(VF);
    std::iota(ResOrder.begin(), ResOrder.end(), 0);
    auto *It = ResOrder.begin();
    for (unsigned K = 0; K < VF; K += Sz) {
      OrdersType CurrentOrder(TE.ReorderIndices);
      SmallVector<int> SubMask{ArrayRef(ReusedMask).slice(K, Sz)};
      if (SubMask.front() == PoisonMaskElem)
        std::iota(SubMask.begin(), SubMask.end(), 0);
      reorderOrder(CurrentOrder, SubMask);
      transform(CurrentOrder, It, [K](unsigned Pos) { return Pos + K; });
      std::advance(It, Sz);
    }
    if (TE.isGather() && all_of(enumerate(ResOrder), [](const auto &Data) {
          return Data.index() == Data.value();
        }))
      return std::nullopt; // No need to reorder.
    return std::move(ResOrder);
  }
  if (TE.State == TreeEntry::StridedVectorize && !TopToBottom &&
      (!TE.UserTreeIndex || !TE.UserTreeIndex.UserTE->hasState() ||
       !Instruction::isBinaryOp(TE.UserTreeIndex.UserTE->getOpcode())) &&
      (TE.ReorderIndices.empty() || isReverseOrder(TE.ReorderIndices)))
    return std::nullopt;
  if (TE.State == TreeEntry::SplitVectorize ||
      ((TE.State == TreeEntry::Vectorize ||
        TE.State == TreeEntry::StridedVectorize ||
        TE.State == TreeEntry::ExpandVectorize ||
        TE.State == TreeEntry::CompressVectorize ||
        TE.State == TreeEntry::BlendedLoadVectorize) &&
       (isa<LoadInst, ExtractElementInst, ExtractValueInst>(TE.getMainOp()) ||
        (TopToBottom && isa<StoreInst, InsertElementInst, InsertValueInst>(
                            TE.getMainOp()))))) {
    assert((TE.State == TreeEntry::SplitVectorize || !TE.isAltShuffle()) &&
           "Alternate instructions are only supported by "
           "BinaryOperator and CastInst.");
    return TE.ReorderIndices;
  }
  if (!TopToBottom && IgnoreReorder && TE.State == TreeEntry::Vectorize &&
      TE.isAltShuffle()) {
    assert(TE.ReuseShuffleIndices.empty() &&
           "ReuseShuffleIndices should be "
           "empty for alternate instructions.");
    SmallVector<int> Mask;
    TE.buildAltOpShuffleMask(
        [&](Instruction *I) {
          assert(TE.getMatchingMainOpOrAltOp(I) &&
                 "Unexpected main/alternate opcode");
          return isAlternateInstruction(I, TE.getMainOp(), TE.getAltOp(), *TLI);
        },
        Mask);
    const int VF = TE.getVectorFactor();
    OrdersType ResOrder(VF, VF);
    for (unsigned I : seq<unsigned>(VF)) {
      if (Mask[I] == PoisonMaskElem)
        continue;
      ResOrder[Mask[I] % VF] = I;
    }
    return std::move(ResOrder);
  }
  if (!TE.ReorderIndices.empty())
    return TE.ReorderIndices;
  if (TE.State == TreeEntry::Vectorize && TE.getOpcode() == Instruction::PHI) {
    if (!TE.ReorderIndices.empty())
      return TE.ReorderIndices;
 
    SmallVector<Instruction *> UserBVHead(TE.Scalars.size());
    for (auto [I, V] : zip(UserBVHead, TE.Scalars)) {
      if (isa<Constant>(V) || !V->hasNUsesOrMore(1))
        continue;
      auto *II = dyn_cast<InsertElementInst>(*V->user_begin());
      if (!II)
        continue;
      Instruction *BVHead = nullptr;
      BasicBlock *BB = II->getParent();
      while (II && II->hasOneUse() && II->getParent() == BB) {
        BVHead = II;
        II = dyn_cast<InsertElementInst>(II->getOperand(0));
      }
      I = BVHead;
    }
 
    auto CompareByBasicBlocks = [&](BasicBlock *BB1, BasicBlock *BB2) {
      assert(BB1 != BB2 && "Expected different basic blocks.");
      if (!DT->isReachableFromEntry(BB1))
        return false;
      if (!DT->isReachableFromEntry(BB2))
        return true;
      auto *NodeA = DT->getNode(BB1);
      auto *NodeB = DT->getNode(BB2);
      assert(NodeA && "Should only process reachable instructions");
      assert(NodeB && "Should only process reachable instructions");
      assert((NodeA == NodeB) ==
                 (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) &&
             "Different nodes should have different DFS numbers");
      return NodeA->getDFSNumIn() < NodeB->getDFSNumIn();
    };
    auto PHICompare = [&](unsigned I1, unsigned I2) {
      Value *V1 = TE.Scalars[I1];
      Value *V2 = TE.Scalars[I2];
      if (V1 == V2 || (V1->use_empty() && V2->use_empty()))
        return false;
      if (isa<PoisonValue>(V1))
        return true;
      if (isa<PoisonValue>(V2))
        return false;
      if (V1->getNumUses() < V2->getNumUses())
        return true;
      if (V1->getNumUses() > V2->getNumUses())
        return false;
      auto *FirstUserOfPhi1 = cast<Instruction>(*V1->user_begin());
      auto *FirstUserOfPhi2 = cast<Instruction>(*V2->user_begin());
      if (FirstUserOfPhi1->getParent() != FirstUserOfPhi2->getParent())
        return CompareByBasicBlocks(FirstUserOfPhi1->getParent(),
                                    FirstUserOfPhi2->getParent());
      auto *IE1 = dyn_cast<InsertElementInst>(FirstUserOfPhi1);
      auto *IE2 = dyn_cast<InsertElementInst>(FirstUserOfPhi2);
      auto *EE1 = dyn_cast<ExtractElementInst>(FirstUserOfPhi1);
      auto *EE2 = dyn_cast<ExtractElementInst>(FirstUserOfPhi2);
      if (IE1 && !IE2)
        return true;
      if (!IE1 && IE2)
        return false;
      if (IE1 && IE2) {
        if (UserBVHead[I1] && !UserBVHead[I2])
          return true;
        if (!UserBVHead[I1])
          return false;
        if (UserBVHead[I1] == UserBVHead[I2])
          return getElementIndex(IE1) < getElementIndex(IE2);
        if (UserBVHead[I1]->getParent() != UserBVHead[I2]->getParent())
          return CompareByBasicBlocks(UserBVHead[I1]->getParent(),
                                      UserBVHead[I2]->getParent());
        return UserBVHead[I1]->comesBefore(UserBVHead[I2]);
      }
      if (EE1 && !EE2)
        return true;
      if (!EE1 && EE2)
        return false;
      if (EE1 && EE2) {
        auto *Inst1 = dyn_cast<Instruction>(EE1->getOperand(0));
        auto *Inst2 = dyn_cast<Instruction>(EE2->getOperand(0));
        auto *P1 = dyn_cast<Argument>(EE1->getOperand(0));
        auto *P2 = dyn_cast<Argument>(EE2->getOperand(0));
        if (!Inst2 && !P2)
          return Inst1 || P1;
        if (EE1->getOperand(0) == EE2->getOperand(0))
          return getElementIndex(EE1) < getElementIndex(EE2);
        if (!Inst1 && Inst2)
          return false;
        if (Inst1 && Inst2) {
          if (Inst1->getParent() != Inst2->getParent())
            return CompareByBasicBlocks(Inst1->getParent(), Inst2->getParent());
          return Inst1->comesBefore(Inst2);
        }
        if (!P1 && P2)
          return false;
        assert(P1 && P2 &&
               "Expected either instructions or arguments vector operands.");
        return P1->getArgNo() < P2->getArgNo();
      }
      return false;
    };
    OrdersType Phis(TE.Scalars.size());
    std::iota(Phis.begin(), Phis.end(), 0);
    stable_sort(Phis, PHICompare);
    if (isIdentityOrder(Phis))
      return std::nullopt; // No need to reorder.
    return std::move(Phis);
  }
  if (TE.isGather() &&
      (!TE.hasState() || !TE.isAltShuffle() ||
       ScalarsInSplitNodes.contains(TE.getMainOp())) &&
      allSameType(TE.Scalars)) {
    // TODO: add analysis of other gather nodes with extractelement
    // instructions and other values/instructions, not only undefs.
    // Nodes with copyable lanes may mix in non-extract lanes, for which the
    // extract-index order is not applicable.
    if (((TE.hasState() && TE.getOpcode() == Instruction::ExtractElement &&
          !TE.hasCopyableElements()) ||
         (all_of(TE.Scalars, IsaPred<UndefValue, ExtractElementInst>) &&
          any_of(TE.Scalars, IsaPred<ExtractElementInst>))) &&
        all_of(TE.Scalars, [](Value *V) {
          auto *EE = dyn_cast<ExtractElementInst>(V);
          return !EE || isa<FixedVectorType>(EE->getVectorOperandType());
        })) {
      // Check that gather of extractelements can be represented as
      // just a shuffle of a single vector.
      OrdersType CurrentOrder;
      bool Reuse =
          canReuseExtract(TE.Scalars, CurrentOrder, /*ResizeAllowed=*/true);
      if (Reuse || !CurrentOrder.empty())
        return std::move(CurrentOrder);
    }
    // If the gather node is <undef, v, .., poison> and
    // insertelement poison, v, 0 [+ permute]
    // is cheaper than
    // insertelement poison, v, n - try to reorder.
    // If rotating the whole graph, exclude the permute cost, the whole graph
    // might be transformed.
    int Sz = TE.Scalars.size();
    if (isSplat(TE.Scalars) && !allConstant(TE.Scalars) &&
        count_if(TE.Scalars, IsaPred<UndefValue>) == Sz - 1) {
      const auto *It = find_if_not(TE.Scalars, isConstant);
      if (It == TE.Scalars.begin())
        return OrdersType();
      auto *Ty =
          cast<VectorType>(getWidenedType(TE.Scalars.front()->getType(), Sz));
      if (It != TE.Scalars.end()) {
        OrdersType Order(Sz, Sz);
        unsigned Idx = std::distance(TE.Scalars.begin(), It);
        Order[Idx] = 0;
        fixupOrderingIndices(Order);
        SmallVector<int> Mask;
        inversePermutation(Order, Mask);
        InstructionCost PermuteCost =
            TopToBottom ? 0
                        : getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc, Ty,
                                         CostKind, Mask);
        InstructionCost InsertFirstCost =
            TTI->getVectorInstrCost(Instruction::InsertElement, Ty, CostKind, 0,
                                    PoisonValue::get(Ty), *It);
        InstructionCost InsertIdxCost =
            TTI->getVectorInstrCost(Instruction::InsertElement, Ty, CostKind,
                                    Idx, PoisonValue::get(Ty), *It);
        if (InsertFirstCost + PermuteCost < InsertIdxCost) {
          OrdersType Order(Sz, Sz);
          Order[Idx] = 0;
          return std::move(Order);
        }
      }
    }
    if (isSplat(TE.Scalars))
      return std::nullopt;
    if (TE.Scalars.size() >= 3)
      if (std::optional<OrdersType> Order = findPartiallyOrderedLoads(TE))
        return Order;
    // Check if can include the order of vectorized loads. For masked gathers do
    // extra analysis later, so include such nodes into a special list.
    if (TE.hasState() && TE.getOpcode() == Instruction::Load) {
      SmallVector<Value *> PointerOps;
      StridedPtrInfo SPtrInfo;
      OrdersType CurrentOrder;
      LoadsState Res = canVectorizeLoads(TE.Scalars, TE.Scalars.front(),
                                         CurrentOrder, PointerOps, SPtrInfo);
      if (Res == LoadsState::Vectorize || Res == LoadsState::StridedVectorize ||
          Res == LoadsState::CompressVectorize ||
          Res == LoadsState::BlendedLoadVectorize)
        return std::move(CurrentOrder);
    }
    if (std::optional<OrdersType> CurrentOrder =
            findReusedOrderedScalars(TE, TopToBottom, IgnoreReorder))
      return CurrentOrder;
  }
  return std::nullopt;
}
 
/// Checks if the given mask is a "clustered" mask with the same clusters of
/// size \p Sz, which are not identity submasks.
static bool isRepeatedNonIdentityClusteredMask(ArrayRef<int> Mask,
                                               unsigned Sz) {
  ArrayRef<int> FirstCluster = Mask.slice(0, Sz);
  if (ShuffleVectorInst::isIdentityMask(FirstCluster, Sz))
    return false;
  for (unsigned I = Sz, E = Mask.size(); I < E; I += Sz) {
    ArrayRef<int> Cluster = Mask.slice(I, Sz);
    if (Cluster != FirstCluster)
      return false;
  }
  return true;
}
 
void BoUpSLP::reorderNodeWithReuses(TreeEntry &TE, ArrayRef<int> Mask) const {
  // Reorder reuses mask.
  reorderReuses(TE.ReuseShuffleIndices, Mask);
  const unsigned Sz = TE.Scalars.size();
  // For vectorized and non-clustered reused no need to do anything else.
  if (!TE.isGather() ||
      !ShuffleVectorInst::isOneUseSingleSourceMask(TE.ReuseShuffleIndices,
                                                   Sz) ||
      !isRepeatedNonIdentityClusteredMask(TE.ReuseShuffleIndices, Sz))
    return;
  SmallVector<int> NewMask;
  inversePermutation(TE.ReorderIndices, NewMask);
  addMask(NewMask, TE.ReuseShuffleIndices);
  // Clear reorder since it is going to be applied to the new mask.
  TE.ReorderIndices.clear();
  // Try to improve gathered nodes with clustered reuses, if possible.
  ArrayRef<int> Slice = ArrayRef(NewMask).slice(0, Sz);
  SmallVector<unsigned> NewOrder(Slice);
  inversePermutation(NewOrder, NewMask);
  reorderScalars(TE.Scalars, NewMask);
  // Fill the reuses mask with the identity submasks.
  for (auto *It = TE.ReuseShuffleIndices.begin(),
            *End = TE.ReuseShuffleIndices.end();
       It != End; std::advance(It, Sz))
    std::iota(It, std::next(It, Sz), 0);
}
 
static void combineOrders(MutableArrayRef<unsigned> Order,
                          ArrayRef<unsigned> SecondaryOrder) {
  assert((SecondaryOrder.empty() || Order.size() == SecondaryOrder.size()) &&
         "Expected same size of orders");
  size_t Sz = Order.size();
  SmallBitVector UsedIndices(Sz);
  for (unsigned Idx : seq<unsigned>(0, Sz)) {
    if (Order[Idx] != Sz)
      UsedIndices.set(Order[Idx]);
  }
  if (SecondaryOrder.empty()) {
    for (unsigned Idx : seq<unsigned>(0, Sz))
      if (Order[Idx] == Sz && !UsedIndices.test(Idx))
        Order[Idx] = Idx;
  } else {
    for (unsigned Idx : seq<unsigned>(0, Sz))
      if (SecondaryOrder[Idx] != Sz && Order[Idx] == Sz &&
          !UsedIndices.test(SecondaryOrder[Idx]))
        Order[Idx] = SecondaryOrder[Idx];
  }
}
 
bool BoUpSLP::isProfitableToReorder() const {
  if (DisableTreeReorder)
    return false;
 
  constexpr unsigned TinyVF = 2;
  constexpr unsigned TinyTree = 10;
  constexpr unsigned PhiOpsLimit = 12;
  constexpr unsigned GatherLoadsLimit = 2;
  if (VectorizableTree.size() <= TinyTree)
    return true;
  if (getRootNode().hasState() && !getRootNode().isGather() &&
      (getRootNode().getOpcode() == Instruction::Store ||
       getRootNode().getOpcode() == Instruction::PHI ||
       (getRootNode().getVectorFactor() <= TinyVF &&
        (getRootNode().getOpcode() == Instruction::PtrToInt ||
         getRootNode().getOpcode() == Instruction::PtrToAddr ||
         getRootNode().getOpcode() == Instruction::ICmp))) &&
      getRootNode().ReorderIndices.empty()) {
    // Check if the tree has only single store and single (unordered) load node,
    // other nodes are phis or geps/binops, combined with phis, and/or single
    // gather load node
    if (getRootNode().hasState() &&
        getRootNode().getOpcode() == Instruction::PHI &&
        getRootNodeScalars().size() == TinyVF &&
        getRootNode().getNumOperands() > PhiOpsLimit)
      return false;
    // Single node, which require reorder - skip.
    if (getRootNode().hasState() &&
        getRootNode().getOpcode() == Instruction::Store &&
        getRootNode().ReorderIndices.empty()) {
      const unsigned ReorderedSplitsCnt =
          count_if(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
            return TE->State == TreeEntry::SplitVectorize &&
                   !TE->ReorderIndices.empty() && TE->UserTreeIndex.UserTE &&
                   TE->UserTreeIndex.UserTE->State == TreeEntry::Vectorize &&
                   isCommutative(TE->UserTreeIndex.UserTE->getMainOp());
          });
      if (ReorderedSplitsCnt <= 1 &&
          static_cast<unsigned>(count_if(
              VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
                return ((!TE->isGather() &&
                         (TE->ReorderIndices.empty() ||
                          (TE->UserTreeIndex.UserTE &&
                           TE->UserTreeIndex.UserTE->State ==
                               TreeEntry::Vectorize &&
                           !TE->UserTreeIndex.UserTE->ReuseShuffleIndices
                                .empty()))) ||
                        (TE->isGather() && TE->ReorderIndices.empty() &&
                         (!TE->hasState() || TE->isAltShuffle() ||
                          TE->getOpcode() == Instruction::Load ||
                          TE->getOpcode() == Instruction::ZExt ||
                          TE->getOpcode() == Instruction::SExt))) &&
                       (getRootNode().getVectorFactor() > TinyVF ||
                        !TE->isGather() || none_of(TE->Scalars, [&](Value *V) {
                          return !isConstant(V) && isVectorized(V);
                        }));
              })) >= VectorizableTree.size() - ReorderedSplitsCnt)
        return false;
    }
    bool HasPhis = false;
    bool HasLoad = true;
    unsigned GatherLoads = 0;
    for (const std::unique_ptr<TreeEntry> &TE :
         ArrayRef(VectorizableTree).drop_front()) {
      if (TE->State == TreeEntry::SplitVectorize)
        continue;
      if (!TE->hasState()) {
        if (all_of(TE->Scalars, IsaPred<Constant, PHINode>) ||
            all_of(TE->Scalars, IsaPred<BinaryOperator, PHINode>))
          continue;
        if (getRootNodeScalars().size() == TinyVF &&
            any_of(TE->Scalars, IsaPred<PHINode, GEPOperator>))
          continue;
        return true;
      }
      if (TE->getOpcode() == Instruction::Load && TE->ReorderIndices.empty()) {
        if (!TE->isGather()) {
          HasLoad = false;
          continue;
        }
        if (HasLoad)
          return true;
        ++GatherLoads;
        if (GatherLoads >= GatherLoadsLimit)
          return true;
      }
      if (TE->getOpcode() == Instruction::GetElementPtr ||
          Instruction::isBinaryOp(TE->getOpcode()))
        continue;
      if (TE->getOpcode() != Instruction::PHI &&
          (!TE->hasCopyableElements() ||
           static_cast<unsigned>(count_if(TE->Scalars, IsaPred<PHINode>)) <
               TE->Scalars.size() / 2))
        return true;
      if (getRootNodeScalars().size() == TinyVF &&
          TE->getNumOperands() > PhiOpsLimit)
        return false;
      HasPhis = true;
    }
    return !HasPhis;
  }
  return true;
}
 
void BoUpSLP::TreeEntry::reorderSplitNode(unsigned Idx, ArrayRef<int> Mask,
                                          ArrayRef<int> MaskOrder) {
  assert(State == TreeEntry::SplitVectorize && "Expected split user node.");
  SmallVector<int> NewMask(getVectorFactor());
  SmallVector<int> NewMaskOrder(getVectorFactor());
  std::iota(NewMask.begin(), NewMask.end(), 0);
  std::iota(NewMaskOrder.begin(), NewMaskOrder.end(), 0);
  if (Idx == 0) {
    copy(Mask, NewMask.begin());
    copy(MaskOrder, NewMaskOrder.begin());
  } else {
    assert(Idx == 1 && "Expected either 0 or 1 index.");
    unsigned Offset = CombinedEntriesWithIndices.back().second;
    for (unsigned I : seq<unsigned>(Mask.size())) {
      NewMask[I + Offset] = Mask[I] + Offset;
      NewMaskOrder[I + Offset] = MaskOrder[I] + Offset;
    }
  }
  reorderScalars(Scalars, NewMask);
  reorderOrder(ReorderIndices, NewMaskOrder, /*BottomOrder=*/true);
  if (!ReorderIndices.empty() && BoUpSLP::isIdentityOrder(ReorderIndices))
    ReorderIndices.clear();
}
 
void BoUpSLP::markCopyableDepsForRecalc(TreeEntry &TE) {
  if (!TE.hasState())
    return;
  if (auto It = BlocksSchedules.find(TE.getMainOp()->getParent());
      It != BlocksSchedules.end())
    It->second->markCopyableDepsForRecalc(TE);
}
 
void BoUpSLP::reorderTopToBottom() {
  // Maps VF to the graph nodes.
  DenseMap<unsigned, SetVector<TreeEntry *>> VFToOrderedEntries;
  // ExtractElement gather nodes which can be vectorized and need to handle
  // their ordering.
  DenseMap<const TreeEntry *, OrdersType> GathersToOrders;
 
  // Phi nodes can have preferred ordering based on their result users
  DenseMap<const TreeEntry *, OrdersType> PhisToOrders;
 
  // AltShuffles can also have a preferred ordering that leads to fewer
  // instructions, e.g., the addsub instruction in x86.
  DenseMap<const TreeEntry *, OrdersType> AltShufflesToOrders;
 
  // Maps a TreeEntry to the reorder indices of external users.
  DenseMap<const TreeEntry *, SmallVector<OrdersType, 1>>
      ExternalUserReorderMap;
  // TODO: Reordering of struct types is not supported.
  if (any_of(VectorizableTree, [](const std::unique_ptr<TreeEntry> &TE) {
        return TE->State == TreeEntry::Vectorize &&
               isa<StructType>(getValueType(TE->Scalars.front(), SLPReVec));
      }))
    return;
  // Compute IgnoreReorder once - it depends only on UserIgnoreList and
  // getRootNode(), which do not change during this loop.
  const bool IgnoreReorder =
      !UserIgnoreList && getRootNode().hasState() &&
      (getRootNode().getOpcode() == Instruction::InsertElement ||
       getRootNode().getOpcode() == Instruction::InsertValue ||
       getRootNode().getOpcode() == Instruction::Store);
  // Find all reorderable nodes with the given VF.
  // Currently the are vectorized stores,loads,extracts + some gathering of
  // extracts.
  for_each(VectorizableTree, [&, &TTIRef = *TTI](
                                 const std::unique_ptr<TreeEntry> &TE) {
    // Look for external users that will probably be vectorized.
    SmallVector<OrdersType, 1> ExternalUserReorderIndices =
        findExternalStoreUsersReorderIndices(TE.get());
    if (!ExternalUserReorderIndices.empty()) {
      VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get());
      ExternalUserReorderMap.try_emplace(TE.get(),
                                         std::move(ExternalUserReorderIndices));
    }
 
    // Patterns like [fadd,fsub] can be combined into a single instruction in
    // x86. Reordering them into [fsub,fadd] blocks this pattern. So we need
    // to take into account their order when looking for the most used order.
    if (TE->hasState() && TE->isAltShuffle() &&
        TE->State != TreeEntry::SplitVectorize) {
      Type *ScalarTy = TE->Scalars[0]->getType();
      auto *VecTy =
          cast<VectorType>(getWidenedType(ScalarTy, TE->Scalars.size()));
      unsigned Opcode0 = TE->getOpcode();
      unsigned Opcode1 = TE->getAltOpcode();
      SmallBitVector OpcodeMask(
          getAltInstrMask(TE->Scalars, ScalarTy, Opcode0, Opcode1));
      // If this pattern is supported by the target then we consider the order.
      if (TTIRef.isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask)) {
        VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get());
        AltShufflesToOrders.try_emplace(TE.get(), OrdersType());
      }
      // TODO: Check the reverse order too.
    }
 
    if (std::optional<OrdersType> CurrentOrder =
            getReorderingData(*TE, /*TopToBottom=*/true, IgnoreReorder)) {
      // Do not include ordering for nodes used in the alt opcode vectorization,
      // better to reorder them during bottom-to-top stage. If follow the order
      // here, it causes reordering of the whole graph though actually it is
      // profitable just to reorder the subgraph that starts from the alternate
      // opcode vectorization node. Such nodes already end-up with the shuffle
      // instruction and it is just enough to change this shuffle rather than
      // rotate the scalars for the whole graph.
      unsigned Cnt = 0;
      const TreeEntry *UserTE = TE.get();
      while (UserTE && Cnt < RecursionMaxDepth) {
        if (!UserTE->UserTreeIndex)
          break;
        if (UserTE->UserTreeIndex.UserTE->State == TreeEntry::Vectorize &&
            UserTE->UserTreeIndex.UserTE->isAltShuffle() &&
            UserTE->UserTreeIndex.UserTE->Idx != 0)
          return;
        UserTE = UserTE->UserTreeIndex.UserTE;
        ++Cnt;
      }
      VFToOrderedEntries[TE->getVectorFactor()].insert(TE.get());
      if (!(TE->State == TreeEntry::Vectorize ||
            TE->State == TreeEntry::StridedVectorize ||
            TE->State == TreeEntry::ExpandVectorize ||
            TE->State == TreeEntry::SplitVectorize ||
            TE->State == TreeEntry::CompressVectorize ||
            TE->State == TreeEntry::BlendedLoadVectorize) ||
          !TE->ReuseShuffleIndices.empty())
        GathersToOrders.try_emplace(TE.get(), *CurrentOrder);
      if (TE->State == TreeEntry::Vectorize &&
          TE->getOpcode() == Instruction::PHI)
        PhisToOrders.try_emplace(TE.get(), *CurrentOrder);
    }
  });
 
  // Reorder the graph nodes according to their vectorization factor.
  for (unsigned VF = getRootNode().getVectorFactor();
       !VFToOrderedEntries.empty() && VF > 1; --VF) {
    auto It = VFToOrderedEntries.find(VF);
    if (It == VFToOrderedEntries.end())
      continue;
    // Try to find the most profitable order. We just are looking for the most
    // used order and reorder scalar elements in the nodes according to this
    // mostly used order.
    ArrayRef<TreeEntry *> OrderedEntries = It->second.getArrayRef();
    // Delete VF entry upon exit.
    llvm::scope_exit Cleanup([&]() { VFToOrderedEntries.erase(It); });
 
    // All operands are reordered and used only in this node - propagate the
    // most used order to the user node.
    MapVector<OrdersType, unsigned,
              DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>>
        OrdersUses;
    for (const TreeEntry *OpTE : OrderedEntries) {
      // No need to reorder this nodes, still need to extend and to use shuffle,
      // just need to merge reordering shuffle and the reuse shuffle.
      if (!OpTE->ReuseShuffleIndices.empty() && !GathersToOrders.count(OpTE) &&
          OpTE->State != TreeEntry::SplitVectorize)
        continue;
      // Count number of orders uses.
      const auto &Order = [OpTE, &GathersToOrders, &AltShufflesToOrders,
                           &PhisToOrders]() -> const OrdersType & {
        if (OpTE->isGather() || !OpTE->ReuseShuffleIndices.empty()) {
          auto It = GathersToOrders.find(OpTE);
          if (It != GathersToOrders.end())
            return It->second;
        }
        if (OpTE->hasState() && OpTE->isAltShuffle()) {
          auto It = AltShufflesToOrders.find(OpTE);
          if (It != AltShufflesToOrders.end())
            return It->second;
        }
        if (OpTE->State == TreeEntry::Vectorize &&
            OpTE->getOpcode() == Instruction::PHI) {
          auto It = PhisToOrders.find(OpTE);
          if (It != PhisToOrders.end())
            return It->second;
        }
        return OpTE->ReorderIndices;
      }();
      // First consider the order of the external scalar users.
      auto It = ExternalUserReorderMap.find(OpTE);
      if (It != ExternalUserReorderMap.end()) {
        const auto &ExternalUserReorderIndices = It->second;
        // If the OpTE vector factor != number of scalars - use natural order,
        // it is an attempt to reorder node with reused scalars but with
        // external uses.
        if (OpTE->getVectorFactor() != OpTE->Scalars.size()) {
          OrdersUses.try_emplace(OrdersType(), 0).first->second +=
              ExternalUserReorderIndices.size();
        } else {
          for (const OrdersType &ExtOrder : ExternalUserReorderIndices)
            ++OrdersUses.try_emplace(ExtOrder, 0).first->second;
        }
        // No other useful reorder data in this entry.
        if (Order.empty())
          continue;
      }
      // Stores actually store the mask, not the order, need to invert.
      if (OpTE->State == TreeEntry::Vectorize &&
          OpTE->getOpcode() == Instruction::Store && !Order.empty()) {
        assert(!OpTE->isAltShuffle() &&
               "Alternate instructions are only supported by BinaryOperator "
               "and CastInst.");
        SmallVector<int> Mask;
        inversePermutation(Order, Mask);
        unsigned E = Order.size();
        OrdersType CurrentOrder(E, E);
        transform(Mask, CurrentOrder.begin(), [E](int Idx) {
          return Idx == PoisonMaskElem ? E : static_cast<unsigned>(Idx);
        });
        fixupOrderingIndices(CurrentOrder);
        ++OrdersUses.try_emplace(CurrentOrder, 0).first->second;
      } else {
        ++OrdersUses.try_emplace(Order, 0).first->second;
      }
    }
    if (OrdersUses.empty())
      continue;
    // Choose the most used order.
    unsigned IdentityCnt = 0;
    unsigned FilledIdentityCnt = 0;
    OrdersType IdentityOrder(VF, VF);
    for (auto &Pair : OrdersUses) {
      if (Pair.first.empty() || isIdentityOrder(Pair.first)) {
        if (!Pair.first.empty())
          FilledIdentityCnt += Pair.second;
        IdentityCnt += Pair.second;
        combineOrders(IdentityOrder, Pair.first);
      }
    }
    MutableArrayRef<unsigned> BestOrder = IdentityOrder;
    unsigned Cnt = IdentityCnt;
    for (auto &Pair : OrdersUses) {
      // Prefer identity order. But, if filled identity found (non-empty order)
      // with same number of uses, as the new candidate order, we can choose
      // this candidate order.
      if (Cnt < Pair.second ||
          (Cnt == IdentityCnt && IdentityCnt == FilledIdentityCnt &&
           Cnt == Pair.second && !BestOrder.empty() &&
           isIdentityOrder(BestOrder))) {
        combineOrders(Pair.first, BestOrder);
        BestOrder = Pair.first;
        Cnt = Pair.second;
      } else {
        combineOrders(BestOrder, Pair.first);
      }
    }
    // Set order of the user node.
    if (isIdentityOrder(BestOrder))
      continue;
    fixupOrderingIndices(BestOrder);
    SmallVector<int> Mask;
    inversePermutation(BestOrder, Mask);
    SmallVector<int> MaskOrder(BestOrder.size(), PoisonMaskElem);
    unsigned E = BestOrder.size();
    transform(BestOrder, MaskOrder.begin(), [E](unsigned I) {
      return I < E ? static_cast<int>(I) : PoisonMaskElem;
    });
    // Do an actual reordering, if profitable.
    for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
      // Just do the reordering for the nodes with the given VF.
      if (TE->Scalars.size() != VF) {
        if (TE->ReuseShuffleIndices.size() == VF &&
            TE->State != TreeEntry::ExpandVectorize) {
          assert(TE->State != TreeEntry::SplitVectorize &&
                 "Split vectorized not expected.");
          // Need to reorder the reuses masks of the operands with smaller VF to
          // be able to find the match between the graph nodes and scalar
          // operands of the given node during vectorization/cost estimation.
          assert(
              (!TE->UserTreeIndex ||
               TE->UserTreeIndex.UserTE->Scalars.size() == VF ||
               TE->UserTreeIndex.UserTE->Scalars.size() == TE->Scalars.size() ||
               TE->UserTreeIndex.UserTE->State == TreeEntry::SplitVectorize) &&
              "All users must be of VF size.");
          if (SLPReVec) {
            assert(SLPReVec && "Only supported by REVEC.");
            // ShuffleVectorInst does not do reorderOperands (and it should not
            // because ShuffleVectorInst supports only a limited set of
            // patterns). Only do reorderNodeWithReuses if the user is not
            // ShuffleVectorInst.
            if (TE->UserTreeIndex && TE->UserTreeIndex.UserTE->hasState() &&
                isa<ShuffleVectorInst>(TE->UserTreeIndex.UserTE->getMainOp()))
              continue;
          }
          // Update ordering of the operands with the smaller VF than the given
          // one.
          reorderNodeWithReuses(*TE, Mask);
          // Update orders in user split vectorize nodes.
          if (TE->UserTreeIndex &&
              TE->UserTreeIndex.UserTE->State == TreeEntry::SplitVectorize)
            TE->UserTreeIndex.UserTE->reorderSplitNode(
                TE->UserTreeIndex.EdgeIdx, Mask, MaskOrder);
        }
        continue;
      }
      if ((TE->State == TreeEntry::SplitVectorize &&
           TE->ReuseShuffleIndices.empty()) ||
          ((TE->State == TreeEntry::Vectorize ||
            TE->State == TreeEntry::StridedVectorize ||
            TE->State == TreeEntry::ExpandVectorize ||
            TE->State == TreeEntry::CompressVectorize ||
            TE->State == TreeEntry::BlendedLoadVectorize) &&
           (isa<ExtractElementInst, ExtractValueInst, LoadInst, StoreInst,
                InsertElementInst, InsertValueInst>(TE->getMainOp()) ||
            (SLPReVec && isa<ShuffleVectorInst>(TE->getMainOp()))))) {
        assert(
            (!TE->isAltShuffle() || (TE->State == TreeEntry::SplitVectorize &&
                                     TE->ReuseShuffleIndices.empty())) &&
            "Alternate instructions are only supported by BinaryOperator "
            "and CastInst.");
        // Build correct orders for extract{element,value}, loads,
        // stores and alternate (split) nodes.
        reorderOrder(TE->ReorderIndices, Mask);
        if (isa<InsertElementInst, InsertValueInst, StoreInst>(
                TE->getMainOp())) {
          TE->reorderOperands(Mask);
          markCopyableDepsForRecalc(*TE);
        }
      } else {
        // Reorder the node and its operands.
        TE->reorderOperands(Mask);
        markCopyableDepsForRecalc(*TE);
        assert(TE->ReorderIndices.empty() &&
               "Expected empty reorder sequence.");
        reorderScalars(TE->Scalars, Mask);
      }
      if (!TE->ReuseShuffleIndices.empty() &&
          TE->State != TreeEntry::ExpandVectorize) {
        // Apply reversed order to keep the original ordering of the reused
        // elements to avoid extra reorder indices shuffling. An ExpandVectorize
        // store keeps its expand mask fixed and carries the reorder in
        // ReorderIndices, so it is excluded here.
        OrdersType CurrentOrder;
        reorderOrder(CurrentOrder, MaskOrder);
        SmallVector<int> NewReuses;
        inversePermutation(CurrentOrder, NewReuses);
        addMask(NewReuses, TE->ReuseShuffleIndices);
        TE->ReuseShuffleIndices.swap(NewReuses);
      } else if (TE->UserTreeIndex &&
                 TE->UserTreeIndex.UserTE->State == TreeEntry::SplitVectorize)
        // Update orders in user split vectorize nodes.
        TE->UserTreeIndex.UserTE->reorderSplitNode(TE->UserTreeIndex.EdgeIdx,
                                                   Mask, MaskOrder);
    }
  }
}
 
void BoUpSLP::buildReorderableOperands(
    TreeEntry *UserTE, SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges,
    const SmallPtrSetImpl<const TreeEntry *> &ReorderableGathers,
    SmallVectorImpl<TreeEntry *> &GatherOps) {
  for (unsigned I : seq<unsigned>(UserTE->getNumOperands())) {
    if (any_of(Edges, [I](const std::pair<unsigned, TreeEntry *> &OpData) {
          return OpData.first == I &&
                 (OpData.second->State == TreeEntry::Vectorize ||
                  OpData.second->State == TreeEntry::StridedVectorize ||
                  OpData.second->State == TreeEntry::ExpandVectorize ||
                  OpData.second->State == TreeEntry::CompressVectorize ||
                  OpData.second->State == TreeEntry::BlendedLoadVectorize ||
                  OpData.second->State == TreeEntry::SplitVectorize);
        }))
      continue;
    // Do not request operands, if they do not exist.
    if (UserTE->hasState()) {
      if (UserTE->getOpcode() == Instruction::ExtractElement ||
          UserTE->getOpcode() == Instruction::ExtractValue)
        continue;
      if ((UserTE->getOpcode() == Instruction::InsertElement ||
           UserTE->getOpcode() == Instruction::InsertValue) &&
          I == 0)
        continue;
      if (UserTE->getOpcode() == Instruction::Store && I == 1 &&
          (UserTE->State == TreeEntry::Vectorize ||
           UserTE->State == TreeEntry::StridedVectorize ||
           UserTE->State == TreeEntry::ExpandVectorize))
        continue;
      if (UserTE->getOpcode() == Instruction::Load &&
          (UserTE->State == TreeEntry::Vectorize ||
           UserTE->State == TreeEntry::StridedVectorize ||
           UserTE->State == TreeEntry::CompressVectorize ||
           UserTE->State == TreeEntry::BlendedLoadVectorize))
        continue;
    }
    TreeEntry *TE = getOperandEntry(UserTE, I);
    assert(TE && "Expected operand entry.");
    if (!TE->isGather()) {
      // Add the node to the list of the ordered nodes with the identity
      // order.
      Edges.emplace_back(I, TE);
      // Add ScatterVectorize nodes to the list of operands, where just
      // reordering of the scalars is required. Similar to the gathers, so
      // simply add to the list of gathered ops.
      // If there are reused scalars, process this node as a regular vectorize
      // node, just reorder reuses mask.
      if (TE->State == TreeEntry::ScatterVectorize &&
          TE->ReuseShuffleIndices.empty() && TE->ReorderIndices.empty())
        GatherOps.push_back(TE);
      continue;
    }
    if (ReorderableGathers.contains(TE))
      GatherOps.push_back(TE);
  }
}
 
void BoUpSLP::reorderBottomToTop(bool IgnoreReorder) {
  struct TreeEntryCompare {
    bool operator()(const TreeEntry *LHS, const TreeEntry *RHS) const {
      if (LHS->UserTreeIndex && RHS->UserTreeIndex)
        return LHS->UserTreeIndex.UserTE->Idx < RHS->UserTreeIndex.UserTE->Idx;
      return LHS->Idx < RHS->Idx;
    }
  };
  PriorityQueue<TreeEntry *, SmallVector<TreeEntry *>, TreeEntryCompare> Queue;
  DenseSet<const TreeEntry *> GathersToOrders;
  // Find all reorderable leaf nodes with the given VF.
  // Currently the are vectorized loads,extracts without alternate operands +
  // some gathering of extracts.
  SmallPtrSet<const TreeEntry *, 4> NonVectorized;
  for (const std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
    if (TE->State != TreeEntry::Vectorize &&
        TE->State != TreeEntry::StridedVectorize &&
        TE->State != TreeEntry::ExpandVectorize &&
        TE->State != TreeEntry::CompressVectorize &&
        TE->State != TreeEntry::BlendedLoadVectorize &&
        TE->State != TreeEntry::SplitVectorize)
      NonVectorized.insert(TE.get());
    if (std::optional<OrdersType> CurrentOrder =
            getReorderingData(*TE, /*TopToBottom=*/false, IgnoreReorder)) {
      Queue.push(TE.get());
      if (!(TE->State == TreeEntry::Vectorize ||
            TE->State == TreeEntry::StridedVectorize ||
            TE->State == TreeEntry::ExpandVectorize ||
            TE->State == TreeEntry::CompressVectorize ||
            TE->State == TreeEntry::BlendedLoadVectorize ||
            TE->State == TreeEntry::SplitVectorize) ||
          !TE->ReuseShuffleIndices.empty())
        GathersToOrders.insert(TE.get());
    }
  }
 
  // 1. Propagate order to the graph nodes, which use only reordered nodes.
  // I.e., if the node has operands, that are reordered, try to make at least
  // one operand order in the natural order and reorder others + reorder the
  // user node itself.
  SmallPtrSet<const TreeEntry *, 4> Visited, RevisitedOps;
  while (!Queue.empty()) {
    // 1. Filter out only reordered nodes.
    std::pair<TreeEntry *, SmallVector<std::pair<unsigned, TreeEntry *>>> Users;
    TreeEntry *TE = Queue.top();
    const TreeEntry *UserTE = TE->UserTreeIndex.UserTE;
    Queue.pop();
    SmallVector<TreeEntry *> OrderedOps(1, TE);
    while (!Queue.empty()) {
      TE = Queue.top();
      if (!UserTE || UserTE != TE->UserTreeIndex.UserTE)
        break;
      Queue.pop();
      OrderedOps.push_back(TE);
    }
    for (TreeEntry *TE : OrderedOps) {
      if (!(TE->State == TreeEntry::Vectorize ||
            TE->State == TreeEntry::StridedVectorize ||
            TE->State == TreeEntry::ExpandVectorize ||
            TE->State == TreeEntry::CompressVectorize ||
            TE->State == TreeEntry::BlendedLoadVectorize ||
            TE->State == TreeEntry::SplitVectorize ||
            (TE->isGather() && GathersToOrders.contains(TE))) ||
          !TE->UserTreeIndex ||
          TE->UserTreeIndex.UserTE->State == TreeEntry::BlendedLoadVectorize ||
          !TE->ReuseShuffleIndices.empty() || !Visited.insert(TE).second)
        continue;
      // Build a map between user nodes and their operands order to speedup
      // search. The graph currently does not provide this dependency directly.
      Users.first = TE->UserTreeIndex.UserTE;
      Users.second.emplace_back(TE->UserTreeIndex.EdgeIdx, TE);
    }
    if (Users.first) {
      auto &Data = Users;
      // TODO: Reordering of struct types is not supported.
      if (Data.first->State == TreeEntry::Vectorize &&
          isa<StructType>(getValueType(Data.first->Scalars.front(), SLPReVec)))
        continue;
      if (Data.first->State == TreeEntry::SplitVectorize) {
        assert(
            Data.second.size() <= 2 &&
            "Expected not greater than 2 operands for split vectorize node.");
        if (any_of(Data.second,
                   [](const auto &Op) { return !Op.second->UserTreeIndex; }))
          continue;
        // Update orders in user split vectorize nodes.
        assert(Data.first->CombinedEntriesWithIndices.size() == 2 &&
               "Expected exactly 2 entries.");
        for (const auto &P : Data.first->CombinedEntriesWithIndices) {
          TreeEntry &OpTE = *VectorizableTree[P.first];
          // The order of an operand that has both reordered and reused scalars
          // cannot be absorbed into the split node cleanly: clearing the
          // reorder indices while keeping the reuse mask (or vice versa)
          // desyncs the split node scalars from the operand effective order.
          // Skip reordering for such operands.
          if (OpTE.State != TreeEntry::SplitVectorize &&
              !OpTE.ReorderIndices.empty() && !OpTE.ReuseShuffleIndices.empty())
            continue;
          OrdersType Order = OpTE.ReorderIndices;
          if (Order.empty() || !OpTE.ReuseShuffleIndices.empty()) {
            if (!OpTE.isGather() && OpTE.ReuseShuffleIndices.empty())
              continue;
            const auto BestOrder =
                getReorderingData(OpTE, /*TopToBottom=*/false, IgnoreReorder);
            if (!BestOrder || BestOrder->empty() || isIdentityOrder(*BestOrder))
              continue;
            Order = *BestOrder;
          }
          fixupOrderingIndices(Order);
          SmallVector<int> Mask;
          inversePermutation(Order, Mask);
          const unsigned E = Order.size();
          SmallVector<int> MaskOrder(E, PoisonMaskElem);
          transform(Order, MaskOrder.begin(), [E](unsigned I) {
            return I < E ? static_cast<int>(I) : PoisonMaskElem;
          });
          Data.first->reorderSplitNode(P.second ? 1 : 0, Mask, MaskOrder);
          // Clear ordering of the operand.
          if (!OpTE.ReorderIndices.empty()) {
            OpTE.ReorderIndices.clear();
          } else if (!OpTE.ReuseShuffleIndices.empty()) {
            reorderReuses(OpTE.ReuseShuffleIndices, Mask);
          } else {
            assert(OpTE.isGather() && "Expected only gather/buildvector node.");
            reorderScalars(OpTE.Scalars, Mask);
          }
        }
        if (Data.first->ReuseShuffleIndices.empty() &&
            !Data.first->ReorderIndices.empty()) {
          // Insert user node to the list to try to sink reordering deeper in
          // the graph.
          Queue.push(Data.first);
        }
        continue;
      }
      // Do not move the operand order to the root PHI node when the root
      // order must be preserved: the root has no user to take over the order
      // and it cannot be dropped at the end of the reordering.
      if (!IgnoreReorder && Data.first == &getRootNode() &&
          !Data.first->UserTreeIndex &&
          Data.first->State == TreeEntry::Vectorize &&
          Data.first->getOpcode() == Instruction::PHI &&
          Data.first->ReuseShuffleIndices.empty())
        continue;
      // Check that operands are used only in the User node.
      SmallVector<TreeEntry *> GatherOps;
      buildReorderableOperands(Data.first, Data.second, NonVectorized,
                               GatherOps);
      // All operands are reordered and used only in this node - propagate the
      // most used order to the user node.
      MapVector<OrdersType, unsigned,
                DenseMap<OrdersType, unsigned, OrdersTypeDenseMapInfo>>
          OrdersUses;
      // Do the analysis for each tree entry only once, otherwise the order of
      // the same node my be considered several times, though might be not
      // profitable.
      SmallPtrSet<const TreeEntry *, 4> VisitedOps;
      SmallPtrSet<const TreeEntry *, 4> VisitedUsers;
      for (const auto &Op : Data.second) {
        TreeEntry *OpTE = Op.second;
        if (!VisitedOps.insert(OpTE).second)
          continue;
        if (!OpTE->ReuseShuffleIndices.empty() && !GathersToOrders.count(OpTE))
          continue;
        const auto Order = [&]() -> const OrdersType {
          if (OpTE->isGather() || !OpTE->ReuseShuffleIndices.empty())
            return getReorderingData(*OpTE, /*TopToBottom=*/false,
                                     IgnoreReorder)
                .value_or(OrdersType(1));
          return OpTE->ReorderIndices;
        }();
        // The order is partially ordered, skip it in favor of fully non-ordered
        // orders.
        if (Order.size() == 1)
          continue;
 
        // Check that the reordering does not increase number of shuffles, i.e.
        // same-values-nodes has same parents or their parents has same parents.
        if (!Order.empty() && !isIdentityOrder(Order)) {
          Value *Root = OpTE->hasState()
                            ? OpTE->getMainOp()
                            : *find_if_not(OpTE->Scalars, isConstant);
          auto GetSameNodesUsers = [&](Value *Root) {
            SmallSetVector<TreeEntry *, 4> Res;
            for (const TreeEntry *TE : ValueToGatherNodes.lookup(Root)) {
              if (TE != OpTE && TE->UserTreeIndex &&
                  TE->getVectorFactor() == OpTE->getVectorFactor() &&
                  TE->Scalars.size() == OpTE->Scalars.size() &&
                  ((TE->ReorderIndices.empty() && OpTE->isSame(TE->Scalars)) ||
                   (OpTE->ReorderIndices.empty() && TE->isSame(OpTE->Scalars))))
                Res.insert(TE->UserTreeIndex.UserTE);
            }
            for (const TreeEntry *TE : getTreeEntries(Root)) {
              if (TE != OpTE && TE->UserTreeIndex &&
                  TE->getVectorFactor() == OpTE->getVectorFactor() &&
                  TE->Scalars.size() == OpTE->Scalars.size() &&
                  ((TE->ReorderIndices.empty() && OpTE->isSame(TE->Scalars)) ||
                   (OpTE->ReorderIndices.empty() && TE->isSame(OpTE->Scalars))))
                Res.insert(TE->UserTreeIndex.UserTE);
            }
            return Res.takeVector();
          };
          auto GetNumOperands = [](const TreeEntry *TE) {
            if (TE->State == TreeEntry::SplitVectorize)
              return TE->getNumOperands();
            if (auto *CI = dyn_cast<CallInst>(TE->getMainOp()); CI)
              return CI->arg_size();
            return TE->getNumOperands();
          };
          auto NodeShouldBeReorderedWithOperands = [&, TTI = TTI](
                                                       const TreeEntry *TE) {
            Intrinsic::ID ID = Intrinsic::not_intrinsic;
            if (auto *CI = dyn_cast<CallInst>(TE->getMainOp()); CI)
              ID = getVectorIntrinsicIDForCall(CI, TLI);
            for (unsigned Idx : seq<unsigned>(GetNumOperands(TE))) {
              if (ID != Intrinsic::not_intrinsic &&
                  isVectorIntrinsicWithScalarOpAtArg(ID, Idx, TTI))
                continue;
              const TreeEntry *Op = getOperandEntry(TE, Idx);
              if (Op->isGather() && Op->hasState()) {
                const TreeEntry *VecOp =
                    getSameValuesTreeEntry(Op->getMainOp(), Op->Scalars);
                if (VecOp)
                  Op = VecOp;
              }
              if (Op->ReorderIndices.empty() && Op->ReuseShuffleIndices.empty())
                return false;
            }
            return true;
          };
          SmallVector<TreeEntry *> Users = GetSameNodesUsers(Root);
          if (!Users.empty() && !all_of(Users, [&](TreeEntry *UTE) {
                if (!RevisitedOps.insert(UTE).second)
                  return false;
                return UTE == Data.first || !UTE->ReorderIndices.empty() ||
                       !UTE->ReuseShuffleIndices.empty() ||
                       (UTE->UserTreeIndex &&
                        UTE->UserTreeIndex.UserTE == Data.first) ||
                       (Data.first->UserTreeIndex &&
                        Data.first->UserTreeIndex.UserTE == UTE) ||
                       (IgnoreReorder && UTE->UserTreeIndex &&
                        UTE->UserTreeIndex.UserTE->Idx == 0) ||
                       NodeShouldBeReorderedWithOperands(UTE);
              }))
            continue;
          for (TreeEntry *UTE : Users) {
            Intrinsic::ID ID = Intrinsic::not_intrinsic;
            if (auto *CI = dyn_cast<CallInst>(UTE->getMainOp()); CI)
              ID = getVectorIntrinsicIDForCall(CI, TLI);
            for (unsigned Idx : seq<unsigned>(GetNumOperands(UTE))) {
              if (ID != Intrinsic::not_intrinsic &&
                  isVectorIntrinsicWithScalarOpAtArg(ID, Idx, TTI))
                continue;
              const TreeEntry *Op = getOperandEntry(UTE, Idx);
              Visited.erase(Op);
              Queue.push(const_cast<TreeEntry *>(Op));
            }
          }
        }
        unsigned NumOps = count_if(
            Data.second, [OpTE](const std::pair<unsigned, TreeEntry *> &P) {
              return P.second == OpTE;
            });
        // Stores actually store the mask, not the order, need to invert.
        if (OpTE->State == TreeEntry::Vectorize &&
            OpTE->getOpcode() == Instruction::Store && !Order.empty()) {
          assert(!OpTE->isAltShuffle() &&
                 "Alternate instructions are only supported by BinaryOperator "
                 "and CastInst.");
          SmallVector<int> Mask;
          inversePermutation(Order, Mask);
          unsigned E = Order.size();
          OrdersType CurrentOrder(E, E);
          transform(Mask, CurrentOrder.begin(), [E](int Idx) {
            return Idx == PoisonMaskElem ? E : static_cast<unsigned>(Idx);
          });
          fixupOrderingIndices(CurrentOrder);
          OrdersUses.try_emplace(CurrentOrder, 0).first->second += NumOps;
        } else {
          OrdersUses.try_emplace(Order, 0).first->second += NumOps;
        }
        auto Res = OrdersUses.try_emplace(OrdersType(), 0);
        const auto AllowsReordering = [&](const TreeEntry *TE) {
          if (!TE->ReorderIndices.empty() || !TE->ReuseShuffleIndices.empty() ||
              (TE->State == TreeEntry::Vectorize && TE->isAltShuffle()) ||
              (IgnoreReorder && TE->Idx == 0))
            return true;
          if (TE->isGather()) {
            if (GathersToOrders.contains(TE))
              return !getReorderingData(*TE, /*TopToBottom=*/false,
                                        IgnoreReorder)
                          .value_or(OrdersType(1))
                          .empty();
            return true;
          }
          return false;
        };
        if (OpTE->UserTreeIndex) {
          TreeEntry *UserTE = OpTE->UserTreeIndex.UserTE;
          if (!VisitedUsers.insert(UserTE).second)
            continue;
          // May reorder user node if it requires reordering, has reused
          // scalars, is an alternate op vectorize node or its op nodes require
          // reordering.
          if (AllowsReordering(UserTE))
            continue;
          // Check if users allow reordering.
          // Currently look up just 1 level of operands to avoid increase of
          // the compile time.
          // Profitable to reorder if definitely more operands allow
          // reordering rather than those with natural order.
          ArrayRef<std::pair<unsigned, TreeEntry *>> Ops = Users.second;
          if (static_cast<unsigned>(count_if(
                  Ops, [UserTE, &AllowsReordering](
                           const std::pair<unsigned, TreeEntry *> &Op) {
                    return AllowsReordering(Op.second) &&
                           Op.second->UserTreeIndex.UserTE == UserTE;
                  })) <= Ops.size() / 2)
            ++Res.first->second;
        }
      }
      if (OrdersUses.empty()) {
        Visited.insert_range(llvm::make_second_range(Data.second));
        continue;
      }
      // Choose the most used order.
      unsigned IdentityCnt = 0;
      unsigned VF = Data.second.front().second->getVectorFactor();
      OrdersType IdentityOrder(VF, VF);
      for (auto &Pair : OrdersUses) {
        if (Pair.first.empty() || isIdentityOrder(Pair.first)) {
          IdentityCnt += Pair.second;
          combineOrders(IdentityOrder, Pair.first);
        }
      }
      MutableArrayRef<unsigned> BestOrder = IdentityOrder;
      unsigned Cnt = IdentityCnt;
      for (auto &Pair : OrdersUses) {
        // Prefer identity order. But, if filled identity found (non-empty
        // order) with same number of uses, as the new candidate order, we can
        // choose this candidate order.
        if (Cnt < Pair.second) {
          combineOrders(Pair.first, BestOrder);
          BestOrder = Pair.first;
          Cnt = Pair.second;
        } else {
          combineOrders(BestOrder, Pair.first);
        }
      }
      // Set order of the user node.
      if (isIdentityOrder(BestOrder)) {
        Visited.insert_range(llvm::make_second_range(Data.second));
        continue;
      }
      fixupOrderingIndices(BestOrder);
      // Erase operands from OrderedEntries list and adjust their orders.
      VisitedOps.clear();
      SmallVector<int> Mask;
      inversePermutation(BestOrder, Mask);
      SmallVector<int> MaskOrder(BestOrder.size(), PoisonMaskElem);
      unsigned E = BestOrder.size();
      transform(BestOrder, MaskOrder.begin(), [E](unsigned I) {
        return I < E ? static_cast<int>(I) : PoisonMaskElem;
      });
      for (const std::pair<unsigned, TreeEntry *> &Op : Data.second) {
        TreeEntry *TE = Op.second;
        if (!VisitedOps.insert(TE).second)
          continue;
        // TODO: Reordering of struct types is not supported.
        if (TE->State == TreeEntry::Vectorize &&
            isa<StructType>(getValueType(TE->Scalars.front(), SLPReVec)))
          continue;
        if (TE->ReuseShuffleIndices.size() == BestOrder.size()) {
          reorderNodeWithReuses(*TE, Mask);
          continue;
        }
        // Gathers are processed separately.
        if (TE->State != TreeEntry::Vectorize &&
            TE->State != TreeEntry::StridedVectorize &&
            TE->State != TreeEntry::ExpandVectorize &&
            TE->State != TreeEntry::CompressVectorize &&
            TE->State != TreeEntry::BlendedLoadVectorize &&
            TE->State != TreeEntry::SplitVectorize &&
            (TE->State != TreeEntry::ScatterVectorize ||
             TE->ReorderIndices.empty()))
          continue;
        assert((BestOrder.size() == TE->ReorderIndices.size() ||
                TE->ReorderIndices.empty()) &&
               "Non-matching sizes of user/operand entries.");
        reorderOrder(TE->ReorderIndices, Mask);
        if (IgnoreReorder && TE == &getRootNode())
          IgnoreReorder = false;
      }
      // For gathers just need to reorder its scalars.
      for (TreeEntry *Gather : GatherOps) {
        assert(Gather->ReorderIndices.empty() &&
               "Unexpected reordering of gathers.");
        if (!Gather->ReuseShuffleIndices.empty()) {
          // Just reorder reuses indices.
          reorderReuses(Gather->ReuseShuffleIndices, Mask);
          continue;
        }
        // A ScatterVectorize (masked gather) node is scheduled, and the
        // scheduler reads its operand list at the same lane where the scalar
        // load sits, so Scalars and the operand list must stay aligned.
        // Record the reorder in ReorderIndices (applied by the final shuffle)
        // instead of physically permuting the scalars, matching how a scatter
        // node with a non-empty order is reordered above.
        if (Gather->State == TreeEntry::ScatterVectorize) {
          reorderOrder(Gather->ReorderIndices, Mask);
          Visited.insert(Gather);
          continue;
        }
        reorderScalars(Gather->Scalars, Mask);
        Visited.insert(Gather);
      }
      // Reorder operands of the user node and set the ordering for the user
      // node itself.
      auto IsNotProfitableAltCodeNode = [](const TreeEntry &TE) {
        return TE.isAltShuffle() &&
               (!TE.ReuseShuffleIndices.empty() || TE.getVectorFactor() == 2 ||
                TE.ReorderIndices.empty());
      };
      if (Data.first->State != TreeEntry::Vectorize ||
          !isa<ExtractElementInst, ExtractValueInst, LoadInst>(
              Data.first->getMainOp()) ||
          IsNotProfitableAltCodeNode(*Data.first)) {
        Data.first->reorderOperands(Mask);
        markCopyableDepsForRecalc(*Data.first);
      }
      if (!isa<InsertElementInst, InsertValueInst, StoreInst>(
              Data.first->getMainOp()) ||
          IsNotProfitableAltCodeNode(*Data.first) ||
          Data.first->State == TreeEntry::CompressVectorize) {
        reorderScalars(Data.first->Scalars, Mask);
        reorderOrder(Data.first->ReorderIndices, MaskOrder,
                     /*BottomOrder=*/true);
        if (Data.first->ReuseShuffleIndices.empty() &&
            !Data.first->ReorderIndices.empty() &&
            !IsNotProfitableAltCodeNode(*Data.first)) {
          // Insert user node to the list to try to sink reordering deeper in
          // the graph.
          Queue.push(Data.first);
        }
      } else {
        reorderOrder(Data.first->ReorderIndices, Mask);
      }
    }
  }
  // If the reordering is unnecessary, just remove the reorder.
  if (IgnoreReorder && !getRootNode().ReorderIndices.empty() &&
      getRootNode().ReuseShuffleIndices.empty())
    getRootNode().ReorderIndices.clear();
}
 
Instruction *BoUpSLP::getRootEntryInstruction(const TreeEntry &Entry) const {
  if (Entry.hasState() &&
      (Entry.getOpcode() == Instruction::Store ||
       Entry.getOpcode() == Instruction::Load) &&
      Entry.State == TreeEntry::StridedVectorize &&
      !Entry.ReorderIndices.empty() && isReverseOrder(Entry.ReorderIndices))
    return dyn_cast<Instruction>(Entry.Scalars[Entry.ReorderIndices.front()]);
  return dyn_cast<Instruction>(Entry.Scalars.front());
}
 
void BoUpSLP::buildExternalUses(
    const ExtraValueToDebugLocsMap &ExternallyUsedValues) {
  const size_t NumVectScalars = ScalarToTreeEntries.size() + 1;
  DenseMap<Value *, unsigned> ScalarToExtUses;
  // Peeled scalars still claimed by the tree (gathered, listed in some
  // entry's scalars, or modeled as a copyable element, which is emitted as
  // a scalar) survive as plain code, along with the peeled scalars
  // in their operand chains, which no vector node can rematerialize.
  KeptReassocScalars.clear();
  SmallVector<const Value *, 8> KeptWorklist;
  for (const auto &[V, Owners] : ReassocScalarToTreeEntries)
    if ((isGathered(V) || !getTreeEntries(V).empty() ||
         any_of(Owners,
                [V = V](const TreeEntry *TE) {
                  return TE->isCopyableElement(const_cast<Value *>(V));
                })) &&
        KeptReassocScalars.insert(V).second)
      KeptWorklist.push_back(V);
  while (!KeptWorklist.empty()) {
    const Value *V = KeptWorklist.pop_back_val();
    for (const Value *Op : cast<Instruction>(V)->operand_values())
      if (ReassocScalarToTreeEntries.contains(Op) &&
          KeptReassocScalars.insert(Op).second)
        KeptWorklist.push_back(Op);
  }
  // Collect the values that we need to extract from the tree.
  for (auto &TEPtr : VectorizableTree) {
    TreeEntry *Entry = TEPtr.get();
 
    // No need to handle users of gathered values.
    if (Entry->isGather() || Entry->State == TreeEntry::SplitVectorize ||
        DeletedNodes.contains(Entry) ||
        TransformedToGatherNodes.contains(Entry))
      continue;
 
    // For each lane:
    for (int Lane = 0, LE = Entry->Scalars.size(); Lane != LE; ++Lane) {
      Value *Scalar = Entry->Scalars[Lane];
      if (!isa<Instruction>(Scalar) || Entry->isCopyableElement(Scalar))
        continue;
      bool IsStructScalar = isa<StructType>(Scalar->getType());
 
      // All uses must be replaced already? No need to do it again.
      auto It = ScalarToExtUses.find(Scalar);
      if (It != ScalarToExtUses.end() && !ExternalUses[It->second].User)
        continue;
 
      if (!IsStructScalar && Scalar->hasNUsesOrMore(NumVectScalars)) {
        unsigned FoundLane = Entry->findLaneForValue(Scalar);
        LLVM_DEBUG(dbgs() << "SLP: Need to extract from lane " << FoundLane
                          << " from " << *Scalar << "for many users.\n");
        It = ScalarToExtUses.try_emplace(Scalar, ExternalUses.size()).first;
        ExternalUses.emplace_back(Scalar, nullptr, *Entry, FoundLane);
        ExternalUsesWithNonUsers.insert(Scalar);
        continue;
      }
 
      // Check if the scalar is externally used as an extra arg.
      const auto ExtI = ExternallyUsedValues.find(Scalar);
      if (ExtI != ExternallyUsedValues.end()) {
        unsigned FoundLane = Entry->findLaneForValue(Scalar);
        LLVM_DEBUG(dbgs() << "SLP: Need to extract: Extra arg from lane "
                          << FoundLane << " from " << *Scalar << ".\n");
        ScalarToExtUses.try_emplace(Scalar, ExternalUses.size());
        ExternalUses.emplace_back(Scalar, nullptr, *Entry, FoundLane);
        continue;
      }
      for (User *U : Scalar->users()) {
        LLVM_DEBUG(dbgs() << "SLP: Checking user:" << *U << ".\n");
 
        Instruction *UserInst = dyn_cast<Instruction>(U);
        if (!UserInst || isDeleted(UserInst))
          continue;
 
        // Ignore users in the user ignore list.
        if (UserIgnoreList && UserIgnoreList->contains(UserInst))
          continue;
 
        // Peeled reassociated scalars are subsumed by the flattened node and
        // erased during vectorization, not external users. Kept ones survive
        // and their uses of erased scalars become extracts like any other.
        if (isReassocScalarVectorized(UserInst) &&
            !KeptReassocScalars.contains(UserInst)) {
          LLVM_DEBUG(dbgs() << "SLP: \tInternal (reassociated) user will be "
                               "removed:"
                            << *U << ".\n");
          continue;
        }
 
        // Skip in-tree scalars that become vectors
        if (ArrayRef<TreeEntry *> UseEntries = getTreeEntries(U);
            any_of(UseEntries, [this](const TreeEntry *UseEntry) {
              return !DeletedNodes.contains(UseEntry) &&
                     !TransformedToGatherNodes.contains(UseEntry);
            })) {
          // Some in-tree scalars will remain as scalar in vectorized
          // instructions. If that is the case, the one in FoundLane will
          // be used.
          if (!((Scalar->getType()->getScalarType()->isPointerTy() &&
                 isa<LoadInst, StoreInst>(UserInst)) ||
                isa<CallInst>(UserInst)) ||
              all_of(UseEntries, [&](TreeEntry *UseEntry) {
                if (DeletedNodes.contains(UseEntry) ||
                    TransformedToGatherNodes.contains(UseEntry))
                  return true;
                return UseEntry->State == TreeEntry::ScatterVectorize ||
                       !doesInTreeUserNeedToExtract(
                           Scalar, getRootEntryInstruction(*UseEntry), TLI,
                           TTI);
              })) {
            LLVM_DEBUG(dbgs() << "SLP: \tInternal user will be removed:" << *U
                              << ".\n");
            assert(none_of(UseEntries,
                           [](TreeEntry *UseEntry) {
                             return UseEntry->isGather();
                           }) &&
                   "Bad state");
            continue;
          }
          if (!IsStructScalar) {
            U = nullptr;
            if (It != ScalarToExtUses.end()) {
              ExternalUses[It->second].User = nullptr;
              break;
            }
          }
        }
 
        if (U && !IsStructScalar && Scalar->hasNUsesOrMore(UsesLimit))
          U = nullptr;
        unsigned FoundLane = Entry->findLaneForValue(Scalar);
        LLVM_DEBUG(dbgs() << "SLP: Need to extract:" << *UserInst
                          << " from lane " << FoundLane << " from " << *Scalar
                          << ".\n");
        It = ScalarToExtUses.try_emplace(Scalar, ExternalUses.size()).first;
        ExternalUses.emplace_back(Scalar, U, *Entry, FoundLane);
        ExternalUsesWithNonUsers.insert(Scalar);
        if (!U)
          break;
      }
    }
  }
 
  // The expansion of the runtime stride may reuse an in-tree instruction with
  // the matching SCEV, which gets erased upon vectorization. Register an
  // external use for it to replace the stride operand with the extract.
  SmallVector<std::pair<const SCEV *, BasicBlock *>> Strides;
  for (const auto &[StridedTE, SPtrInfo] : TreeEntryToStridedPtrInfoMap)
    if (SPtrInfo.StrideSCEV && !SPtrInfo.StrideVal &&
        !DeletedNodes.contains(StridedTE) &&
        !TransformedToGatherNodes.contains(StridedTE))
      Strides.emplace_back(SPtrInfo.StrideSCEV,
                           StridedTE->getMainOp()->getParent());
  if (Strides.empty())
    return;
  for (const std::unique_ptr<TreeEntry> &TEPtr : VectorizableTree) {
    TreeEntry *Entry = TEPtr.get();
    if (Entry->isGather() || Entry->State == TreeEntry::SplitVectorize ||
        DeletedNodes.contains(Entry) ||
        TransformedToGatherNodes.contains(Entry))
      continue;
    for (Value *Scalar : Entry->Scalars) {
      auto *I = dyn_cast<Instruction>(Scalar);
      if (!I || Entry->isCopyableElement(I) || !SE->isSCEVable(I->getType()))
        continue;
      const SCEV *ScalarSCEV = SE->getSCEV(I);
      if (isa<SCEVConstant>(ScalarSCEV) ||
          none_of(Strides, [&](const auto &Stride) {
            return DT->dominates(I, Stride.second->getTerminator()) &&
                   SCEVExprContains(Stride.first, [ScalarSCEV](const SCEV *S) {
                     return S == ScalarSCEV;
                   });
          }))
        continue;
      auto It = ScalarToExtUses.find(Scalar);
      if (It != ScalarToExtUses.end()) {
        // Replace all uses: the stride operand is emitted later, during the
        // codegen.
        ExternalUses[It->second].User = nullptr;
        continue;
      }
      unsigned FoundLane = Entry->findLaneForValue(Scalar);
      LLVM_DEBUG(dbgs() << "SLP: Need to extract: strided load stride from "
                           "lane "
                        << FoundLane << " from " << *Scalar << ".\n");
      ScalarToExtUses.try_emplace(Scalar, ExternalUses.size());
      ExternalUses.emplace_back(Scalar, nullptr, *Entry, FoundLane);
      ExternalUsesWithNonUsers.insert(Scalar);
    }
  }
}
 
SmallVector<SmallVector<StoreInst *>>
BoUpSLP::collectUserStores(const BoUpSLP::TreeEntry *TE) const {
  SmallDenseMap<std::tuple<BasicBlock *, Type *, Value *>,
                SmallVector<StoreInst *>, 8>
      PtrToStoresMap;
  for (unsigned Lane : seq<unsigned>(0, TE->Scalars.size())) {
    Value *V = TE->Scalars[Lane];
    // Don't iterate over the users of constant data.
    if (!isa<Instruction>(V))
      continue;
    // To save compilation time we don't visit if we have too many users.
    if (V->hasNUsesOrMore(UsesLimit))
      break;
 
    // Collect stores per pointer object.
    for (User *U : V->users()) {
      auto *SI = dyn_cast<StoreInst>(U);
      // Test whether we can handle the store. V might be a global, which could
      // be used in a different function.
      if (SI == nullptr || !SI->isSimple() || SI->getFunction() != F ||
          !isValidElementType(SI->getValueOperand()->getType(), SLPReVec))
        continue;
      // Skip entry if already
      if (isVectorized(U))
        continue;
 
      Value *Ptr =
          getUnderlyingObject(SI->getPointerOperand(), RecursionMaxDepth);
      auto &StoresVec = PtrToStoresMap[{SI->getParent(),
                                        SI->getValueOperand()->getType(), Ptr}];
      // For now just keep one store per pointer object per lane.
      // TODO: Extend this to support multiple stores per pointer per lane
      if (StoresVec.size() > Lane)
        continue;
      if (!StoresVec.empty()) {
        std::optional<int64_t> Diff = getPointersDiff(
            SI->getValueOperand()->getType(), SI->getPointerOperand(),
            SI->getValueOperand()->getType(),
            StoresVec.front()->getPointerOperand(), *DL, *SE,
            /*StrictCheck=*/true);
        // We failed to compare the pointers so just abandon this store.
        if (!Diff)
          continue;
      }
      StoresVec.push_back(SI);
    }
  }
  SmallVector<SmallVector<StoreInst *>> Res(PtrToStoresMap.size());
  unsigned I = 0;
  for (auto &P : PtrToStoresMap) {
    Res[I].swap(P.second);
    ++I;
  }
  return Res;
}
 
bool BoUpSLP::canFormVector(ArrayRef<StoreInst *> StoresVec,
                            OrdersType &ReorderIndices) const {
  // We check whether the stores in StoreVec can form a vector by sorting them
  // and checking whether they are consecutive.
 
  // To avoid calling getPointersDiff() while sorting we create a vector of
  // pairs {store, offset from first} and sort this instead.
  SmallVector<std::pair<int64_t, unsigned>> StoreOffsetVec;
  StoreInst *S0 = StoresVec[0];
  StoreOffsetVec.emplace_back(0, 0);
  Type *S0Ty = S0->getValueOperand()->getType();
  Value *S0Ptr = S0->getPointerOperand();
  for (unsigned Idx : seq<unsigned>(1, StoresVec.size())) {
    StoreInst *SI = StoresVec[Idx];
    std::optional<int64_t> Diff =
        getPointersDiff(S0Ty, S0Ptr, SI->getValueOperand()->getType(),
                        SI->getPointerOperand(), *DL, *SE,
                        /*StrictCheck=*/true);
    StoreOffsetVec.emplace_back(*Diff, Idx);
  }
 
  // Check if the stores are consecutive by checking if their difference is 1.
  if (StoreOffsetVec.size() != StoresVec.size())
    return false;
  sort(StoreOffsetVec, llvm::less_first());
  unsigned Idx = 0;
  int64_t PrevDist = 0;
  for (const auto &P : StoreOffsetVec) {
    if (Idx > 0 && P.first != PrevDist + 1)
      return false;
    PrevDist = P.first;
    ++Idx;
  }
 
  // Calculate the shuffle indices according to their offset against the sorted
  // StoreOffsetVec.
  ReorderIndices.assign(StoresVec.size(), 0);
  bool IsIdentity = true;
  for (auto [I, P] : enumerate(StoreOffsetVec)) {
    ReorderIndices[P.second] = I;
    IsIdentity &= P.second == I;
  }
  // Identity order (e.g., {0,1,2,3}) is modeled as an empty OrdersType in
  // reorderTopToBottom() and reorderBottomToTop(), so we are following the
  // same convention here.
  if (IsIdentity)
    ReorderIndices.clear();
 
  return true;
}
 
#ifndef NDEBUG
LLVM_DUMP_METHOD static void dumpOrder(const BoUpSLP::OrdersType &Order) {
  for (unsigned Idx : Order)
    dbgs() << Idx << ", ";
  dbgs() << "\n";
}
#endif
 
SmallVector<BoUpSLP::OrdersType, 1>
BoUpSLP::findExternalStoreUsersReorderIndices(TreeEntry *TE) const {
  unsigned NumLanes = TE->Scalars.size();
 
  SmallVector<SmallVector<StoreInst *>> Stores = collectUserStores(TE);
 
  // Holds the reorder indices for each candidate store vector that is a user of
  // the current TreeEntry.
  SmallVector<OrdersType, 1> ExternalReorderIndices;
 
  // Now inspect the stores collected per pointer and look for vectorization
  // candidates. For each candidate calculate the reorder index vector and push
  // it into `ExternalReorderIndices`
  for (ArrayRef<StoreInst *> StoresVec : Stores) {
    // If we have fewer than NumLanes stores, then we can't form a vector.
    if (StoresVec.size() != NumLanes)
      continue;
 
    // If the stores are not consecutive then abandon this StoresVec.
    OrdersType ReorderIndices;
    if (!canFormVector(StoresVec, ReorderIndices))
      continue;
 
    // We now know that the scalars in StoresVec can form a vector instruction,
    // so set the reorder indices.
    ExternalReorderIndices.push_back(ReorderIndices);
  }
  return ExternalReorderIndices;
}
 
void BoUpSLP::buildTree(ArrayRef<Value *> Roots,
                        const SmallDenseSet<Value *> &UserIgnoreLst) {
  deleteTree();
  assert(TreeEntryToStridedPtrInfoMap.empty() &&
         "TreeEntryToStridedPtrInfoMap is not cleared");
  UserIgnoreList = &UserIgnoreLst;
  if (!allSameType(Roots))
    return;
  buildTreeRec(Roots, 0, EdgeInfo());
  // Build splat-gather subtrees here so the reordering passes cover them too.
  tryToVectorizeSplatGatheredScalars();
}
 
void BoUpSLP::buildTree(ArrayRef<Value *> Roots) {
  deleteTree();
  assert(TreeEntryToStridedPtrInfoMap.empty() &&
         "TreeEntryToStridedPtrInfoMap is not cleared");
  if (!allSameType(Roots))
    return;
  buildTreeRec(Roots, 0, EdgeInfo());
  // Build splat-gather subtrees here so the reordering passes cover them too.
  tryToVectorizeSplatGatheredScalars();
}
 
/// Tries to find subvector of loads and builds new vector of only loads if can
/// be profitable.
static void gatherPossiblyVectorizableLoads(
    const BoUpSLP &R, ArrayRef<Value *> VL, const DataLayout &DL,
    ScalarEvolution &SE, const TargetTransformInfo &TTI,
    SmallVectorImpl<SmallVector<std::pair<LoadInst *, int64_t>>> &GatheredLoads,
    bool AddNew = true) {
  if (VL.empty())
    return;
  Type *ScalarTy = getValueType(VL.front(), SLPReVec);
  if (!isValidElementType(ScalarTy, SLPReVec))
    return;
  SmallVector<SmallVector<std::pair<LoadInst *, int64_t>>> ClusteredLoads;
  SmallVector<DenseMap<int64_t, LoadInst *>> ClusteredDistToLoad;
  for (Value *V : VL) {
    auto *LI = dyn_cast<LoadInst>(V);
    if (!LI)
      continue;
    if (R.isDeleted(LI) || R.isVectorized(LI) || !LI->isSimple())
      continue;
    bool IsFound = false;
    for (auto [Map, Data] : zip(ClusteredDistToLoad, ClusteredLoads)) {
      assert(LI->getParent() == Data.front().first->getParent() &&
             LI->getType() == Data.front().first->getType() &&
             getUnderlyingObject(LI->getPointerOperand(), RecursionMaxDepth) ==
                 getUnderlyingObject(Data.front().first->getPointerOperand(),
                                     RecursionMaxDepth) &&
             "Expected loads with the same type, same parent and same "
             "underlying pointer.");
      std::optional<int64_t> Dist = getPointersDiff(
          LI->getType(), LI->getPointerOperand(), Data.front().first->getType(),
          Data.front().first->getPointerOperand(), DL, SE,
          /*StrictCheck=*/true);
      if (!Dist)
        continue;
      auto It = Map.find(*Dist);
      if (It != Map.end() && It->second != LI)
        continue;
      if (It == Map.end()) {
        Data.emplace_back(LI, *Dist);
        Map.try_emplace(*Dist, LI);
      }
      IsFound = true;
      break;
    }
    if (!IsFound) {
      ClusteredLoads.emplace_back().emplace_back(LI, 0);
      ClusteredDistToLoad.emplace_back().try_emplace(0, LI);
    }
  }
  auto FindMatchingLoads =
      [&](ArrayRef<std::pair<LoadInst *, int64_t>> Loads,
          SmallVectorImpl<SmallVector<std::pair<LoadInst *, int64_t>>>
              &GatheredLoads,
          SetVector<unsigned> &ToAdd, SetVector<unsigned> &Repeated,
          int64_t &Offset, unsigned &Start) {
        if (Loads.empty())
          return GatheredLoads.end();
        LoadInst *LI = Loads.front().first;
        for (auto [Idx, Data] : enumerate(GatheredLoads)) {
          if (Idx < Start)
            continue;
          ToAdd.clear();
          if (LI->getParent() != Data.front().first->getParent() ||
              LI->getType() != Data.front().first->getType())
            continue;
          std::optional<int64_t> Dist =
              getPointersDiff(LI->getType(), LI->getPointerOperand(),
                              Data.front().first->getType(),
                              Data.front().first->getPointerOperand(), DL, SE,
                              /*StrictCheck=*/true);
          if (!Dist)
            continue;
          SmallSet<int64_t, 4> DataDists;
          SmallPtrSet<LoadInst *, 4> DataLoads;
          for (std::pair<LoadInst *, int64_t> P : Data) {
            DataDists.insert(P.second);
            DataLoads.insert(P.first);
          }
          // Found matching gathered loads - check if all loads are unique or
          // can be effectively vectorized.
          unsigned NumUniques = 0;
          for (auto [Cnt, Pair] : enumerate(Loads)) {
            bool Used = DataLoads.contains(Pair.first);
            if (!Used && !DataDists.contains(*Dist + Pair.second)) {
              ++NumUniques;
              ToAdd.insert(Cnt);
            } else if (Used) {
              Repeated.insert(Cnt);
            }
          }
          if (NumUniques > 0 &&
              (Loads.size() == NumUniques ||
               (Loads.size() - NumUniques >= 2 &&
                Loads.size() - NumUniques >= Loads.size() / 2 &&
                (has_single_bit(Data.size() + NumUniques) ||
                 bit_ceil(Data.size()) <
                     bit_ceil(Data.size() + NumUniques))))) {
            Offset = *Dist;
            Start = Idx + 1;
            return std::next(GatheredLoads.begin(), Idx);
          }
        }
        ToAdd.clear();
        return GatheredLoads.end();
      };
  for (ArrayRef<std::pair<LoadInst *, int64_t>> Data : ClusteredLoads) {
    unsigned Start = 0;
    SetVector<unsigned> ToAdd, LocalToAdd, Repeated;
    int64_t Offset = 0;
    auto *It = FindMatchingLoads(Data, GatheredLoads, LocalToAdd, Repeated,
                                 Offset, Start);
    while (It != GatheredLoads.end()) {
      assert(!LocalToAdd.empty() && "Expected some elements to add.");
      for (unsigned Idx : LocalToAdd)
        It->emplace_back(Data[Idx].first, Data[Idx].second + Offset);
      ToAdd.insert_range(LocalToAdd);
      It = FindMatchingLoads(Data, GatheredLoads, LocalToAdd, Repeated, Offset,
                             Start);
    }
    if (any_of(seq<unsigned>(Data.size()), [&](unsigned Idx) {
          return !ToAdd.contains(Idx) && !Repeated.contains(Idx);
        })) {
      auto AddNewLoads =
          [&](SmallVectorImpl<std::pair<LoadInst *, int64_t>> &Loads) {
            for (unsigned Idx : seq<unsigned>(Data.size())) {
              if (ToAdd.contains(Idx) || Repeated.contains(Idx))
                continue;
              Loads.push_back(Data[Idx]);
            }
          };
      if (!AddNew) {
        LoadInst *LI = Data.front().first;
        It = find_if(
            GatheredLoads, [&](ArrayRef<std::pair<LoadInst *, int64_t>> PD) {
              return PD.front().first->getParent() == LI->getParent() &&
                     PD.front().first->getType() == LI->getType();
            });
        while (It != GatheredLoads.end()) {
          AddNewLoads(*It);
          It = std::find_if(
              std::next(It), GatheredLoads.end(),
              [&](ArrayRef<std::pair<LoadInst *, int64_t>> PD) {
                return PD.front().first->getParent() == LI->getParent() &&
                       PD.front().first->getType() == LI->getType();
              });
        }
      }
      GatheredLoads.emplace_back().append(Data.begin(), Data.end());
      AddNewLoads(GatheredLoads.emplace_back());
    }
  }
}
 
void BoUpSLP::tryToVectorizeGatheredLoads(
    const SmallMapVector<
        std::tuple<BasicBlock *, Value *, Type *>,
        SmallVector<SmallVector<std::pair<LoadInst *, int64_t>>>, 8>
        &GatheredLoads) {
  GatheredLoadsEntriesFirst = VectorizableTree.size();
 
  SmallVector<SmallPtrSet<const Value *, 4>> LoadSetsToVectorize(
      LoadEntriesToVectorize.size());
  for (auto [Idx, Set] : zip(LoadEntriesToVectorize, LoadSetsToVectorize))
    Set.insert_range(VectorizableTree[Idx]->Scalars);
 
  // Sort loads by distance.
  auto LoadSorter = [](const std::pair<LoadInst *, int64_t> &L1,
                       const std::pair<LoadInst *, int64_t> &L2) {
    return L1.second > L2.second;
  };
 
  auto IsMaskedGatherSupported = [&, TTI = TTI](ArrayRef<LoadInst *> Loads) {
    ArrayRef<Value *> Values(reinterpret_cast<Value *const *>(Loads.begin()),
                             Loads.size());
    Align Alignment = computeCommonAlignment<LoadInst>(Values);
    auto *Ty = cast<VectorType>(
        getWidenedType(Loads.front()->getType(), Loads.size()));
    return TTI->isLegalMaskedGather(Ty, Alignment) &&
           !TTI->forceScalarizeMaskedGather(Ty, Alignment);
  };
 
  auto GetVectorizedRanges = [this](ArrayRef<LoadInst *> Loads,
                                    BoUpSLP::ValueSet &VectorizedLoads,
                                    SmallVectorImpl<LoadInst *> &NonVectorized,
                                    bool Final, unsigned MaxVF) {
    SmallVector<std::pair<ArrayRef<Value *>, LoadsState>> Results;
    unsigned StartIdx = 0;
    SmallVector<int> CandidateVFs;
    if (isAllowedNonPowerOf2VF(MaxVF))
      CandidateVFs.push_back(MaxVF);
    for (int NumElts = getFloorFullVectorNumberOfElements(
             *TTI, Loads.front()->getType(), MaxVF, SLPReVec);
         NumElts > 1;
         NumElts = getFloorFullVectorNumberOfElements(
             *TTI, Loads.front()->getType(), NumElts - 1, SLPReVec)) {
      CandidateVFs.push_back(NumElts);
      if (VectorizeNonPowerOf2 && NumElts > 2)
        CandidateVFs.push_back(NumElts - 1);
    }
 
    if (Final && CandidateVFs.empty())
      return Results;
 
    unsigned BestVF = Final ? CandidateVFs.back() : 0;
    for (unsigned NumElts : CandidateVFs) {
      if (Final && NumElts > BestVF)
        continue;
      SmallVector<unsigned> MaskedGatherVectorized;
      for (unsigned Cnt = StartIdx, E = Loads.size(); Cnt < E;
           ++Cnt) {
        ArrayRef<LoadInst *> Slice =
            ArrayRef(Loads).slice(Cnt, std::min(NumElts, E - Cnt));
        if (VectorizedLoads.count(Slice.front()) ||
            VectorizedLoads.count(Slice.back()) ||
            areKnownNonVectorizableLoads(Slice))
          continue;
        // Check if it is profitable to try vectorizing gathered loads. It is
        // profitable if we have more than 3 consecutive loads or if we have
        // less but all users are vectorized or deleted.
        bool AllowToVectorize = false;
        // Check if it is profitable to vectorize 2-elements loads.
        if (NumElts == 2) {
          bool IsLegalBroadcastLoad = TTI->isLegalBroadcastLoad(
              Slice.front()->getType(), ElementCount::getFixed(NumElts));
          auto CheckIfAllowed = [=](ArrayRef<LoadInst *> Slice) {
            for (LoadInst *LI : Slice) {
              // If single use/user - allow to vectorize.
              if (LI->hasOneUse())
                continue;
              // 1. Check if number of uses equals number of users.
              // 2. All users are deleted.
              // 3. The load broadcasts are not allowed or the load is not
              // broadcasted.
              if (static_cast<unsigned int>(std::distance(
                      LI->user_begin(), LI->user_end())) != LI->getNumUses())
                return false;
              if (!IsLegalBroadcastLoad)
                continue;
              if (LI->hasNUsesOrMore(UsesLimit))
                return false;
              for (User *U : LI->users()) {
                if (auto *UI = dyn_cast<Instruction>(U); UI && isDeleted(UI))
                  continue;
                for (const TreeEntry *UTE : getTreeEntries(U)) {
                  for (int I : seq<int>(UTE->getNumOperands())) {
                    if (all_of(UTE->getOperand(I), [LI](Value *V) {
                          return V == LI || isa<PoisonValue>(V);
                        }))
                      // Found legal broadcast - do not vectorize.
                      return false;
                  }
                }
              }
            }
            return true;
          };
          AllowToVectorize = CheckIfAllowed(Slice);
        } else {
          AllowToVectorize =
              NumElts >= 3 ||
              any_of(ValueToGatherNodes.at(Slice.front()),
                     [=](const TreeEntry *TE) {
                       return TE->Scalars.size() == 2 &&
                              ((TE->Scalars.front() == Slice.front() &&
                                TE->Scalars.back() == Slice.back()) ||
                               (TE->Scalars.front() == Slice.back() &&
                                TE->Scalars.back() == Slice.front()));
                     });
        }
        if (AllowToVectorize) {
          SmallVector<Value *> PointerOps;
          OrdersType CurrentOrder;
          // Try to build vector load.
          ArrayRef<Value *> Values(
              reinterpret_cast<Value *const *>(Slice.begin()), Slice.size());
          StridedPtrInfo SPtrInfo;
          LoadsState LS = canVectorizeLoads(Values, Slice.front(), CurrentOrder,
                                            PointerOps, SPtrInfo, &BestVF);
          if (LS != LoadsState::Gather ||
              (BestVF > 1 && static_cast<unsigned>(NumElts) == 2 * BestVF)) {
            if (LS == LoadsState::ScatterVectorize) {
              if (MaskedGatherVectorized.empty() ||
                  Cnt >= MaskedGatherVectorized.back() + NumElts)
                MaskedGatherVectorized.push_back(Cnt);
              continue;
            }
            if (LS != LoadsState::Gather) {
              Results.emplace_back(Values, LS);
              VectorizedLoads.insert_range(Slice);
              // If we vectorized initial block, no need to try to vectorize it
              // again.
              if (Cnt == StartIdx)
                StartIdx += NumElts;
            }
            // Check if the whole array was vectorized already - exit.
            if (StartIdx >= Loads.size())
              break;
            // Erase last masked gather candidate, if another candidate within
            // the range is found to be better.
            if (!MaskedGatherVectorized.empty() &&
                Cnt < MaskedGatherVectorized.back() + NumElts)
              MaskedGatherVectorized.pop_back();
            Cnt += NumElts - 1;
            continue;
          }
        }
        if (!AllowToVectorize || BestVF == 0)
          registerNonVectorizableLoads(Slice);
      }
      // Mark masked gathers candidates as vectorized, if any.
      for (unsigned Cnt : MaskedGatherVectorized) {
        ArrayRef<LoadInst *> Slice = ArrayRef(Loads).slice(
            Cnt, std::min<unsigned>(NumElts, Loads.size() - Cnt));
        ArrayRef<Value *> Values(
            reinterpret_cast<Value *const *>(Slice.begin()), Slice.size());
        Results.emplace_back(Values, LoadsState::ScatterVectorize);
        VectorizedLoads.insert_range(Slice);
        // If we vectorized initial block, no need to try to vectorize it again.
        if (Cnt == StartIdx)
          StartIdx += NumElts;
      }
    }
    for (LoadInst *LI : Loads) {
      if (!VectorizedLoads.contains(LI))
        NonVectorized.push_back(LI);
    }
    return Results;
  };
  auto ProcessGatheredLoads =
      [&, &TTI = *TTI](
          ArrayRef<SmallVector<std::pair<LoadInst *, int64_t>>> GatheredLoads,
          bool Final = false) {
        SmallVector<LoadInst *> NonVectorized;
        for (ArrayRef<std::pair<LoadInst *, int64_t>> LoadsDists :
             GatheredLoads) {
          if (LoadsDists.size() <= 1) {
            NonVectorized.push_back(LoadsDists.back().first);
            continue;
          }
          SmallVector<std::pair<LoadInst *, int64_t>> LocalLoadsDists(
              LoadsDists);
          SmallVector<LoadInst *> OriginalLoads(make_first_range(LoadsDists));
          stable_sort(LocalLoadsDists, LoadSorter);
          SmallVector<LoadInst *> Loads;
          unsigned MaxConsecutiveDistance = 0;
          unsigned CurrentConsecutiveDist = 1;
          int64_t LastDist = LocalLoadsDists.front().second;
          bool AllowMaskedGather = IsMaskedGatherSupported(OriginalLoads);
          for (const std::pair<LoadInst *, int64_t> &L : LocalLoadsDists) {
            if (isVectorized(L.first))
              continue;
            assert(LastDist >= L.second &&
                   "Expected first distance always not less than second");
            if (static_cast<uint64_t>(LastDist - L.second) ==
                CurrentConsecutiveDist) {
              ++CurrentConsecutiveDist;
              MaxConsecutiveDistance =
                  std::max(MaxConsecutiveDistance, CurrentConsecutiveDist);
              Loads.push_back(L.first);
              continue;
            }
            if (!AllowMaskedGather && CurrentConsecutiveDist == 1 &&
                !Loads.empty())
              Loads.pop_back();
            CurrentConsecutiveDist = 1;
            LastDist = L.second;
            Loads.push_back(L.first);
          }
          if (Loads.size() <= 1)
            continue;
          if (AllowMaskedGather)
            MaxConsecutiveDistance = Loads.size();
          else if (MaxConsecutiveDistance < 2)
            continue;
          BoUpSLP::ValueSet VectorizedLoads;
          SmallVector<LoadInst *> SortedNonVectorized;
          SmallVector<std::pair<ArrayRef<Value *>, LoadsState>> Results =
              GetVectorizedRanges(Loads, VectorizedLoads, SortedNonVectorized,
                                  Final, MaxConsecutiveDistance);
          if (!Results.empty() && !SortedNonVectorized.empty() &&
              OriginalLoads.size() == Loads.size() &&
              MaxConsecutiveDistance == Loads.size() &&
              all_of(Results,
                     [](const std::pair<ArrayRef<Value *>, LoadsState> &P) {
                       return P.second == LoadsState::ScatterVectorize;
                     })) {
            VectorizedLoads.clear();
            SmallVector<LoadInst *> UnsortedNonVectorized;
            SmallVector<std::pair<ArrayRef<Value *>, LoadsState>>
                UnsortedResults =
                    GetVectorizedRanges(OriginalLoads, VectorizedLoads,
                                        UnsortedNonVectorized, Final,
                                        OriginalLoads.size());
            if (SortedNonVectorized.size() >= UnsortedNonVectorized.size()) {
              SortedNonVectorized.swap(UnsortedNonVectorized);
              Results.swap(UnsortedResults);
            }
          }
          for (auto [Slice, _] : Results) {
            LLVM_DEBUG(dbgs() << "SLP: Trying to vectorize gathered loads ("
                              << Slice.size() << ")\n");
            if (any_of(Slice, [&](Value *V) { return isVectorized(V); })) {
              for (Value *L : Slice)
                if (!isVectorized(L))
                  SortedNonVectorized.push_back(cast<LoadInst>(L));
              continue;
            }
 
            // Select maximum VF as a maximum of user gathered nodes and
            // distance between scalar loads in these nodes.
            unsigned MaxVF = Slice.size();
            unsigned UserMaxVF = 0;
            unsigned InterleaveFactor = 0;
            if (MaxVF == 2) {
              UserMaxVF = MaxVF;
            } else {
              // Found distance between segments of the interleaved loads.
              std::optional<unsigned> InterleavedLoadsDistance = 0;
              unsigned Order = 0;
              std::optional<unsigned> CommonVF = 0;
              DenseMap<const TreeEntry *, unsigned> EntryToPosition;
              SmallPtrSet<const TreeEntry *, 8> DeinterleavedNodes;
              for (auto [Idx, V] : enumerate(Slice)) {
                for (const TreeEntry *E : ValueToGatherNodes.at(V)) {
                  UserMaxVF = std::max<unsigned>(UserMaxVF, E->Scalars.size());
                  unsigned Pos =
                      EntryToPosition.try_emplace(E, Idx).first->second;
                  UserMaxVF = std::max<unsigned>(UserMaxVF, Idx - Pos + 1);
                  if (CommonVF) {
                    if (*CommonVF == 0) {
                      CommonVF = E->Scalars.size();
                      continue;
                    }
                    if (*CommonVF != E->Scalars.size())
                      CommonVF.reset();
                  }
                  // Check if the load is the part of the interleaved load.
                  if (Pos != Idx && InterleavedLoadsDistance) {
                    if (!DeinterleavedNodes.contains(E) &&
                        any_of(E->Scalars, [&, Slice = Slice](Value *V) {
                          if (isa<Constant>(V))
                            return false;
                          if (isVectorized(V))
                            return true;
                          const auto &Nodes = ValueToGatherNodes.at(V);
                          return (Nodes.size() != 1 || !Nodes.contains(E)) &&
                                 !is_contained(Slice, V);
                        })) {
                      InterleavedLoadsDistance.reset();
                      continue;
                    }
                    DeinterleavedNodes.insert(E);
                    if (*InterleavedLoadsDistance == 0) {
                      InterleavedLoadsDistance = Idx - Pos;
                      continue;
                    }
                    if ((Idx - Pos) % *InterleavedLoadsDistance != 0 ||
                        (Idx - Pos) / *InterleavedLoadsDistance < Order)
                      InterleavedLoadsDistance.reset();
                    Order = (Idx - Pos) / InterleavedLoadsDistance.value_or(1);
                  }
                }
              }
              DeinterleavedNodes.clear();
              // Check if the large load represents interleaved load operation.
              if (InterleavedLoadsDistance.value_or(0) > 1 &&
                  CommonVF.value_or(0) != 0) {
                InterleaveFactor = bit_ceil(*InterleavedLoadsDistance);
                unsigned VF = *CommonVF;
                OrdersType Order;
                SmallVector<Value *> PointerOps;
                StridedPtrInfo SPtrInfo;
                // Segmented load detected - vectorize at maximum vector factor.
                if (InterleaveFactor <= Slice.size() &&
                    TTI.isLegalInterleavedAccessType(
                        cast<VectorType>(
                            getWidenedType(Slice.front()->getType(), VF)),
                        InterleaveFactor,
                        cast<LoadInst>(Slice.front())->getAlign(),
                        cast<LoadInst>(Slice.front())
                            ->getPointerAddressSpace()) &&
                    canVectorizeLoads(Slice, Slice.front(), Order, PointerOps,
                                      SPtrInfo) == LoadsState::Vectorize) {
                  UserMaxVF = InterleaveFactor * VF;
                } else {
                  InterleaveFactor = 0;
                }
              }
              // Cannot represent the loads as consecutive vectorizable nodes -
              // just exit.
              unsigned ConsecutiveNodesSize = 0;
              if (!LoadEntriesToVectorize.empty() && InterleaveFactor == 0 &&
                  any_of(zip(LoadEntriesToVectorize, LoadSetsToVectorize),
                         [&, Slice = Slice](const auto &P) {
                           const auto *It = find_if(Slice, [&](Value *V) {
                             return std::get<1>(P).contains(V);
                           });
                           if (It == Slice.end())
                             return false;
                           const TreeEntry &TE =
                               *VectorizableTree[std::get<0>(P)];
                           ArrayRef<Value *> VL = TE.Scalars;
                           OrdersType Order;
                           SmallVector<Value *> PointerOps;
                           StridedPtrInfo SPtrInfo;
                           LoadsState State = canVectorizeLoads(
                               VL, VL.front(), Order, PointerOps, SPtrInfo);
                           if (State == LoadsState::ScatterVectorize ||
                               State == LoadsState::CompressVectorize ||
                               State == LoadsState::BlendedLoadVectorize)
                             return false;
                           ConsecutiveNodesSize += VL.size();
                           size_t Start = std::distance(Slice.begin(), It);
                           size_t Sz = Slice.size() - Start;
                           return Sz < VL.size() ||
                                  Slice.slice(Start, VL.size()) != VL;
                         }))
                continue;
              // Try to build long masked gather loads.
              UserMaxVF = bit_ceil(UserMaxVF);
              if (InterleaveFactor == 0 &&
                  any_of(seq<unsigned>(Slice.size() / UserMaxVF),
                         [&, Slice = Slice](unsigned Idx) {
                           OrdersType Order;
                           SmallVector<Value *> PointerOps;
                           StridedPtrInfo SPtrInfo;
                           return canVectorizeLoads(
                                      Slice.slice(Idx * UserMaxVF, UserMaxVF),
                                      Slice[Idx * UserMaxVF], Order, PointerOps,
                                      SPtrInfo) == LoadsState::ScatterVectorize;
                         }))
                UserMaxVF = MaxVF;
              if (Slice.size() != ConsecutiveNodesSize)
                MaxVF = std::min<unsigned>(MaxVF, UserMaxVF);
            }
            for (unsigned VF = MaxVF; VF >= 2; VF /= 2) {
              bool IsVectorized = true;
              for (unsigned I = 0, E = Slice.size(); I < E; I += VF) {
                ArrayRef<Value *> SubSlice =
                    Slice.slice(I, std::min(VF, E - I));
                if (isVectorized(SubSlice.front()))
                  continue;
                // Check if the subslice is to be-vectorized entry, which is not
                // equal to entry.
                if (any_of(zip(LoadEntriesToVectorize, LoadSetsToVectorize),
                           [&](const auto &P) {
                             return !SubSlice.equals(
                                        VectorizableTree[std::get<0>(P)]
                                            ->Scalars) &&
                                    set_is_subset(SubSlice, std::get<1>(P));
                           }))
                  continue;
                unsigned Sz = VectorizableTree.size();
                // A chunk smaller than InterleaveFactor cannot form an
                // interleave group; keep it non-interleaved instead.
                buildTreeRec(
                    SubSlice, 0, EdgeInfo(),
                    SubSlice.size() >= InterleaveFactor ? InterleaveFactor : 0);
                if (Sz == VectorizableTree.size()) {
                  IsVectorized = false;
                  // Try non-interleaved vectorization with smaller vector
                  // factor.
                  if (InterleaveFactor > 0) {
                    VF = 2 * (MaxVF / InterleaveFactor);
                    InterleaveFactor = 0;
                  }
                  continue;
                }
              }
              if (IsVectorized)
                break;
            }
          }
          NonVectorized.append(SortedNonVectorized);
        }
        return NonVectorized;
      };
  for (const auto &GLs : GatheredLoads) {
    const auto &Ref = GLs.second;
    SmallVector<LoadInst *> NonVectorized = ProcessGatheredLoads(Ref);
    if (!Ref.empty() && !NonVectorized.empty() &&
        accumulate(
            Ref, 0u,
            [](unsigned S, ArrayRef<std::pair<LoadInst *, int64_t>> LoadsDists)
                -> unsigned { return S + LoadsDists.size(); }) !=
            NonVectorized.size() &&
        IsMaskedGatherSupported(NonVectorized)) {
      SmallVector<SmallVector<std::pair<LoadInst *, int64_t>>>
          FinalGatheredLoads;
      for (LoadInst *LI : NonVectorized) {
        // Reinsert non-vectorized loads to other list of loads with the same
        // base pointers.
        gatherPossiblyVectorizableLoads(*this, LI, *DL, *SE, *TTI,
                                        FinalGatheredLoads,
                                        /*AddNew=*/false);
      }
      // Final attempt to vectorize non-vectorized loads.
      (void)ProcessGatheredLoads(FinalGatheredLoads, /*Final=*/true);
    }
  }
  // Try to vectorize postponed load entries, previously marked as gathered.
  for (unsigned Idx : LoadEntriesToVectorize) {
    const TreeEntry &E = *VectorizableTree[Idx];
    SmallVector<Value *> GatheredScalars(E.Scalars.begin(), E.Scalars.end());
    // Avoid reordering, if possible.
    if (!E.ReorderIndices.empty()) {
      // Build a mask out of the reorder indices and reorder scalars per this
      // mask.
      SmallVector<int> ReorderMask;
      inversePermutation(E.ReorderIndices, ReorderMask);
      reorderScalars(GatheredScalars, ReorderMask);
    }
    buildTreeRec(GatheredScalars, 0, EdgeInfo());
  }
  // If no new entries created, consider it as no gathered loads entries must be
  // handled.
  if (static_cast<unsigned>(*GatheredLoadsEntriesFirst) ==
      VectorizableTree.size())
    GatheredLoadsEntriesFirst.reset();
}
 
/// Generates key/subkey pair for the given value to provide effective sorting
/// of the values and better detection of the vectorizable values sequences. The
/// keys/subkeys can be used for better sorting of the values themselves (keys)
/// and in values subgroups (subkeys).
static std::pair<size_t, size_t> generateKeySubkey(
    Value *V, const TargetLibraryInfo *TLI,
    function_ref<hash_code(size_t, LoadInst *)> LoadsSubkeyGenerator,
    bool AllowAlternate) {
  hash_code Key = hash_value(V->getValueID() + 2);
  hash_code SubKey = hash_value(0);
  // Sort the loads by the distance between the pointers.
  if (auto *LI = dyn_cast<LoadInst>(V)) {
    Key = hash_combine(LI->getType(), hash_value(Instruction::Load), Key);
    if (LI->isSimple())
      SubKey = hash_value(LoadsSubkeyGenerator(Key, LI));
    else
      Key = SubKey = hash_value(LI);
  } else if (isVectorLikeInstWithConstOps(V)) {
    // Sort extracts by the vector operands.
    if (isa<ExtractElementInst, UndefValue>(V))
      Key = hash_value(Value::UndefValueVal + 1);
    if (auto *EI = dyn_cast<ExtractElementInst>(V)) {
      if (!isUndefVector(EI->getVectorOperand()).all() &&
          !isa<UndefValue>(EI->getIndexOperand()))
        SubKey = hash_value(EI->getVectorOperand());
    }
  } else if (auto *I = dyn_cast<Instruction>(V)) {
    // Sort other instructions just by the opcodes except for CMPInst.
    // For CMP also sort by the predicate kind.
    if ((isa<BinaryOperator, CastInst>(I)) &&
        isValidForAlternation(I->getOpcode())) {
      if (AllowAlternate)
        Key = hash_value(isa<BinaryOperator>(I) ? 1 : 0);
      else
        Key = hash_combine(hash_value(I->getOpcode()), Key);
      SubKey = hash_combine(
          hash_value(I->getOpcode()), hash_value(I->getType()),
          hash_value(isa<BinaryOperator>(I)
                         ? I->getType()
                         : cast<CastInst>(I)->getOperand(0)->getType()));
      // For casts, look through the only operand to improve compile time.
      if (isa<CastInst>(I)) {
        std::pair<size_t, size_t> OpVals =
            generateKeySubkey(I->getOperand(0), TLI, LoadsSubkeyGenerator,
                              /*AllowAlternate=*/true);
        Key = hash_combine(OpVals.first, Key);
        SubKey = hash_combine(OpVals.first, SubKey);
      }
    } else if (auto *CI = dyn_cast<CmpInst>(I)) {
      CmpInst::Predicate Pred = CI->getPredicate();
      if (CI->isCommutative())
        Pred = std::min(Pred, CmpInst::getInversePredicate(Pred));
      CmpInst::Predicate SwapPred = CmpInst::getSwappedPredicate(Pred);
      SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Pred),
                            hash_value(SwapPred),
                            hash_value(CI->getOperand(0)->getType()));
    } else if (auto *Call = dyn_cast<CallInst>(I)) {
      Intrinsic::ID ID = getVectorIntrinsicIDForCall(Call, TLI);
      if (isTriviallyVectorizable(ID)) {
        if (ID == Intrinsic::fmuladd)
          ID = Intrinsic::fma;
        SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(ID),
                              hash_value(I->getType()));
      } else if (!VFDatabase(*Call).getMappings(*Call).empty()) {
        SubKey = hash_combine(hash_value(I->getOpcode()),
                              hash_value(Call->getCalledFunction()),
                              hash_value(I->getType()));
      } else {
        Key = hash_combine(hash_value(Call), Key);
        SubKey = hash_combine(hash_value(I->getOpcode()), hash_value(Call));
      }
      for (const CallBase::BundleOpInfo &Op : Call->bundle_op_infos())
        SubKey = hash_combine(hash_value(Op.Begin), hash_value(Op.End),
                              hash_value(Op.Tag), SubKey);
    } else if (auto *Gep = dyn_cast<GetElementPtrInst>(I)) {
      if (Gep->getNumOperands() == 2 && isa<ConstantInt>(Gep->getOperand(1)))
        SubKey = hash_value(Gep->getPointerOperand());
      else
        SubKey = hash_value(Gep);
    } else if (BinaryOperator::isIntDivRem(I->getOpcode()) &&
               !isa<ConstantInt>(I->getOperand(1))) {
      // Do not try to vectorize instructions with potentially high cost.
      SubKey = hash_value(I);
    } else {
      SubKey =
          hash_combine(hash_value(I->getOpcode()), hash_value(I->getType()));
    }
    Key = hash_combine(hash_value(I->getParent()->getNumber()), Key);
  }
  return std::make_pair(Key, SubKey);
}
 
/// Checks if the specified instruction \p I is an main operation for the given
/// \p MainOp and \p AltOp instructions.
static bool isMainInstruction(Instruction *I, Instruction *MainOp,
                              Instruction *AltOp, const TargetLibraryInfo &TLI);
 
/// Builds the arguments types vector for the given call instruction with the
/// given \p ID for the specified vector factor.
static SmallVector<Type *>
buildIntrinsicArgTypes(const CallInst *CI, const Intrinsic::ID ID,
                       const unsigned VF, unsigned MinBW,
                       const TargetTransformInfo *TTI) {
  SmallVector<Type *> ArgTys;
  for (auto [Idx, Arg] : enumerate(CI->args())) {
    if (ID != Intrinsic::not_intrinsic) {
      if (isVectorIntrinsicWithScalarOpAtArg(ID, Idx, TTI)) {
        ArgTys.push_back(Arg->getType());
        continue;
      }
      if (MinBW > 0) {
        ArgTys.push_back(
            getWidenedType(IntegerType::get(CI->getContext(), MinBW), VF));
        continue;
      }
    }
    ArgTys.push_back(getWidenedType(Arg->getType(), VF));
  }
  return ArgTys;
}
 
/// Calculates the costs of vectorized intrinsic (if possible) and vectorized
/// function (if possible) calls. Returns invalid cost for the corresponding
/// calls, if they cannot be vectorized/will be scalarized.
static std::pair<InstructionCost, InstructionCost>
getVectorCallCosts(CallInst *CI, Type *VecTy, const TargetTransformInfo *TTI,
                   const TargetLibraryInfo *TLI, ArrayRef<Type *> ArgTys,
                   const TTI::TargetCostKind CostKind) {
  auto Shape = VFShape::get(CI->getFunctionType(),
                            ElementCount::getFixed(getNumElements(VecTy)),
                            false /*HasGlobalPred*/);
  Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape);
  auto LibCost = InstructionCost::getInvalid();
  if (!CI->isNoBuiltin() && VecFunc) {
    // Calculate the cost of the vector library call.
    // If the corresponding vector call is cheaper, return its cost.
    LibCost = TTI->getCallInstrCost(nullptr, VecTy, ArgTys, CostKind);
  }
  Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
 
  // Calculate the cost of the vector intrinsic call.
  FastMathFlags FMF;
  if (auto *FPCI = dyn_cast<FPMathOperator>(CI))
    FMF = FPCI->getFastMathFlags();
  const InstructionCost ScalarLimit = 10000;
  IntrinsicCostAttributes CostAttrs(ID, VecTy, ArgTys, FMF, nullptr,
                                    LibCost.isValid() ? LibCost : ScalarLimit);
  auto IntrinsicCost = TTI->getIntrinsicInstrCost(CostAttrs, CostKind);
  if (LibCost.isValid()) {
    if (IntrinsicCost > LibCost)
      IntrinsicCost = InstructionCost::getInvalid();
  } else if (IntrinsicCost > ScalarLimit) {
    // A type-based query always scalarizes struct-returning intrinsics (e.g.
    // llvm.sincos), which do not have a VFDatabase name mapping. Retry with an
    // argument-aware query (as the loop vectorizer does) so such lowerings are
    // taken into account.
    SmallVector<const Value *> Args(CI->args());
    IntrinsicCostAttributes ArgAwareAttrs(
        ID, VecTy, Args, ArgTys, FMF, dyn_cast<IntrinsicInst>(CI), ScalarLimit);
    IntrinsicCost = TTI->getIntrinsicInstrCost(ArgAwareAttrs, CostKind);
    if (IntrinsicCost > ScalarLimit)
      IntrinsicCost = InstructionCost::getInvalid();
  }
 
  return {IntrinsicCost, LibCost};
}
 
/// \returns the reciprocal-throughput cost of \p I widened to \p VF lanes (an
/// arithmetic op or a vectorizable call).
static InstructionCost getVectorOpCost(Instruction *I, unsigned VF,
                                       const TargetTransformInfo &TTI,
                                       const TargetLibraryInfo &TLI,
                                       const TTI::TargetCostKind CostKind) {
  assert((isa<BinaryOperator, CallInst>(I)) &&
         "getVectorOpCost expects an arithmetic op or a vectorizable call.");
  Type *VecTy = getWidenedType(I->getType(), VF);
  if (auto *CI = dyn_cast<CallInst>(I)) {
    Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &TLI);
    SmallVector<Type *> ArgTys = buildIntrinsicArgTypes(CI, ID, VF, 0, &TTI);
    auto [IntrCost, LibCost] =
        getVectorCallCosts(CI, VecTy, &TTI, &TLI, ArgTys, CostKind);
    return std::min(IntrCost, LibCost);
  }
  return TTI.getArithmeticInstrCost(I->getOpcode(), VecTy, CostKind);
}
 
/// Packs a type's kind and scalar width into one key, so an opcode/intrinsic
/// id reused across scalar widths (e.g. fdiv float vs. fdiv double) is cached
/// per width instead of sharing one verdict.
static unsigned getPoorThroughputTypeKey(Type *Ty) {
  return (static_cast<unsigned>(Ty->getTypeID()) << 16) |
         Ty->getScalarSizeInBits();
}
 
namespace {
/// Memoizes opcode / intrinsic / callee kinds already ruled out as seeds.
struct PoorThroughputOpCache {
  SmallDenseSet<std::pair<unsigned, unsigned>> CheapOpcodes;
  SmallDenseSet<std::pair<Intrinsic::ID, unsigned>> CheapIntrinsics;
  // Keyed on the callee alone: a Function's signature is fixed, so its type
  // never varies between calls.
  SmallDenseSet<const Function *> CheapCallees;
};
 
/// Groups poor-throughput seeds for tryToVectorizeSequence.
struct SeedGroupKey {
  unsigned Opcode = 0;
  Intrinsic::ID IntrID = Intrinsic::not_intrinsic;
  StringRef CalleeName;
 
  bool less(const SeedGroupKey &O) const {
    if (Opcode != O.Opcode)
      return Opcode < O.Opcode;
    if (IntrID != O.IntrID)
      return IntrID < O.IntrID;
    return CalleeName < O.CalleeName;
  }
};
 
static SeedGroupKey getSeedGroupKey(const Instruction *I,
                                    const TargetLibraryInfo &TLI) {
  SeedGroupKey K;
  K.Opcode = I->getOpcode();
  if (auto *CI = dyn_cast<CallInst>(I)) {
    K.IntrID = getVectorIntrinsicIDForCall(CI, &TLI);
    if (const Function *F = CI->getCalledFunction())
      K.CalleeName = F->getName();
  }
  return K;
}
} // namespace
 
/// Returns true if \p I is an expensive scalar op whose vector form is cheaper
/// per lane (e.g. fdiv, frem, fsqrt).
static bool isPoorThroughputOp(Instruction *I, const TargetTransformInfo &TTI,
                               const TargetLibraryInfo &TLI,
                               PoorThroughputOpCache &Cache,
                               const TTI::TargetCostKind CostKind) {
  if (!isa<BinaryOperator, CallInst>(I))
    return false;
  Type *Ty = I->getType();
  if ((Ty->isVectorTy() && !SLPReVec) || Ty->isAggregateType() ||
      !isValidElementType(Ty, SLPReVec))
    return false;
  auto Analyze = [&]() {
    InstructionCost ScalarCost = TTI.getInstructionCost(I, CostKind);
    if (ScalarCost < TTI::TCC_Expensive)
      return false;
    constexpr unsigned MinVF = 2;
    return getVectorOpCost(I, MinVF, TTI, TLI, CostKind) < ScalarCost * MinVF;
  };
  auto CheckCached = [&](bool IsCheap, llvm::function_ref<void()> MarkCheap) {
    if (IsCheap)
      return false;
    if (Analyze())
      return true;
    MarkCheap();
    return false;
  };
  if (auto *CI = dyn_cast<CallInst>(I)) {
    if (any_of(CI->args(), [](const Value *Arg) {
          return !isValidElementType(Arg->getType(), SLPReVec);
        }))
      return false;
    if (Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &TLI)) {
      auto Key = std::make_pair(ID, getPoorThroughputTypeKey(Ty));
      return CheckCached(Cache.CheapIntrinsics.contains(Key),
                         [&] { Cache.CheapIntrinsics.insert(Key); });
    }
    const Function *Callee = CI->getCalledFunction();
    if (!Callee)
      return Analyze();
    return CheckCached(Cache.CheapCallees.contains(Callee),
                       [&] { Cache.CheapCallees.insert(Callee); });
  }
  auto Key = std::make_pair(I->getOpcode(), getPoorThroughputTypeKey(Ty));
  return CheckCached(Cache.CheapOpcodes.contains(Key),
                     [&] { Cache.CheapOpcodes.insert(Key); });
}
 
/// Find the innermost loop starting from \p L, for which at least a single
/// value in \p VL is not invariant.
static const Loop *findInnermostNonInvariantLoop(const Loop *L,
                                                 ArrayRef<Value *> VL) {
  assert(L && "Expected valid loop");
  auto IsLoopInvariant = [&](const Loop *L, ArrayRef<Value *> VL) {
    return all_of(VL, [&](Value *V) {
      return isa<Constant>(V) || !isa<Instruction>(V) || L->isLoopInvariant(V);
    });
  };
  while (L && IsLoopInvariant(L, VL))
    L = L->getParentLoop();
  return L;
}
 
/// Get the loop nest for the given loop.
ArrayRef<const Loop *> BoUpSLP::getLoopNest(const Loop *L) {
  assert(L && "Expected valid loop");
  if (LoopAwareTripCount == 0)
    return {};
  SmallVector<const Loop *> &Res =
      LoopToLoopNest.try_emplace(L).first->getSecond();
  if (!Res.empty())
    return Res;
  SmallVector<const Loop *> LoopNest;
  while (L) {
    LoopNest.push_back(L);
    L = L->getParentLoop();
  }
  Res.assign(LoopNest.rbegin(), LoopNest.rend());
  return Res;
}
 
/// Detects an extractvalue bundle that can be widened by vectorizing the
/// underlying struct-returning calls.
///
/// \p VL is a bundle whose state \p S is Instruction::ExtractValue. The
/// bundle is acceptable for widening into one struct-of-vectors call only
/// when:
///   - every element of \p VL is an ExtractValueInst,
///   - every ExtractValueInst extracts the same struct field (its
///     getIndices() matches the main op's indices),
///   - the aggregate operands form a uniform set of CallInsts (per
///     getSameOpcode) that is not an alt-shuffle and whose return type is
///     a literal struct, and
///   - every user of every such call is itself an ExtractValueInst, so the
///     external-use extraction code can rebuild scalars via extractvalue +
///     extractelement without needing an insertvalue chain.
///
/// On success returns true and fills \p Indices with the common field
/// index path and \p Calls with the per-lane aggregate calls (in VL order),
/// for the caller to feed as the operand of the new tree entry. Otherwise
/// returns false and leaves the output parameters untouched.
static bool checkEVsForVecCalls(ArrayRef<Value *> VL,
                                const InstructionsState &S,
                                const TargetLibraryInfo &TLI,
                                SmallVectorImpl<unsigned> &Indices,
                                SmallVectorImpl<Value *> &Calls) {
  assert(S && S.getOpcode() == Instruction::ExtractValue &&
         "Expected extractvalue instruction state.");
  if (!all_of(VL, IsaPred<ExtractValueInst>))
    return false;
  auto *VL0 = cast<ExtractValueInst>(S.getMainOp());
  ArrayRef<unsigned> VL0Indices = VL0->getIndices();
  SmallVector<Value *> Aggregates;
  for (Value *V : VL) {
    if (V == VL0) {
      Aggregates.push_back(VL0->getAggregateOperand());
      continue;
    }
    auto *IV = cast<ExtractValueInst>(V);
    if (IV->getIndices() != VL0Indices ||
        isa<ScalableVectorType>(IV->getType()))
      return false;
    Value *Agg = IV->getAggregateOperand();
    Aggregates.push_back(Agg);
  }
  const InstructionsState AggState = getSameOpcode(Aggregates, TLI);
  if (AggState && AggState.getOpcode() == Instruction::Call &&
      !AggState.isAltShuffle() &&
      isa<StructType>(AggState.getMainOp()->getType()) &&
      allSameBlock(Aggregates)) {
    // The struct-returning call may have non-bundle users too. The external
    // extraction code rebuilds scalars by extractvalue + extractelement,
    // which only works when every user of the call is an ExtractValueInst.
    // Bail out if any aggregate has a different kind of user.
    for (Value *Agg : Aggregates) {
      if (!all_of(Agg->users(), IsaPred<ExtractValueInst>))
        return false;
    }
    Indices.assign(VL0Indices.begin(), VL0Indices.end());
    Calls.swap(Aggregates);
    return true;
  }
  return false;
}
 
static bool allStructUsersAreExtractValueInsts(ArrayRef<Value *> VL) {
  return all_of(VL, [](Value *V) {
    return !isa<StructType>(V->getType()) ||
           all_of(V->users(), IsaPred<ExtractValueInst>);
  });
}
 
BoUpSLP::TreeEntry::EntryState BoUpSLP::getScalarsVectorizationState(
    const InstructionsState &S, ArrayRef<Value *> VL,
    bool IsScatterVectorizeUserTE, OrdersType &CurrentOrder,
    SmallVectorImpl<Value *> &PointerOps, StridedPtrInfo &SPtrInfo,
    SmallVectorImpl<int> &ReuseShuffleIndices) {
  assert(S.getMainOp() &&
         "Expected instructions with same/alternate opcodes only.");
 
  unsigned ShuffleOrOp =
      S.isAltShuffle() ? (unsigned)Instruction::ShuffleVector : S.getOpcode();
  Instruction *VL0 = S.getMainOp();
  switch (ShuffleOrOp) {
  case Instruction::PHI: {
    // Too many operands - gather, most probably won't be vectorized.
    if (VL0->getNumOperands() > MaxPHINumOperands)
      return TreeEntry::NeedToGather;
    // Check for terminator values (e.g. invoke).
    for (Value *V : VL) {
      auto *PHI = dyn_cast<PHINode>(V);
      if (!PHI)
        continue;
      for (Value *Incoming : PHI->incoming_values()) {
        Instruction *Term = dyn_cast<Instruction>(Incoming);
        if (Term && Term->isTerminator()) {
          LLVM_DEBUG(dbgs()
                     << "SLP: Need to swizzle PHINodes (terminator use).\n");
          return TreeEntry::NeedToGather;
        }
      }
    }
 
    return TreeEntry::Vectorize;
  }
  case Instruction::ExtractElement:
    if (any_of(VL, [&](Value *V) {
          if (S.isCopyableElement(V) || isa<PoisonValue>(V))
            return false;
          auto *EI = dyn_cast<ExtractElementInst>(V);
          if (!EI)
            return true;
          Value *Op = EI->getOperand(0);
          if (isVectorized(Op))
            return true;
          auto *OpI = dyn_cast<Instruction>(Op);
          return OpI && OpI->isTerminator();
        }))
      return TreeEntry::NeedToGather;
    [[fallthrough]];
  case Instruction::ExtractValue: {
    bool Reuse = canReuseExtract(VL, CurrentOrder);
    // Copyable lanes are inserted into the reused source vector at their own
    // index, which is correct only for the identity extract order (empty
    // CurrentOrder) and without a reuse shuffle; other cases fall back to
    // gather.
    if (S.areInstructionsWithCopyableElements() &&
        (!Reuse || !ReuseShuffleIndices.empty()))
      return TreeEntry::NeedToGather;
    if (Reuse || !CurrentOrder.empty())
      return TreeEntry::Vectorize;
    SmallVector<unsigned> Indices;
    SmallVector<Value *> Calls;
    if (ShuffleOrOp == Instruction::ExtractValue &&
        checkEVsForVecCalls(VL, S, *TLI, Indices, Calls))
      return TreeEntry::Vectorize;
    LLVM_DEBUG(dbgs() << "SLP: Gather extract sequence.\n");
    return TreeEntry::NeedToGather;
  }
  case Instruction::InsertValue:
    // Handle only simple insertvalue building a homogeneous aggregate from
    // scalars: exactly one index and an inserted value whose type is a valid
    // vector element type. Vectors and aggregates (structs/arrays) are
    // rejected, since the inserted operand type is later widened into the
    // result FixedVectorType.
    // TODO: Support more complex insertvalues.
    if (any_of(VL,
               [](Value *V) {
                 auto *IV = dyn_cast<InsertValueInst>(V);
                 return IV && (IV->getNumIndices() != 1 ||
                               !::isValidElementType(
                                   IV->getOperand(1)->getType(), SLPReVec) ||
                               IV->getOperand(1)->getType()->isVectorTy());
               }) ||
        none_of(VL, [](Value *V) {
          auto *IV = dyn_cast<InsertValueInst>(V);
          return IV && isa<UndefValue>(IV->getAggregateOperand());
        }))
      return TreeEntry::NeedToGather;
    [[fallthrough]];
  case Instruction::InsertElement: {
    // Check that we have a buildvector and not a shuffle of 2 or more
    // different vectors.
    ValueSet SourceVectors;
    for (Value *V : VL) {
      if (isa<PoisonValue>(V)) {
        LLVM_DEBUG(
            dbgs()
            << "SLP: Gather of insertelement/insertvalue/poison vector.\n");
        return TreeEntry::NeedToGather;
      }
      SourceVectors.insert(cast<Instruction>(V)->getOperand(0));
      assert(getElementIndex(V) != std::nullopt &&
             "Non-constant or undef index?");
    }
 
    if (count_if(VL, [&SourceVectors](Value *V) {
          return !SourceVectors.contains(V);
        }) >= 2) {
      // Found 2nd source vector - cancel.
      LLVM_DEBUG(
          dbgs() << "SLP: Gather of insertelement/insertvalue vectors with "
                    "different source vectors.\n");
      return TreeEntry::NeedToGather;
    }
 
    if (any_of(VL, [&SourceVectors](Value *V) {
          // The last InsertElement/InsertValue can have multiple uses.
          return SourceVectors.contains(V) && !V->hasOneUse();
        })) {
      assert((SLPReVec || ShuffleOrOp == Instruction::InsertValue) &&
             "Only supported by REVEC or InsertValue.");
      LLVM_DEBUG(
          dbgs() << "SLP: Gather of insertelement/insertvalue vectors with "
                    "multiple uses.\n");
      return TreeEntry::NeedToGather;
    }
 
    return TreeEntry::Vectorize;
  }
  case Instruction::Load: {
    // Check that a vectorized load would load the same memory as a scalar
    // load. For example, we don't want to vectorize loads that are smaller
    // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM
    // treats loading/storing it as an i8 struct. If we vectorize loads/stores
    // from such a struct, we read/write packed bits disagreeing with the
    // unvectorized version.
    auto IsGatheredNode = [&]() {
      if (!GatheredLoadsEntriesFirst)
        return false;
      return all_of(VL, [&](Value *V) {
        if (isa<PoisonValue>(V))
          return true;
        return any_of(getTreeEntries(V), [&](const TreeEntry *TE) {
          return TE->Idx >= *GatheredLoadsEntriesFirst;
        });
      });
    };
    switch (canVectorizeLoads(VL, VL0, CurrentOrder, PointerOps, SPtrInfo)) {
    case LoadsState::Vectorize:
      return TreeEntry::Vectorize;
    case LoadsState::CompressVectorize:
      if (!IsGraphTransformMode && VectorizableTree.size() > 1) {
        // Delay slow vectorized nodes for better vectorization attempts.
        LoadEntriesToVectorize.insert(VectorizableTree.size());
        return TreeEntry::NeedToGather;
      }
      return IsGatheredNode() ? TreeEntry::NeedToGather
                              : TreeEntry::CompressVectorize;
    case LoadsState::ScatterVectorize:
      if (!IsGraphTransformMode && VectorizableTree.size() > 1) {
        // Delay slow vectorized nodes for better vectorization attempts.
        LoadEntriesToVectorize.insert(VectorizableTree.size());
        return TreeEntry::NeedToGather;
      }
      return IsGatheredNode() ? TreeEntry::NeedToGather
                              : TreeEntry::ScatterVectorize;
    case LoadsState::StridedVectorize:
      if (!IsGraphTransformMode && VectorizableTree.size() > 1) {
        // Delay slow vectorized nodes for better vectorization attempts.
        LoadEntriesToVectorize.insert(VectorizableTree.size());
        return TreeEntry::NeedToGather;
      }
      return IsGatheredNode() ? TreeEntry::NeedToGather
                              : TreeEntry::StridedVectorize;
    case LoadsState::BlendedLoadVectorize:
      if (!IsGraphTransformMode && VectorizableTree.size() > 1) {
        // Delay slow vectorized nodes for better vectorization attempts.
        LoadEntriesToVectorize.insert(VectorizableTree.size());
        return TreeEntry::NeedToGather;
      }
      return IsGatheredNode() ? TreeEntry::NeedToGather
                              : TreeEntry::BlendedLoadVectorize;
    case LoadsState::Gather:
#ifndef NDEBUG
      Type *ScalarTy = VL0->getType();
      if (DL->getTypeSizeInBits(ScalarTy) !=
          DL->getTypeAllocSizeInBits(ScalarTy))
        LLVM_DEBUG(dbgs() << "SLP: Gathering loads of non-packed type.\n");
      else if (any_of(VL, [](Value *V) {
                 auto *LI = dyn_cast<LoadInst>(V);
                 return !LI || !LI->isSimple();
               }))
        LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple loads.\n");
      else
        LLVM_DEBUG(dbgs() << "SLP: Gathering non-consecutive loads.\n");
#endif // NDEBUG
      registerNonVectorizableLoads(VL);
      return TreeEntry::NeedToGather;
    }
    llvm_unreachable("Unexpected state of loads");
  }
  case Instruction::ZExt:
  case Instruction::SExt:
  case Instruction::FPToUI:
  case Instruction::FPToSI:
  case Instruction::FPExt:
  case Instruction::PtrToInt:
  case Instruction::PtrToAddr:
  case Instruction::IntToPtr:
  case Instruction::SIToFP:
  case Instruction::UIToFP:
  case Instruction::Trunc:
  case Instruction::FPTrunc:
  case Instruction::BitCast: {
    Type *SrcTy = VL0->getOperand(0)->getType();
    for (Value *V : VL) {
      if (isa<PoisonValue>(V))
        continue;
      Type *Ty = cast<Instruction>(V)->getOperand(0)->getType();
      if (Ty != SrcTy || !isValidElementType(Ty, SLPReVec)) {
        LLVM_DEBUG(
            dbgs() << "SLP: Gathering casts with different src types.\n");
        return TreeEntry::NeedToGather;
      }
    }
    return TreeEntry::Vectorize;
  }
  case Instruction::ICmp:
  case Instruction::FCmp: {
    // Check that all of the compares have the same predicate.
    CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate();
    CmpInst::Predicate SwapP0 = CmpInst::getSwappedPredicate(P0);
    Type *ComparedTy = VL0->getOperand(0)->getType();
    for (Value *V : VL) {
      if (isa<PoisonValue>(V))
        continue;
      auto *Cmp = cast<CmpInst>(V);
      if ((Cmp->getPredicate() != P0 && Cmp->getPredicate() != SwapP0 &&
           !CmpSamePredicateHelper::canConvertTo(Cmp, P0)) ||
          Cmp->getOperand(0)->getType() != ComparedTy) {
        LLVM_DEBUG(dbgs() << "SLP: Gathering cmp with different predicate.\n");
        return TreeEntry::NeedToGather;
      }
    }
    return TreeEntry::Vectorize;
  }
  case Instruction::Select:
    if (SLPReVec) {
      SmallPtrSet<Type *, 4> CondTypes;
      for (Value *V : VL) {
        Value *Cond;
        if (!match(V, m_Select(m_Value(Cond), m_Value(), m_Value())) &&
            !match(V, m_ZExt(m_Value(Cond))))
          continue;
        CondTypes.insert(Cond->getType());
      }
      if (CondTypes.size() > 1) {
        LLVM_DEBUG(
            dbgs()
            << "SLP: Gathering select with different condition types.\n");
        return TreeEntry::NeedToGather;
      }
    }
    [[fallthrough]];
  case Instruction::FNeg:
  case Instruction::Add:
  case Instruction::FAdd:
  case Instruction::Sub:
  case Instruction::FSub:
  case Instruction::Mul:
  case Instruction::FMul:
  case Instruction::UDiv:
  case Instruction::SDiv:
  case Instruction::FDiv:
  case Instruction::URem:
  case Instruction::SRem:
  case Instruction::FRem:
  case Instruction::Shl:
  case Instruction::LShr:
  case Instruction::AShr:
  case Instruction::And:
  case Instruction::Or:
  case Instruction::Xor:
  case Instruction::Freeze:
    if (S.getMainOp()->getType()->isFloatingPointTy() &&
        TTI->isFPVectorizationPotentiallyUnsafe() && any_of(VL, [](Value *V) {
          auto *I = dyn_cast<Instruction>(V);
          return I && I->isBinaryOp() && !I->isFast();
        }))
      return TreeEntry::NeedToGather;
    return TreeEntry::Vectorize;
  case Instruction::GetElementPtr: {
    // We don't combine GEPs with complicated (nested) indexing.
    for (Value *V : VL) {
      auto *I = dyn_cast<GetElementPtrInst>(V);
      if (!I)
        continue;
      if (I->getNumOperands() != 2) {
        LLVM_DEBUG(dbgs() << "SLP: not-vectorizable GEP (nested indexes).\n");
        return TreeEntry::NeedToGather;
      }
    }
 
    // We can't combine several GEPs into one vector if they operate on
    // different types.
    Type *Ty0 = cast<GEPOperator>(VL0)->getSourceElementType();
    for (Value *V : VL) {
      auto *GEP = dyn_cast<GEPOperator>(V);
      if (!GEP)
        continue;
      Type *CurTy = GEP->getSourceElementType();
      if (Ty0 != CurTy) {
        LLVM_DEBUG(dbgs() << "SLP: not-vectorizable GEP (different types).\n");
        return TreeEntry::NeedToGather;
      }
    }
 
    // We don't combine GEPs with non-constant indexes.
    Type *Ty1 = VL0->getOperand(1)->getType();
    for (Value *V : VL) {
      auto *I = dyn_cast<GetElementPtrInst>(V);
      if (!I)
        continue;
      auto *Op = I->getOperand(1);
      if ((!IsScatterVectorizeUserTE && !isa<ConstantInt>(Op)) ||
          (Op->getType() != Ty1 &&
           ((IsScatterVectorizeUserTE && !isa<ConstantInt>(Op)) ||
            Op->getType()->getScalarSizeInBits() >
                DL->getIndexSizeInBits(
                    V->getType()->getPointerAddressSpace())))) {
        LLVM_DEBUG(
            dbgs() << "SLP: not-vectorizable GEP (non-constant indexes).\n");
        return TreeEntry::NeedToGather;
      }
    }
 
    return TreeEntry::Vectorize;
  }
  case Instruction::Store: {
    // Check if the stores are consecutive or if we need to swizzle them.
    llvm::Type *ScalarTy = cast<StoreInst>(VL0)->getValueOperand()->getType();
    // Avoid types that are padded when being allocated as scalars, while
    // being packed together in a vector (such as i1).
    if (DL->getTypeSizeInBits(ScalarTy) !=
        DL->getTypeAllocSizeInBits(ScalarTy)) {
      LLVM_DEBUG(dbgs() << "SLP: Gathering stores of non-packed type.\n");
      return TreeEntry::NeedToGather;
    }
    // Make sure all stores in the bundle are simple - we can't vectorize
    // atomic or volatile stores.
    for (Value *V : VL) {
      auto *SI = cast<StoreInst>(V);
      if (!SI->isSimple()) {
        LLVM_DEBUG(dbgs() << "SLP: Gathering non-simple stores.\n");
        return TreeEntry::NeedToGather;
      }
      PointerOps.push_back(SI->getPointerOperand());
    }
 
    // Check the order of pointer operands.
    if (llvm::sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, CurrentOrder)) {
      Value *Ptr0;
      Value *PtrN;
      if (CurrentOrder.empty()) {
        Ptr0 = PointerOps.front();
        PtrN = PointerOps.back();
      } else {
        Ptr0 = PointerOps[CurrentOrder.front()];
        PtrN = PointerOps[CurrentOrder.back()];
      }
      Align CommonAlignment = computeCommonAlignment<StoreInst>(VL);
      std::optional<int64_t> Dist =
          getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, *DL, *SE);
      // Check that the sorted pointer operands are consecutive.
      if (static_cast<uint64_t>(*Dist) == VL.size() - 1)
        return TreeEntry::Vectorize;
      if (EnableStridedStores &&
          analyzeConstantStrideCandidate(PointerOps, ScalarTy, CommonAlignment,
                                         CurrentOrder, *Dist, Ptr0, SPtrInfo))
        return TreeEntry::StridedVectorize;
      // If the stores are not consecutive but the target supports masked stores
      // for the widened type, lower them as a masked store.
      FixedVectorType *StoreVecTy = nullptr;
      if (EnableMaskedStores &&
          isMaskedStoreCompress(VL, PointerOps, CurrentOrder, *TTI, *DL, *SE,
                                CommonAlignment, ReuseShuffleIndices,
                                StoreVecTy)) {
        SPtrInfo.Ty = StoreVecTy;
        return TreeEntry::ExpandVectorize;
      }
    }
 
    LLVM_DEBUG(dbgs() << "SLP: Non-consecutive store.\n");
    return TreeEntry::NeedToGather;
  }
  case Instruction::Call: {
    if (S.getMainOp()->getType()->isFloatingPointTy() &&
        TTI->isFPVectorizationPotentiallyUnsafe() && any_of(VL, [](Value *V) {
          auto *FPOp = dyn_cast<FPMathOperator>(V);
          return FPOp && !FPOp->isFast();
        }))
      return TreeEntry::NeedToGather;
    // Check if the calls are all to the same vectorizable intrinsic or
    // library function.
    CallInst *CI = cast<CallInst>(VL0);
    Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
 
    VFShape Shape = VFShape::get(
        CI->getFunctionType(),
        ElementCount::getFixed(static_cast<unsigned int>(VL.size())),
        false /*HasGlobalPred*/);
    Function *VecFunc = VFDatabase(*CI).getVectorizedFunction(Shape);
 
    if (!VecFunc && !isTriviallyVectorizable(ID)) {
      LLVM_DEBUG(dbgs() << "SLP: Non-vectorizable call.\n");
      return TreeEntry::NeedToGather;
    }
    if (isa<StructType>(CI->getType()) &&
        !allStructUsersAreExtractValueInsts(VL)) {
      LLVM_DEBUG(dbgs() << "SLP: Struct-returning calls have non-extractvalue "
                           "users.\n");
      return TreeEntry::NeedToGather;
    }
    Function *F = CI->getCalledFunction();
    unsigned NumArgs = CI->arg_size();
    SmallVector<Value *, 4> ScalarArgs(NumArgs, nullptr);
    for (unsigned J = 0; J != NumArgs; ++J)
      if (isVectorIntrinsicWithScalarOpAtArg(ID, J, TTI))
        ScalarArgs[J] = CI->getArgOperand(J);
    for (Value *V : VL) {
      // Skip copyables (idempotent stand-ins) and poisons.
      if (isa<PoisonValue>(V) || S.isCopyableElement(V))
        continue;
      CallInst *CI2 = dyn_cast<CallInst>(V);
      Intrinsic::ID ID2 = CI2 ? getVectorIntrinsicIDForCall(CI2, TLI)
                              : Intrinsic::not_intrinsic;
      Intrinsic::ID Equivalent = isEquivalentIntrinsicID(ID, ID2);
      if (!CI2 ||
          (CI2->getCalledFunction() != F &&
           isEquivalentIntrinsicID(Equivalent, Intrinsic::fmuladd) ==
               Intrinsic::not_intrinsic) ||
          (ID != ID2 && Equivalent == Intrinsic::not_intrinsic) ||
          (VecFunc &&
           VecFunc != VFDatabase(*CI2).getVectorizedFunction(Shape)) ||
          !CI->hasIdenticalOperandBundleSchema(*CI2)) {
        LLVM_DEBUG(dbgs() << "SLP: mismatched calls:" << *CI << "!=" << *V
                          << "\n");
        return TreeEntry::NeedToGather;
      }
      // Some intrinsics have scalar arguments and should be same in order for
      // them to be vectorized.
      for (unsigned J = 0; J != NumArgs; ++J) {
        if (isVectorIntrinsicWithScalarOpAtArg(ID, J, TTI)) {
          Value *A1J = CI2->getArgOperand(J);
          if (ScalarArgs[J] != A1J) {
            LLVM_DEBUG(dbgs()
                       << "SLP: mismatched arguments in call:" << *CI
                       << " argument " << ScalarArgs[J] << "!=" << A1J << "\n");
            return TreeEntry::NeedToGather;
          }
        }
      }
      // Verify that the bundle operands are identical between the two calls.
      if (CI->hasOperandBundles() &&
          !std::equal(CI->op_begin() + CI->getBundleOperandsStartIndex(),
                      CI->op_begin() + CI->getBundleOperandsEndIndex(),
                      CI2->op_begin() + CI2->getBundleOperandsStartIndex())) {
        LLVM_DEBUG(dbgs() << "SLP: mismatched bundle operands in calls:" << *CI
                          << "!=" << *V << '\n');
        return TreeEntry::NeedToGather;
      }
    }
    SmallVector<Type *> ArgTys =
        buildIntrinsicArgTypes(CI, ID, VL.size(), 0, TTI);
    auto *VecTy = getWidenedType(S.getMainOp()->getType(), VL.size());
    auto VecCallCosts =
        getVectorCallCosts(CI, VecTy, TTI, TLI, ArgTys, CostKind);
    if (!VecCallCosts.first.isValid() && !VecCallCosts.second.isValid())
      return TreeEntry::NeedToGather;
 
    return TreeEntry::Vectorize;
  }
  case Instruction::ShuffleVector: {
    if (!S.isAltShuffle()) {
      // REVEC can support non alternate shuffle.
      if (SLPReVec && getShufflevectorNumGroups(VL))
        return TreeEntry::Vectorize;
      // If this is not an alternate sequence of opcode like add-sub
      // then do not vectorize this instruction.
      LLVM_DEBUG(dbgs() << "SLP: ShuffleVector are not vectorized.\n");
      return TreeEntry::NeedToGather;
    }
 
    return TreeEntry::Vectorize;
  }
  default:
    LLVM_DEBUG(dbgs() << "SLP: Gathering unknown instruction.\n");
    return TreeEntry::NeedToGather;
  }
}
 
namespace {
/// Allows to correctly handle operands of the phi nodes based on the \p Main
/// PHINode order of incoming basic blocks/values.
class PHIHandler {
  DominatorTree &DT;
  PHINode *Main = nullptr;
  SmallVector<Value *> Phis;
  SmallVector<SmallVector<Value *>> Operands;
 
public:
  PHIHandler() = delete;
  PHIHandler(DominatorTree &DT, PHINode *Main, ArrayRef<Value *> Phis)
      : DT(DT), Main(Main), Phis(Phis),
        Operands(Main->getNumIncomingValues(),
                 SmallVector<Value *>(Phis.size(), nullptr)) {}
  void buildOperands() {
    constexpr unsigned FastLimit = 4;
    if (Main->getNumIncomingValues() <= FastLimit) {
      for (unsigned I : seq<unsigned>(0, Main->getNumIncomingValues())) {
        BasicBlock *InBB = Main->getIncomingBlock(I);
        if (!DT.isReachableFromEntry(InBB)) {
          Operands[I].assign(Phis.size(), PoisonValue::get(Main->getType()));
          continue;
        }
        // Prepare the operand vector.
        for (auto [Idx, V] : enumerate(Phis)) {
          auto *P = dyn_cast<PHINode>(V);
          if (!P) {
            assert(isa<PoisonValue>(V) &&
                   "Expected isa instruction or poison value.");
            Operands[I][Idx] = V;
            continue;
          }
          if (P->getIncomingBlock(I) == InBB)
            Operands[I][Idx] = P->getIncomingValue(I);
          else
            Operands[I][Idx] = P->getIncomingValueForBlock(InBB);
        }
      }
      return;
    }
    SmallMapVector<BasicBlock *, SmallVector<unsigned>, 4>
        Blocks;
    for (unsigned I : seq<unsigned>(Main->getNumIncomingValues())) {
      BasicBlock *InBB = Main->getIncomingBlock(I);
      if (!DT.isReachableFromEntry(InBB)) {
        Operands[I].assign(Phis.size(), PoisonValue::get(Main->getType()));
        continue;
      }
      Blocks.try_emplace(InBB).first->second.push_back(I);
    }
    for (auto [Idx, V] : enumerate(Phis)) {
      if (isa<PoisonValue>(V)) {
        for (unsigned I : seq<unsigned>(Main->getNumIncomingValues()))
          Operands[I][Idx] = V;
        continue;
      }
      auto *P = cast<PHINode>(V);
      for (unsigned I : seq<unsigned>(P->getNumIncomingValues())) {
        BasicBlock *InBB = P->getIncomingBlock(I);
        if (InBB == Main->getIncomingBlock(I)) {
          if (isa_and_nonnull<PoisonValue>(Operands[I][Idx]))
            continue;
          Operands[I][Idx] = P->getIncomingValue(I);
          continue;
        }
        auto *It = Blocks.find(InBB);
        if (It == Blocks.end())
          continue;
        Operands[It->second.front()][Idx] = P->getIncomingValue(I);
      }
    }
    for (const auto &P : Blocks) {
      ArrayRef<unsigned> IncomingValues = P.second;
      if (IncomingValues.size() <= 1)
        continue;
      unsigned BasicI = IncomingValues.consume_front();
      for (unsigned I : IncomingValues) {
        assert(all_of(enumerate(Operands[I]),
                      [&](const auto &Data) {
                        return !Data.value() ||
                               Data.value() == Operands[BasicI][Data.index()];
                      }) &&
               "Expected empty operands list.");
        Operands[I] = Operands[BasicI];
      }
    }
  }
  ArrayRef<Value *> getOperands(unsigned I) const { return Operands[I]; }
};
} // namespace
 
/// Returns main/alternate instructions for the given \p VL. Unlike
/// getSameOpcode supports non-compatible instructions for better SplitVectorize
/// node support.
/// \returns first main/alt instructions, if only poisons and instruction with
/// only 2 opcodes exists. Returns pair of nullptr otherwise.
static std::pair<Instruction *, Instruction *>
getMainAltOpsNoStateVL(ArrayRef<Value *> VL) {
  Instruction *MainOp = nullptr;
  Instruction *AltOp = nullptr;
  for (Value *V : VL) {
    if (isa<PoisonValue>(V))
      continue;
    auto *I = dyn_cast<Instruction>(V);
    if (!I)
      return {};
    if (!MainOp) {
      MainOp = I;
      continue;
    }
    if (MainOp->getOpcode() == I->getOpcode()) {
      if (I->getParent() != MainOp->getParent())
        return {};
      continue;
    }
    if (!AltOp) {
      AltOp = I;
      continue;
    }
    if (AltOp->getOpcode() == I->getOpcode()) {
      if (I->getParent() != AltOp->getParent())
        return {};
      continue;
    }
    return {};
  }
  if (!AltOp)
    return {};
  assert(MainOp && AltOp && MainOp->getOpcode() != AltOp->getOpcode() &&
         "Expected different main and alt instructions.");
  return std::make_pair(MainOp, AltOp);
}
 
/// Checks that every instruction appears once in the list and if not, packs
/// them, building \p ReuseShuffleIndices mask and mutating \p VL. The list of
/// unique scalars is extended by poison values to the whole register size.
///
/// \returns false if \p VL could not be uniquified, in which case \p VL is
/// unchanged and \p ReuseShuffleIndices is empty.
static bool tryToFindDuplicates(SmallVectorImpl<Value *> &VL,
                                SmallVectorImpl<int> &ReuseShuffleIndices,
                                const TargetTransformInfo &TTI,
                                const TargetLibraryInfo &TLI,
                                const InstructionsState &S,
                                const BoUpSLP::EdgeInfo &UserTreeIdx,
                                const BoUpSLP &R, bool BuildGatherOnly = true) {
  // TODO: Reordering of struct types is not supported.
  if (isa<StructType>(getValueType(VL.front(), SLPReVec))) {
    LLVM_DEBUG(dbgs() << "SLP: struct type in bundle.\n");
    ReuseShuffleIndices.clear();
    return true;
  }
  // Check that every instruction appears once in this bundle.
  SmallVector<Value *> UniqueValues;
  SmallDenseMap<Value *, unsigned, 16> UniquePositions(VL.size());
  for (Value *V : VL) {
    if (isConstant(V)) {
      // Constants are always considered distinct, even if the same constant
      // appears multiple times in VL.
      ReuseShuffleIndices.emplace_back(
          isa<PoisonValue>(V) ? PoisonMaskElem : UniqueValues.size());
      UniqueValues.emplace_back(V);
      continue;
    }
    auto Res = UniquePositions.try_emplace(V, UniqueValues.size());
    ReuseShuffleIndices.emplace_back(Res.first->second);
    if (Res.second)
      UniqueValues.emplace_back(V);
  }
 
  // Check if we need to schedule the scalars. If no, can keep original scalars
  // and avoid extra shuffles.
  bool RequireScheduling = S && S.getOpcode() != Instruction::PHI &&
                           !isVectorLikeInstWithConstOps(S.getMainOp()) &&
                           (S.areInstructionsWithCopyableElements() ||
                            !doesNotNeedToSchedule(UniqueValues));
  // Compute this flag BEFORE the tail-poison erase below - it must reflect
  // the state of the original VL (for the InsertsCost call), not the
  // potentially-shrunk UniqueValues.
  bool AreAllValuesNonConst = UniquePositions.size() == UniqueValues.size();
  // Drop tail poisons, if the values can be vectorized.
  if (RequireScheduling) {
    const auto EndIt =
        find_if_not(make_range(UniqueValues.rbegin(), UniqueValues.rend()),
                    IsaPred<PoisonValue>);
    assert(EndIt != UniqueValues.rend() && "Expected at least one non-poison.");
    UniqueValues.erase(EndIt.base(), UniqueValues.end());
  }
  unsigned NumUniqueScalarValues = UniqueValues.size();
  if (NumUniqueScalarValues == VL.size()) {
    ReuseShuffleIndices.clear();
    return true;
  }
 
  // For VL=4 with 3 unique values: keep originals. A <3 x T> vector is
  // always widened to <4 x T> on hardware, so the packing just adds an
  // extra expand shuffle. Does not apply to loads (a <3 x T> load is a
  // single memory access) or PHIs (benefit from compact packing in loops).
  constexpr unsigned SmallVecWidth = 4;
  constexpr unsigned SmallVecUniqueThreshold = 3;
  if (VL.size() == SmallVecWidth &&
      NumUniqueScalarValues == SmallVecUniqueThreshold && !BuildGatherOnly &&
      !(S && (S.getOpcode() == Instruction::Load ||
              S.getOpcode() == Instruction::PHI))) {
    // Keep originals with identity reuse - no packing, no extra shuffle.
    ReuseShuffleIndices.clear();
    return true;
  }
 
  // Checks if unique inserts + shuffle is more profitable than just inserts or
  // vectorized values.
  auto EstimatePackPlusShuffleVsInserts = [&]() {
    // Single instruction/argument insert - no shuffle.
    if (UniquePositions.size() == 1 &&
        (NumUniqueScalarValues == 1 ||
         all_of(UniqueValues, IsaPred<UndefValue, Instruction, Argument>)))
      return std::make_pair(false, false);
    // For large gathers with power-of-2 VL where packing would produce
    // non-power-of-2, reject if most scalars are constants - the packing
    // overhead (non-power-of-2 split + shuffles) outweighs the benefit.
    constexpr unsigned MinVLForConstGatherCheck = 4;
    if (BuildGatherOnly && VL.size() > MinVLForConstGatherCheck &&
        has_single_bit(static_cast<unsigned>(VL.size())) &&
        !has_single_bit(NumUniqueScalarValues) &&
        UniquePositions.size() * 2 < NumUniqueScalarValues)
      return std::make_pair(false, false);
    auto CheckLoads = [&](ArrayRef<Value *> Loads, bool IncludeGather) {
      assert(S && S.getOpcode() == Instruction::Load && "Expected load.");
      BoUpSLP::OrdersType Order;
      SmallVector<Value *> PointerOps;
      BoUpSLP::StridedPtrInfo SPtrInfo;
      BoUpSLP::LoadsState Res = R.canVectorizeLoads(Loads, S.getMainOp(), Order,
                                                    PointerOps, SPtrInfo);
      return (IncludeGather && Res == BoUpSLP::LoadsState::Gather) ||
             Res == BoUpSLP::LoadsState::ScatterVectorize ||
             Res == BoUpSLP::LoadsState::StridedVectorize ||
             Res == BoUpSLP::LoadsState::CompressVectorize ||
             Res == BoUpSLP::LoadsState::BlendedLoadVectorize;
    };
    // Operand of the root tree entry on the vectorize path: always pack the
    // scalars (PackProfitable=true). Choose between keeping the original VL
    // and packing the unique values:
    // - For loads, prefer the originals only when both the deduplicated and
    //   the full sequence can be vectorized non-Gather, or when the reuse
    //   mask is the identity (the shuffle is free).
    // - For everything else (including !S, where RequireScheduling is forced
    //   to false above), keep originals iff no scheduling is required.
    bool IsRootOperand =
        UserTreeIdx.UserTE && UserTreeIdx.UserTE->Idx == 0 && !BuildGatherOnly;
    if (IsRootOperand) {
      if (S && S.getOpcode() == Instruction::Load) {
        bool UseOrig = (CheckLoads(UniqueValues, /*IncludeGather=*/true) &&
                        CheckLoads(VL, /*IncludeGather=*/false)) ||
                       ShuffleVectorInst::isIdentityMask(
                           ReuseShuffleIndices, ReuseShuffleIndices.size());
        return std::make_pair(true, UseOrig);
      }
      return std::make_pair(true, !RequireScheduling);
    }
    APInt DemandedElts = APInt::getZero(VL.size());
    for (auto [Idx, Val] : enumerate(ReuseShuffleIndices))
      if (Val != PoisonMaskElem && UniquePositions.contains(UniqueValues[Val]))
        DemandedElts.setBit(Idx);
    Type *ScalarTy = ::getValueType(UniqueValues.front(), SLPReVec);
    auto *VecTy = cast<VectorType>(getWidenedType(ScalarTy, VL.size()));
    auto *UniquesVecTy =
        cast<VectorType>(getWidenedType(ScalarTy, NumUniqueScalarValues));
    const unsigned NumParts = ::getNumberOfParts(TTI, VecTy, ScalarTy);
    const unsigned UniquesNumParts =
        ::getNumberOfParts(TTI, UniquesVecTy, ScalarTy);
    // No need to schedule scalars and only single register used? Use original
    // scalars, do not pack.
    if (!RequireScheduling) {
      if (VL.size() / NumUniqueScalarValues == 1 &&
          (NumParts <= 1 || UniquesNumParts >= NumParts))
        return std::make_pair(true, true);
      // For PHI operands, prefer packing with reuse shuffle - the PHI
      // carries the vector through the loop cheaply.
      if (S && S.getOpcode() == Instruction::PHI && NumUniqueScalarValues > 1 &&
          UniquesNumParts <= NumParts)
        return std::make_pair(true, false);
    }
    const TTI::TargetCostKind CostKind = R.getCostKind();
    InstructionCost ReusesCost = getShuffleCost(
        TTI, TTI::SK_PermuteSingleSrc, VecTy, CostKind,
        NumUniqueScalarValues > VL.size() / 2 ? ArrayRef<int>()
                                              : ArrayRef(ReuseShuffleIndices),
        /*Index=*/0, UniquesVecTy);
    // For vectorizable (non-gather) nodes with low duplication, prefer keeping
    // the original values over packing uniques + reshuffling:
    // - A single duplicate (non-load) adds negligible overhead.
    // - When most values are already unique (>50%), or exactly half are unique
    //   for some ops (GEPs, non-alt-shuffle casts), the reshuffle cost may
    //   exceed the savings from a smaller packed vector - check against a
    //   per-register-part threshold (stricter for wider vectors).
    if (S && !BuildGatherOnly) {
      bool HasOneDup = S.getOpcode() != Instruction::Load &&
                       NumUniqueScalarValues + 1 == VL.size();
      bool MostlyUnique = NumUniqueScalarValues * 2 > VL.size();
      bool IsHalfUniqueValues =
          NumUniqueScalarValues * 2 == VL.size() &&
          (S.getOpcode() == Instruction::GetElementPtr ||
           (isa<CastInst>(S.getMainOp()) && !S.isAltShuffle()));
      InstructionCost CostThreshold =
          NumParts * (VL.size() > SmallVecWidth ? 1 : 2);
      if (HasOneDup ||
          ((MostlyUnique || IsHalfUniqueValues) && ReusesCost > CostThreshold))
        return std::make_pair(true, true);
    }
    // For loads, check if either the deduplicated or the full (with
    // duplicates) set can be scatter/compress-vectorized. Prefer the unique
    // loads (pack + reshuffle) when possible, otherwise use the originals.
    if (S && S.getOpcode() == Instruction::Load) {
      bool UniquesVectorized =
          CheckLoads(UniqueValues, /*IncludeGather=*/false);
      if (UniquesVectorized || CheckLoads(VL, /*IncludeGather=*/false))
        return std::make_pair(true, !UniquesVectorized);
    }
    bool CanSkipBVCost =
        (!BuildGatherOnly && !RequireScheduling) || R.hasSameNode(S, VL);
    InstructionCost InsertsCost =
        CanSkipBVCost
            ? InstructionCost(TTI::TCC_Free)
            : getScalarizationOverhead(TTI, SLPReVec, ScalarTy, VecTy,
                                       DemandedElts,
                                       /*Insert=*/true, /*Extract=*/false,
                                       CostKind, AreAllValuesNonConst, VL);
    APInt UniquesDemandedElts = APInt::getAllOnes(NumUniqueScalarValues);
    for (const auto [Idx, V] : enumerate(UniqueValues))
      if (isConstant(V))
        UniquesDemandedElts.clearBit(Idx);
    InstructionCost UniquesCost =
        CanSkipBVCost
            ? InstructionCost(TTI::TCC_Free)
            : getScalarizationOverhead(TTI, SLPReVec, ScalarTy, UniquesVecTy,
                                       UniquesDemandedElts, /*Insert=*/true,
                                       /*Extract=*/false, CostKind,
                                       AreAllValuesNonConst, UniqueValues);
    UniquesCost += ReusesCost;
    if (UniquesCost <= InsertsCost) {
      // Packing to a vector that occupies the same register width as the
      // original only adds the reshuffle cost; keep the originals. Loads may
      // instead benefit from a wider contiguous load.
      if (S && S.getOpcode() != Instruction::Load) {
        unsigned EltBits =
            S.getMainOp()->getDataLayout().getTypeSizeInBits(ScalarTy);
        unsigned MinVF = R.getMinVF(EltBits);
        auto RegWidth = [&](unsigned N) {
          return std::max(
              getFullVectorNumberOfElements(TTI, ScalarTy, N, SLPReVec), MinVF);
        };
        // Keeping the originals just moves the reshuffle to the operand
        // columns with duplicates; keep them only if at most one column has
        // duplicates, so the total number of reshuffles does not grow.
        auto HasExtraReshuffle = [&]() {
          if (BuildGatherOnly)
            return false;
          unsigned NumDupColumns = 0;
          for (unsigned OpIdx :
               seq<unsigned>(S.getMainOp()->getNumOperands())) {
            SmallPtrSet<const Value *, 16> UniqueOps;
            for (Value *V : VL) {
              if (isa<PoisonValue>(V))
                continue;
              Value *Op;
              if (S.isCopyableElement(V)) {
                if (OpIdx != 0)
                  continue;
                Op = V;
              } else {
                auto *I = dyn_cast<Instruction>(V);
                if (!I || OpIdx >= I->getNumOperands())
                  continue;
                Op = I->getOperand(OpIdx);
              }
              if (!isConstant(Op) && !UniqueOps.insert(Op).second) {
                ++NumDupColumns;
                break;
              }
            }
            if (NumDupColumns > 1)
              return true;
          }
          return false;
        };
        return std::make_pair(true, RegWidth(NumUniqueScalarValues) >=
                                            RegWidth(VL.size()) &&
                                        !HasExtraReshuffle());
      }
      return std::make_pair(true, false);
    }
    InstructionCost CostDiff = UniquesCost - InsertsCost;
    if (CostDiff < TTI::TCC_Expensive ||
        (R.getTreeSize() == 0 && R.isReductionTree() &&
         CostDiff == TTI::TCC_Expensive))
      return std::make_pair(S && (!S.isAltShuffle() || !BuildGatherOnly),
                            false);
    // Otherwise, use original values, if values do not require scheduling and
    // pass still try to vectorize them.
    bool KeepOriginal = !BuildGatherOnly && !RequireScheduling;
    return std::make_pair(KeepOriginal, KeepOriginal);
  };
 
  const auto [PackProfitable, UseOriginal] = EstimatePackPlusShuffleVsInserts();
 
  if (PackProfitable) {
    if (UseOriginal) {
      // Prefer original scalars - avoid shuffling.
      ReuseShuffleIndices.clear();
    } else {
      // Better to use uniques + reshuffle.
      LLVM_DEBUG(dbgs() << "SLP: Shuffle for reused scalars.\n");
      VL = std::move(UniqueValues);
    }
    return true;
  }
 
  // Buildvector/gather of the original scalars.
  LLVM_DEBUG(dbgs() << "SLP: Scalar used twice in bundle.\n");
  ReuseShuffleIndices.clear();
  return false;
}
 
bool BoUpSLP::canBuildSplitNode(ArrayRef<Value *> VL,
                                const InstructionsState &LocalState,
                                SmallVectorImpl<Value *> &Op1,
                                SmallVectorImpl<Value *> &Op2,
                                OrdersType &ReorderIndices) const {
  constexpr unsigned SmallNodeSize = 4;
  if (VL.size() <= SmallNodeSize || TTI->preferAlternateOpcodeVectorization() ||
      !SplitAlternateInstructions)
    return false;
 
  // Check if this is a duplicate of another split entry.
  LLVM_DEBUG(dbgs() << "SLP: \tChecking bundle: " << *LocalState.getMainOp()
                    << ".\n");
  for (TreeEntry *E : getSplitTreeEntries(LocalState.getMainOp())) {
    if (E->isSame(VL)) {
      LLVM_DEBUG(dbgs() << "SLP: Perfect diamond merge at "
                        << *LocalState.getMainOp() << ".\n");
      return false;
    }
    SmallPtrSet<Value *, 8> Values(llvm::from_range, E->Scalars);
    if (all_of(VL, [&](Value *V) {
          return isa<PoisonValue>(V) || Values.contains(V);
        })) {
      LLVM_DEBUG(dbgs() << "SLP: Gathering due to full overlap.\n");
      return false;
    }
  }
 
  ReorderIndices.assign(VL.size(), VL.size());
  SmallBitVector Op1Indices(VL.size());
  for (auto [Idx, V] : enumerate(VL)) {
    auto *I = dyn_cast<Instruction>(V);
    if (!I) {
      Op1.push_back(V);
      Op1Indices.set(Idx);
      continue;
    }
    if ((LocalState.getAltOpcode() != LocalState.getOpcode() &&
         isMainInstruction(I, LocalState.getMainOp(), LocalState.getAltOp(),
                           *TLI)) ||
        (LocalState.getAltOpcode() == LocalState.getOpcode() &&
         !isAlternateInstruction(I, LocalState.getMainOp(),
                                 LocalState.getAltOp(), *TLI))) {
      Op1.push_back(V);
      Op1Indices.set(Idx);
      continue;
    }
    Op2.push_back(V);
  }
  Type *ScalarTy = getValueType(VL.front(), SLPReVec);
  auto *VecTy = cast<VectorType>(getWidenedType(ScalarTy, VL.size()));
  unsigned Opcode0 = LocalState.getOpcode();
  unsigned Opcode1 = LocalState.getAltOpcode();
  SmallBitVector OpcodeMask(getAltInstrMask(VL, ScalarTy, Opcode0, Opcode1));
  // Enable split node, only if all nodes do not form legal alternate
  // instruction (like X86 addsub).
  SmallPtrSet<Value *, 4> UOp1(llvm::from_range, Op1);
  SmallPtrSet<Value *, 4> UOp2(llvm::from_range, Op2);
  if (UOp1.size() <= 1 || UOp2.size() <= 1 ||
      TTI->isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask))
    return false;
  unsigned Op1Cnt = 0, Op2Cnt = Op1.size();
  for (unsigned Idx : seq<unsigned>(VL.size())) {
    if (Op1Indices.test(Idx)) {
      ReorderIndices[Op1Cnt] = Idx;
      ++Op1Cnt;
    } else {
      ReorderIndices[Op2Cnt] = Idx;
      ++Op2Cnt;
    }
  }
  if (isIdentityOrder(ReorderIndices))
    ReorderIndices.clear();
  // When VL fills a power-of-2 register but the split halves do not, the
  // reorder shuffle makes the split unprofitable - reject.
  else if (hasFullVectorsOrPowerOf2(*TTI, Op1.front()->getType(), VL.size(),
                                    SLPReVec) &&
           (!hasFullVectorsOrPowerOf2(*TTI, Op1.front()->getType(), Op1.size(),
                                      SLPReVec) ||
            !hasFullVectorsOrPowerOf2(*TTI, Op2.front()->getType(), Op2.size(),
                                      SLPReVec)))
    return false;
  SmallVector<int> Mask;
  if (!ReorderIndices.empty())
    inversePermutation(ReorderIndices, Mask);
  unsigned NumParts = TTI->getNumberOfParts(VecTy);
  auto *Op1VecTy = cast<VectorType>(getWidenedType(ScalarTy, Op1.size()));
  auto *Op2VecTy = cast<VectorType>(getWidenedType(ScalarTy, Op2.size()));
  // Check non-profitable single register ops, which better to be represented
  // as alternate ops.
  if (NumParts >= VL.size())
    return false;
  InstructionCost InsertCost = getShuffleCost(
      *TTI, TTI::SK_InsertSubvector, VecTy, CostKind, {}, Op1.size(), Op2VecTy);
  auto *SubVecTy = cast<VectorType>(
      getWidenedType(ScalarTy, std::max(Op1.size(), Op2.size())));
  InstructionCost NewShuffleCost =
      getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc, SubVecTy, CostKind, Mask);
  if (!LocalState.isCmpOp() && NumParts <= 1 &&
      (Mask.empty() || InsertCost >= NewShuffleCost))
    return false;
  if ((LocalState.getMainOp()->isBinaryOp() &&
       LocalState.getAltOp()->isBinaryOp() &&
       (LocalState.isShiftOp() || LocalState.isBitwiseLogicOp() ||
        LocalState.isAddSubLikeOp() || LocalState.isMulDivLikeOp())) ||
      (LocalState.getMainOp()->isCast() && LocalState.getAltOp()->isCast()) ||
      (LocalState.getMainOp()->isUnaryOp() &&
       LocalState.getAltOp()->isUnaryOp())) {
    InstructionCost OriginalVecOpsCost =
        TTI->getArithmeticInstrCost(Opcode0, VecTy, CostKind) +
        TTI->getArithmeticInstrCost(Opcode1, VecTy, CostKind);
    SmallVector<int> OriginalMask(VL.size(), PoisonMaskElem);
    for (unsigned Idx : seq<unsigned>(VL.size())) {
      if (isa<PoisonValue>(VL[Idx]))
        continue;
      OriginalMask[Idx] = Idx + (Op1Indices.test(Idx) ? 0 : VL.size());
    }
    InstructionCost OriginalCost =
        OriginalVecOpsCost + getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc, VecTy,
                                            CostKind, OriginalMask);
    InstructionCost NewVecOpsCost =
        TTI->getArithmeticInstrCost(Opcode0, Op1VecTy, CostKind) +
        TTI->getArithmeticInstrCost(Opcode1, Op2VecTy, CostKind);
    InstructionCost NewCost =
        NewVecOpsCost + InsertCost +
        (!VectorizableTree.empty() && getRootNode().hasState() &&
                 getRootNode().getOpcode() == Instruction::Store
             ? NewShuffleCost
             : 0);
    // If not profitable to split - exit.
    if (NewCost >= OriginalCost)
      return false;
  }
  return true;
}
 
namespace {
/// Class accepts incoming list of values, checks if it is able to model
/// "copyable" values as compatible operations, and generates the list of values
/// for scheduling and list of operands doe the new nodes.
class InstructionsCompatibilityAnalysis {
  DominatorTree &DT;
  const DataLayout &DL;
  const TargetTransformInfo &TTI;
  const TargetLibraryInfo &TLI;
  unsigned MainOpcode = 0;
  Instruction *MainOp = nullptr;
  /// Whether every copyable in the current value list is an absorbable
  /// single-use fmul/fadd. Computed once per buildInstructionsState call.
  bool AbsorbCopyableFMulOrFAdds = false;
 
  /// Checks if the opcode is supported as the main opcode for copyable
  /// elements.
  static bool isSupportedOpcode(const unsigned Opcode) {
    return Opcode == Instruction::Add || Opcode == Instruction::Sub ||
           Opcode == Instruction::Mul || Opcode == Instruction::AShr ||
           Opcode == Instruction::LShr || Opcode == Instruction::Shl ||
           Opcode == Instruction::SDiv || Opcode == Instruction::UDiv ||
           Opcode == Instruction::And || Opcode == Instruction::Or ||
           Opcode == Instruction::Xor || Opcode == Instruction::FAdd ||
           Opcode == Instruction::FSub || Opcode == Instruction::FMul ||
           Opcode == Instruction::FDiv;
  }
 
  /// Checks if \p I can be the main op for copyable analysis: a supported
  /// binary operator, fmuladd, or an integer min/max intrinsic, the only
  /// call with a well-defined idempotent value (FP min/max lacks one because of
  /// NaNs). An extractelement with constant index from a fixed vector is also
  /// supported: the matching lanes reuse the source vector and the copyable
  /// lanes are inserted into it.
  static bool isSupportedMainOp(Instruction *I) {
    return isSupportedOpcode(I->getOpcode()) || isa<MinMaxIntrinsic>(I) ||
           RecurrenceDescriptor::isFMulAddIntrinsic(I) ||
           (isa<ExtractElementInst>(I) && isVectorLikeInstWithConstOps(I) &&
            isa<FixedVectorType>(I->getOperand(0)->getType()));
  }
 
  /// Identifies the best candidate value, which represents main opcode
  /// operation.
  /// Currently the best candidate is the Add instruction with the parent
  /// block with the highest DFS incoming number (block, that dominates other).
  void findAndSetMainInstruction(ArrayRef<Value *> VL, const BoUpSLP &R) {
    BasicBlock *Parent = nullptr;
    // Checks if the instruction has supported opcode.
    auto IsSupportedInstruction = [&](Instruction *I, bool AnyUndef) {
      if (AnyUndef && (I->isIntDivRem() || I->isFPDivRem() || isa<CallInst>(I)))
        return false;
      return I && isSupportedMainOp(I) &&
             (!doesNotNeedToBeScheduled(I) || !R.isVectorized(I));
    };
    // Exclude operands instructions immediately to improve compile time, it
    // will be unable to schedule anyway.
    SmallDenseSet<Value *, 8> Operands;
    SmallMapVector<unsigned, SmallVector<Instruction *>, 4> Candidates;
    bool AnyUndef = false;
    for (Value *V : VL) {
      auto *I = dyn_cast<Instruction>(V);
      if (!I) {
        AnyUndef |= isa<UndefValue>(V);
        continue;
      }
      if (!DT.isReachableFromEntry(I->getParent()))
        continue;
      if (Candidates.empty()) {
        Candidates.try_emplace(I->getOpcode()).first->second.push_back(I);
        Parent = I->getParent();
        Operands.insert(I->op_begin(), I->op_end());
        continue;
      }
      if (Parent == I->getParent()) {
        Candidates.try_emplace(I->getOpcode()).first->second.push_back(I);
        Operands.insert(I->op_begin(), I->op_end());
        continue;
      }
      auto *NodeA = DT.getNode(Parent);
      auto *NodeB = DT.getNode(I->getParent());
      assert(NodeA && "Should only process reachable instructions");
      assert(NodeB && "Should only process reachable instructions");
      assert((NodeA == NodeB) ==
                 (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) &&
             "Different nodes should have different DFS numbers");
      if (NodeA->getDFSNumIn() < NodeB->getDFSNumIn()) {
        Candidates.clear();
        Candidates.try_emplace(I->getOpcode()).first->second.push_back(I);
        Parent = I->getParent();
        Operands.clear();
        Operands.insert(I->op_begin(), I->op_end());
      }
    }
    unsigned BestOpcodeNum = 0;
    MainOp = nullptr;
    bool UsedOutside = false;
    for (const auto &P : Candidates) {
      bool PUsedOutside = all_of(P.second, isUsedOutsideBlock);
      if (UsedOutside && !PUsedOutside)
        continue;
      if (!UsedOutside && PUsedOutside)
        BestOpcodeNum = 0;
      if (P.second.size() < BestOpcodeNum)
        continue;
      // If have inner dependencies - skip.
      if (!PUsedOutside && any_of(P.second, [&](Instruction *I) {
            return Operands.contains(I);
          }))
        continue;
      // On a tie, keep the outer binary op as MainOp rather than replacing it
      // with an inner op that appears as its direct operand. For example, in
      // (2.0f * A) + B the fadd and fmul each appear once in VL; without this
      // check the fmul could win and prevent vectorization of the fadd pair.
      if (P.second.size() == BestOpcodeNum) {
        auto *I = P.second.front();
        if (auto *MainBO = dyn_cast<BinaryOperator>(MainOp)) {
          auto *MainBOOp0 = dyn_cast<Instruction>(MainBO->getOperand(0));
          auto *MainBOOp1 = dyn_cast<Instruction>(MainBO->getOperand(1));
          if (MainBOOp0 && MainBOOp0->getOpcode() == I->getOpcode() &&
              MainBOOp0->getParent() == I->getParent())
            continue;
          if (MainBOOp1 && MainBOOp1->getOpcode() == I->getOpcode() &&
              MainBOOp1->getParent() == I->getParent())
            continue;
        }
        // Keep fmuladd over fmul/fadd on a tie only when every copyable is
        // an absorbed fmul/fadd.
        if (RecurrenceDescriptor::isFMulAddIntrinsic(MainOp) &&
            (I->getOpcode() == Instruction::FMul ||
             I->getOpcode() == Instruction::FAdd) &&
            AbsorbCopyableFMulOrFAdds)
          continue;
        // Same check when fmuladd replaces fmul/fadd on a tie.
        if ((MainOp->getOpcode() == Instruction::FMul ||
             MainOp->getOpcode() == Instruction::FAdd) &&
            RecurrenceDescriptor::isFMulAddIntrinsic(I) &&
            !AbsorbCopyableFMulOrFAdds)
          continue;
      }
      UsedOutside = PUsedOutside;
      for (Instruction *I : P.second) {
        if (IsSupportedInstruction(I, AnyUndef)) {
          MainOp = I;
          BestOpcodeNum = P.second.size();
          break;
        }
      }
    }
    if (MainOp) {
      // Do not match, if any copyable is a terminator from the same block as
      // the main operation.
      if (any_of(VL, [&](Value *V) {
            auto *I = dyn_cast<Instruction>(V);
            return I && I->getParent() == MainOp->getParent() &&
                   I->isTerminator();
          })) {
        MainOp = nullptr;
        return;
      }
      MainOpcode = MainOp->getOpcode();
    }
  }
 
  /// Returns the idempotent value for MainOp. For Add, returns 0. For Or, it
  /// should choose between false and the operand itself, since V or V == V.
  /// For the min/max intrinsics, returns the corresponding limit value (e.g. 0
  /// for umax, since umax(V, 0) == V).
  Value *selectBestIdempotentValue() const {
    assert(isSupportedMainOp(MainOp) && "Unsupported opcode");
    return ConstantExpr::getIdentity(MainOp, MainOp->getType(),
                                     !MainOp->isCommutative());
  }
 
  /// Returns the value and operands for the \p V, considering if it is original
  /// instruction and its actual operands should be returned, or it is a
  /// copyable element and its should be represented as idempotent instruction.
  SmallVector<Value *> getOperands(const InstructionsState &S, Value *V) const {
    if (isa<PoisonValue>(V))
      return {V, V};
    if (!S.isCopyableElement(V))
      return convertTo(cast<Instruction>(V), S).second;
    if (RecurrenceDescriptor::isFMulAddIntrinsic(MainOp)) {
      Type *Ty = MainOp->getType();
      if (S.hasAbsorbedCopyableFMulOrFAdd() &&
          isAbsorbableCopyableFMulOrFAdd(S, V)) {
        auto *I = cast<Instruction>(V);
        // fmuladd(a, b, -0.0) == fmul a, b.
        if (I->getOpcode() == Instruction::FMul)
          return {I->getOperand(0), I->getOperand(1),
                  ConstantFP::getNegativeZero(Ty)};
        // fmuladd(1.0, a, b) == fadd a, b.
        return {ConstantFP::get(Ty, 1.0), I->getOperand(0), I->getOperand(1)};
      }
      // fmuladd(V, 1.0, -0.0) == V.
      if (S.getCopyableOpIdx() == 0)
        return {V, ConstantFP::get(Ty, 1.0), ConstantFP::getNegativeZero(Ty)};
      // fmuladd(0.0, -0.0, V) == V.
      return {ConstantFP::getZero(Ty), ConstantFP::getNegativeZero(Ty), V};
    }
    assert(isSupportedMainOp(MainOp) && "Unsupported opcode");
    return {V, selectBestIdempotentValue()};
  }
 
  /// Builds operands for the original instructions.
  void
  buildOriginalOperands(const InstructionsState &S, ArrayRef<Value *> VL,
                        SmallVectorImpl<BoUpSLP::ValueList> &Operands) const {
 
    unsigned ShuffleOrOp =
        S.isAltShuffle() ? (unsigned)Instruction::ShuffleVector : S.getOpcode();
    Instruction *VL0 = S.getMainOp();
 
    switch (ShuffleOrOp) {
    case Instruction::PHI: {
      auto *PH = cast<PHINode>(VL0);
 
      // Keeps the reordered operands to avoid code duplication.
      PHIHandler Handler(DT, PH, VL);
      Handler.buildOperands();
      Operands.assign(PH->getNumOperands(), {});
      for (unsigned I : seq<unsigned>(PH->getNumOperands()))
        Operands[I].assign(Handler.getOperands(I).begin(),
                           Handler.getOperands(I).end());
      return;
    }
    case Instruction::ExtractValue: {
      SmallVector<unsigned> Indices;
      SmallVector<Value *> Calls;
      if (checkEVsForVecCalls(VL, S, TLI, Indices, Calls)) {
        Operands.assign(1, {});
        Operands[0].swap(Calls);
        return;
      }
      [[fallthrough]];
    }
    case Instruction::ExtractElement:
      // This is a special case, as it does not gather, but at the same time
      // we are not extending buildTree_rec() towards the operands.
      Operands.assign(1, {VL.size(), VL0->getOperand(0)});
      return;
    case Instruction::InsertValue:
    case Instruction::InsertElement:
      Operands.assign(2, {VL.size(), nullptr});
      for (auto [Idx, V] : enumerate(VL)) {
        auto *IE = cast<Instruction>(V);
        for (auto [OpIdx, Ops] : enumerate(Operands))
          Ops[Idx] = IE->getOperand(OpIdx);
      }
      return;
    case Instruction::Load:
      Operands.assign(
          1, {VL.size(),
              PoisonValue::get(cast<LoadInst>(VL0)->getPointerOperandType())});
      for (auto [V, Op] : zip(VL, Operands.back())) {
        auto *LI = dyn_cast<LoadInst>(V);
        if (!LI)
          continue;
        Op = LI->getPointerOperand();
      }
      return;
    case Instruction::ZExt:
    case Instruction::SExt:
    case Instruction::FPToUI:
    case Instruction::FPToSI:
    case Instruction::FPExt:
    case Instruction::PtrToInt:
    case Instruction::PtrToAddr:
    case Instruction::IntToPtr:
    case Instruction::SIToFP:
    case Instruction::UIToFP:
    case Instruction::Trunc:
    case Instruction::FPTrunc:
    case Instruction::BitCast:
    case Instruction::ICmp:
    case Instruction::FCmp:
    case Instruction::FNeg:
    case Instruction::Add:
    case Instruction::FAdd:
    case Instruction::Sub:
    case Instruction::FSub:
    case Instruction::Mul:
    case Instruction::FMul:
    case Instruction::UDiv:
    case Instruction::SDiv:
    case Instruction::FDiv:
    case Instruction::URem:
    case Instruction::SRem:
    case Instruction::FRem:
    case Instruction::Shl:
    case Instruction::LShr:
    case Instruction::AShr:
    case Instruction::And:
    case Instruction::Or:
    case Instruction::Xor:
    case Instruction::Freeze:
    case Instruction::Store:
    case Instruction::ShuffleVector:
      Operands.assign(VL0->getNumOperands(), {VL.size(), nullptr});
      for (auto [Idx, V] : enumerate(VL)) {
        auto *I = dyn_cast<Instruction>(V);
        if (!I) {
          for (auto [OpIdx, Ops] : enumerate(Operands))
            Ops[Idx] = PoisonValue::get(VL0->getOperand(OpIdx)->getType());
          continue;
        }
        auto [Op, ConvertedOps] = convertTo(I, S);
        for (auto [OpIdx, Ops] : enumerate(Operands))
          Ops[Idx] = ConvertedOps[OpIdx];
      }
      return;
    case Instruction::Select:
      Operands.assign(VL0->getNumOperands(), {VL.size(), nullptr});
      for (auto [Idx, V] : enumerate(VL)) {
        auto *I = dyn_cast<Instruction>(V);
        if (!I) {
          for (auto [OpIdx, Ops] : enumerate(Operands))
            Ops[Idx] = PoisonValue::get(VL0->getOperand(OpIdx)->getType());
          continue;
        }
        if (isa<ZExtInst>(I)) {
          // Special case for select + zext i1 to avoid explosion of different
          // types. We want to keep the condition as i1 to be able to match
          // different selects together and reuse the vectorized condition
          // rather than trying to gather it.
          Operands[0][Idx] = I->getOperand(0);
          Operands[1][Idx] = ConstantInt::get(I->getType(), 1);
          Operands[2][Idx] = ConstantInt::getNullValue(I->getType());
          continue;
        }
        auto [Op, ConvertedOps] = convertTo(I, S);
        for (auto [OpIdx, Ops] : enumerate(Operands))
          Ops[Idx] = ConvertedOps[OpIdx];
      }
      return;
    case Instruction::GetElementPtr: {
      Operands.assign(2, {VL.size(), nullptr});
      // Need to cast all indices to the same type before vectorization to
      // avoid crash.
      // Required to be able to find correct matches between different gather
      // nodes and reuse the vectorized values rather than trying to gather them
      // again.
      const unsigned IndexIdx = 1;
      Type *VL0Ty = VL0->getOperand(IndexIdx)->getType();
      Type *Ty =
          all_of(VL,
                 [&](Value *V) {
                   auto *GEP = dyn_cast<GetElementPtrInst>(V);
                   return !GEP || VL0Ty == GEP->getOperand(IndexIdx)->getType();
                 })
              ? VL0Ty
              : DL.getIndexType(cast<GetElementPtrInst>(VL0)
                                    ->getPointerOperandType()
                                    ->getScalarType());
      for (auto [Idx, V] : enumerate(VL)) {
        auto *GEP = dyn_cast<GetElementPtrInst>(V);
        if (!GEP) {
          Operands[0][Idx] = V;
          Operands[1][Idx] = ConstantInt::getNullValue(Ty);
          continue;
        }
        Operands[0][Idx] = GEP->getPointerOperand();
        auto *Op = GEP->getOperand(IndexIdx);
        auto *CI = dyn_cast<ConstantInt>(Op);
        Operands[1][Idx] = CI ? ConstantFoldIntegerCast(
                                    CI, Ty, CI->getValue().isSignBitSet(), DL)
                              : Op;
      }
      return;
    }
    case Instruction::Call: {
      auto *CI = cast<CallInst>(VL0);
      Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, &TLI);
      for (unsigned Idx : seq<unsigned>(CI->arg_size())) {
        if (isVectorIntrinsicWithScalarOpAtArg(ID, Idx, &TTI))
          continue;
        auto &Ops = Operands.emplace_back();
        for (Value *V : VL) {
          auto *I = dyn_cast<Instruction>(V);
          Ops.push_back(I ? I->getOperand(Idx)
                          : PoisonValue::get(VL0->getOperand(Idx)->getType()));
        }
      }
      return;
    }
    default:
      break;
    }
    llvm_unreachable("Unexpected vectorization of the instructions.");
  }
 
  /// Check if the specified \p VL list of values is better to represent as
  /// uniform with copyables, as modeled via \p CopyableS, or as alternate (or
  /// uniform with compatible ops), modeled via \p S.
  /// Performs the analysis of the operands, choosing the preferred main
  /// instruction and checking the matching of the operands for the main
  /// instruction and copyable elements.
  bool isCopyablePreferable(ArrayRef<Value *> VL, const BoUpSLP &R,
                            const InstructionsState &S,
                            const InstructionsState &CopyableS) {
    // If all elements are vectorized already - keep as is.
    if (all_of(VL, [&](Value *V) {
          return isa<PoisonValue>(V) || R.isVectorized(V);
        }))
      return false;
    Instruction *SMain = S.getMainOp();
    Instruction *SAlt = S.isAltShuffle() ? S.getAltOp() : nullptr;
    const bool IsCommutative = isCommutative(SMain);
    const bool IsAltCommutative =
        S.isAltShuffle() ? isCommutative(SAlt) : false;
    const bool IsMainCommutative = isCommutative(MainOp);
    SmallVector<BoUpSLP::ValueList> Ops;
    buildOriginalOperands(S, SMain, Ops);
    // Support only binary operations for now.
    if (Ops.size() != 2)
      return false;
    // Try to find better candidate for S main instruction, which operands have
    // better matching.
    auto CheckOperands = [](Value *Op, Value *SMainOp) {
      auto *OpI = dyn_cast<BinaryOperator>(Op);
      if (!OpI)
        return false;
      auto *SMainOpI = dyn_cast<BinaryOperator>(SMainOp);
      if (!SMainOpI)
        return true;
      return any_of(OpI->operands(), [&](Value *V) {
        auto *I = dyn_cast<Instruction>(V);
        return I && I->getOpcode() == SMainOpI->getOpcode();
      });
    };
    SmallPtrSet<Value *, 8> Operands;
    for (Value *V : VL) {
      auto *I = dyn_cast<Instruction>(V);
      if (!I || I == SMain)
        continue;
      Instruction *MatchingOp = S.getMatchingMainOpOrAltOp(I);
      if (MatchingOp != SMain)
        continue;
      SmallVector<BoUpSLP::ValueList> VOps;
      buildOriginalOperands(S, I, VOps);
      Operands.insert(I->op_begin(), I->op_end());
      assert(VOps.size() == 2 && Ops.size() == 2 &&
             "Expected binary operations only.");
      if (CheckOperands(VOps[0][0], Ops[0][0]) ||
          CheckOperands(VOps[1][0], Ops[1][0]) ||
          (IsCommutative && (CheckOperands(VOps[0][0], Ops[1][0]) ||
                             CheckOperands(VOps[1][0], Ops[0][0])))) {
        SMain = I;
        Ops.swap(VOps);
        break;
      }
    }
    SmallVector<BoUpSLP::ValueList> MainOps;
    buildOriginalOperands(S, MainOp, MainOps);
 
    auto BuildFirstOperandCandidates =
        [&](SmallVectorImpl<std::pair<Value *, Value *>> &Candidates,
            ArrayRef<BoUpSLP::ValueList> Ops, Value *Op0, Value *Op1,
            bool IsCommutative) {
          Candidates.emplace_back(Ops[0][0], Op0);
          if (IsCommutative)
            Candidates.emplace_back(Ops[0][0], Op1);
        };
 
    auto BuildSecondOperandCandidates =
        [&](SmallVectorImpl<std::pair<Value *, Value *>> &Candidates,
            ArrayRef<BoUpSLP::ValueList> Ops, int PrevBestIdx, Value *Op0,
            Value *Op1, bool IsCommutative) {
          if (PrevBestIdx != 1)
            Candidates.emplace_back(Ops[1][0], Op1);
          if (PrevBestIdx != 0 && IsCommutative)
            Candidates.emplace_back(Ops[1][0], Op0);
        };
 
    auto FindBestCandidate =
        [&](ArrayRef<std::pair<Value *, Value *>> Candidates, bool &IsConst,
            int &Score) {
          auto Res = R.findBestRootPair(Candidates);
          Score = Res.second;
          IsConst =
              Res.second == BoUpSLP::LookAheadHeuristics::ScoreConstants &&
              isConstant(Candidates[Res.first.value_or(0)].first) &&
              isConstant(Candidates[Res.first.value_or(0)].second);
          if (IsConst) {
            // Check if there are splat candidates and consider them better
            // option.
            for (const auto [Idx, P] : enumerate(Candidates)) {
              if (!isConstant(P.first) && !isConstant(P.second) &&
                  P.second == P.first) {
                Res.first = Idx;
                IsConst = false;
                Score = isa<LoadInst>(Candidates[Res.first.value_or(0)].first)
                            ? BoUpSLP::LookAheadHeuristics::ScoreSplatLoads
                            : BoUpSLP::LookAheadHeuristics::ScoreSplat;
                break;
              }
            }
          }
          return Res.first;
        };
 
    for (Value *V : VL) {
      auto *I = dyn_cast<Instruction>(V);
      if (!I || (I == MainOp && (!S.isAltShuffle() || I == SMain)) ||
          (!S.isAltShuffle() && I == SMain))
        continue;
      SmallVector<BoUpSLP::ValueList> VOps;
      buildOriginalOperands(S, I == SMain ? MainOp : I, VOps);
      SmallVector<Value *> CopyableOps =
          getOperands(CopyableS, I == MainOp ? SMain : I);
      if (CopyableOps.size() == VOps.size() &&
          all_of(zip(CopyableOps, VOps), [&](const auto &P) {
            return std::get<0>(P) == std::get<1>(P)[0];
          }))
        continue;
      SmallVector<std::pair<Value *, Value *>> Candidates;
      BuildFirstOperandCandidates(Candidates, MainOps, CopyableOps[0],
                                  CopyableOps[1], IsMainCommutative);
      const unsigned OpSize = Candidates.size();
      Instruction *MatchingOp =
          S.getMatchingMainOpOrAltOp(I) == S.getMainOp() ? SMain : SAlt;
      const bool IsCommutativeInst =
          (MatchingOp == SMain ? IsCommutative : IsAltCommutative) ||
          isCommutative(I, MatchingOp);
      if (S.isAltShuffle() && MatchingOp == SAlt &&
          any_of(VOps, [&](const BoUpSLP::ValueList &Ops) {
            auto *I = dyn_cast<BinaryOperator>(Ops[0]);
            return I && Operands.contains(I);
          }))
        return false;
      if (S.isAltShuffle() && MatchingOp == SMain)
        Operands.insert(I->op_begin(), I->op_end());
      BuildFirstOperandCandidates(Candidates, Ops, VOps[0][0], VOps[1][0],
                                  IsCommutativeInst);
      bool IsBestConst;
      int Score;
      std::optional<int> BestOp =
          FindBestCandidate(Candidates, IsBestConst, Score);
      const bool IsOriginalBetter =
          static_cast<unsigned>(BestOp.value_or(OpSize)) >= OpSize;
      Candidates.clear();
      BuildSecondOperandCandidates(
          Candidates, MainOps, IsOriginalBetter ? -1 : *BestOp, CopyableOps[0],
          CopyableOps[1], IsMainCommutative);
      const unsigned SecondOpSize = Candidates.size();
      BuildSecondOperandCandidates(
          Candidates, Ops,
          IsOriginalBetter ? BestOp.value_or(OpSize - 1) - OpSize : -1,
          VOps[0][0], VOps[1][0], IsCommutativeInst);
      bool IsSecondBestConst;
      int SecondScore;
      std::optional<int> SecondBestOp =
          FindBestCandidate(Candidates, IsSecondBestConst, SecondScore);
      // No best candidates.
      if (!BestOp && !SecondBestOp)
        return false;
      // Original better in both ops combinations.
      const bool IsSecondOriginalBetter =
          static_cast<unsigned>(SecondBestOp.value_or(SecondOpSize)) >=
          SecondOpSize;
      if (IsOriginalBetter && IsSecondOriginalBetter)
        return false;
      // Original is better in second combination, but in the first combination
      // no best candidates.
      if (!BestOp && IsSecondOriginalBetter)
        return false;
      // Original is better in first combination, but in the second combination
      // no best candidates.
      if (!SecondBestOp && IsOriginalBetter)
        return false;
      // Copyable is best in the first combination, but it is constant, but
      // original is better in second non-constant combination.
      if (!IsOriginalBetter && IsBestConst && IsSecondOriginalBetter &&
          !IsSecondBestConst)
        return false;
      // Copyable is best in the second combination, but it is constant, but
      // original is better in the first non-constant combination.
      if (BestOp && IsOriginalBetter && !IsBestConst &&
          !IsSecondOriginalBetter && IsSecondBestConst)
        return false;
      // Original combination score is better.
      if (((Score > SecondScore ||
            (Score <= BoUpSLP::LookAheadHeuristics::ScoreAltOpcodes &&
             Score == SecondScore)) &&
           IsOriginalBetter) ||
          (IsSecondOriginalBetter &&
           (SecondScore > Score ||
            (Score <= BoUpSLP::LookAheadHeuristics::ScoreAltOpcodes &&
             Score == SecondScore))))
        return false;
    }
    return true;
  }
 
public:
  InstructionsCompatibilityAnalysis(DominatorTree &DT, const DataLayout &DL,
                                    const TargetTransformInfo &TTI,
                                    const TargetLibraryInfo &TLI)
      : DT(DT), DL(DL), TTI(TTI), TLI(TLI) {}
 
  InstructionsState buildInstructionsState(ArrayRef<Value *> VL,
                                           const BoUpSLP &R,
                                           bool WithProfitabilityCheck = false,
                                           bool SkipSameCodeCheck = false) {
    InstructionsState S = (SkipSameCodeCheck || !allSameBlock(VL))
                              ? InstructionsState::invalid()
                              : getSameOpcode(VL, TLI);
    // Check if series of selects + zext i1 %x to in can be combined into
    // selects + select %x, i32 1, i32 0.
    Instruction *SelectOp = nullptr;
    if (!S && allSameBlock(VL) && all_of(VL, [&](Value *V) {
          if (match(V, m_Select(m_Value(), m_Value(), m_Value()))) {
            if (!SelectOp)
              SelectOp = cast<Instruction>(V);
            return true;
          }
          auto *ZExt = dyn_cast<ZExtInst>(V);
          return (ZExt && ZExt->getSrcTy()->isIntegerTy(1)) ||
                 isa<PoisonValue>(V);
        })) {
      if (SelectOp)
        return InstructionsState(SelectOp, SelectOp);
    }
    if (S && S.isAltShuffle()) {
      Type *ScalarTy = S.getMainOp()->getType();
      auto *VecTy = cast<VectorType>(getWidenedType(ScalarTy, VL.size()));
      unsigned Opcode0 = S.getOpcode();
      unsigned Opcode1 = S.getAltOpcode();
      SmallBitVector OpcodeMask(
          getAltInstrMask(VL, ScalarTy, Opcode0, Opcode1));
      // If this pattern is supported by the target then we consider the order.
      if (TTI.isLegalAltInstr(VecTy, Opcode0, Opcode1, OpcodeMask))
        return S;
    } else if (S &&
               (!VectorizeCopyableElements ||
                !isSupportedMainOp(S.getMainOp()) || all_of(VL, [&](Value *V) {
                  auto *I = dyn_cast<Instruction>(V);
                  return !I || I->getOpcode() == S.getOpcode() ||
                         (S.getOpcode() == Instruction::Add &&
                          I->getOpcode() == Instruction::Shl);
                }))) {
      return S;
    }
    if (!VectorizeCopyableElements)
      return S;
    AbsorbCopyableFMulOrFAdds = hasOnlyAbsorbableCopyableFMulOrFAdds(VL);
    findAndSetMainInstruction(VL, R);
    if (!MainOp)
      return S;
    InstructionsState OrigS = S;
    S = InstructionsState(MainOp, MainOp, /*HasCopyables=*/true);
    if (OrigS && !isCopyablePreferable(VL, R, OrigS, S))
      return OrigS;
    if (!WithProfitabilityCheck)
      return S;
    // ExtractElement copyable nodes reuse the source vector and insert the
    // copyable lanes; the binary-operator operand heuristics below do not
    // apply, so defer profitability to the full tree cost.
    if (isa<ExtractElementInst>(MainOp)) {
      // A load as a copyable lane would be pulled out of the consecutive-load
      // vectorization; keep the original state so the node can be split.
      if (any_of(VL, [&](Value *V) {
            return S.isCopyableElement(V) && isa<LoadInst>(V);
          }))
        return OrigS;
      return S;
    }
    // Check if it is profitable to vectorize the instruction.
    unsigned CopyableNum =
        count_if(VL, [&](Value *V) { return S.isCopyableElement(V); });
    // Absorb copyable single-use fmuls/fadds as fmuladd(a, b, -0.0) or
    // fmuladd(1.0, a, b) when every copyable is such a binop: the binops die
    // instead of being computed and gathered.
    if (RecurrenceDescriptor::isFMulAddIntrinsic(MainOp) &&
        AbsorbCopyableFMulOrFAdds)
      S.setAbsorbCopyableFMulOrFAdd(true);
    SmallVector<BoUpSLP::ValueList> Operands = buildOperands(S, VL);
    auto BuildCandidates =
        [](SmallVectorImpl<std::pair<Value *, Value *>> &Candidates, Value *V1,
           Value *V2) {
          if (V1 != V2 && isa<PHINode>(V1))
            return;
          auto *I1 = dyn_cast<Instruction>(V1);
          auto *I2 = dyn_cast<Instruction>(V2);
          if (I1 && I2 && I1->getOpcode() == I2->getOpcode() &&
              I1->getParent() != I2->getParent())
            return;
          Candidates.emplace_back(V1, (I1 || I2) ? V2 : V1);
        };
    if (VL.size() == 2) {
      // The operand-pairing heuristic below does not apply to calls; defer
      // to the full tree cost computation instead of pre-rejecting here.
      if (MainOpcode == Instruction::Call)
        return S;
      // Check if the operands allow better vectorization.
      SmallVector<std::pair<Value *, Value *>, 4> Candidates1, Candidates2;
      BuildCandidates(Candidates1, Operands[0][0], Operands[0][1]);
      BuildCandidates(Candidates2, Operands[1][0], Operands[1][1]);
      bool Res = !Candidates1.empty() && !Candidates2.empty() &&
                 R.findBestRootPair(Candidates1).first &&
                 R.findBestRootPair(Candidates2).first;
      if (!Res && isCommutative(MainOp)) {
        Candidates1.clear();
        Candidates2.clear();
        BuildCandidates(Candidates1, Operands[0][0], Operands[1][1]);
        BuildCandidates(Candidates2, Operands[1][0], Operands[0][1]);
        Res = !Candidates1.empty() && !Candidates2.empty() &&
              R.findBestRootPair(Candidates1).first &&
              R.findBestRootPair(Candidates2).first;
      }
      if (!Res)
        return OrigS;
      const TTI::TargetCostKind Kind = R.getCostKind();
      InstructionCost ScalarCost = TTI.getInstructionCost(S.getMainOp(), Kind);
      InstructionCost VectorCost;
      auto *VecTy = getWidenedType(S.getMainOp()->getType(), VL.size());
      switch (MainOpcode) {
      case Instruction::Add:
      case Instruction::Sub:
      case Instruction::Mul:
      case Instruction::AShr:
      case Instruction::LShr:
      case Instruction::Shl:
      case Instruction::SDiv:
      case Instruction::UDiv:
      case Instruction::And:
      case Instruction::Or:
      case Instruction::Xor:
      case Instruction::FAdd:
      case Instruction::FMul:
      case Instruction::FSub:
      case Instruction::FDiv:
        VectorCost = TTI.getArithmeticInstrCost(MainOpcode, VecTy, Kind);
        break;
      default:
        // Calls (min/max, fmuladd) return above before reaching this switch.
        llvm_unreachable("Unexpected instruction.");
      }
      if (VectorCost > ScalarCost)
        return OrigS;
      return S;
    }
    // fmuladd is the only 3-operand copyable.
    assert((Operands.size() == 2 ||
            (Operands.size() == 3 &&
             RecurrenceDescriptor::isFMulAddIntrinsic(MainOp))) &&
           "Unexpected number of operands!");
    if (CopyableNum < VL.size() / 2)
      return S;
    // Too many phi copyables - exit.
    const unsigned Limit = VL.size() / 24;
    if ((CopyableNum >= VL.size() - Limit ||
         (CopyableNum >= VL.size() - 1 && VL.size() > 4) ||
         CopyableNum >= MaxPHINumOperands) &&
        all_of(VL, [&](Value *V) {
          return isa<PHINode>(V) || !S.isCopyableElement(V);
        }))
      return OrigS;
    // Check profitability if number of copyables > VL.size() / 2.
    // 1. Reorder operands for better matching.
    if (Operands.size() == 2 && isCommutative(MainOp)) {
      Value *BestFrontOp = nullptr;
      for (auto [OpL, OpR] : zip(Operands.front(), Operands.back())) {
        // Make instructions the first operands.
        if (!isa<Instruction>(OpL) && isa<Instruction>(OpR)) {
          BestFrontOp = OpR;
          std::swap(OpL, OpR);
          continue;
        }
        // Make constants the second operands.
        if ((isa<Constant>(OpL) && !match(OpR, m_Zero())) ||
            match(OpL, m_Zero())) {
          if (isa<Instruction>(OpR))
            BestFrontOp = OpR;
          std::swap(OpL, OpR);
          continue;
        }
        if (isa<Instruction>(OpL))
          BestFrontOp = OpL;
      }
      // If some of the RHS operands better match most of LHS - swap such
      // operands to increase matching rate.
      if (auto *BestLHS = dyn_cast_if_present<Instruction>(BestFrontOp)) {
        const unsigned BestOpcode = BestLHS->getOpcode();
        for (auto [OpL, OpR] : zip(Operands.front(), Operands.back())) {
          auto *OpRI = dyn_cast<Instruction>(OpR);
          if (!OpRI)
            continue;
          if (OpRI->getOpcode() == BestOpcode)
            std::swap(OpL, OpR);
        }
      }
    }
    // 2. Check, if operands can be vectorized. Skip for fmuladd; the
    // copyable operand is checked below and may legitimately hold many
    // instructions.
    if (Operands.size() == 2 &&
        count_if(Operands.back(), IsaPred<Instruction>) > 1)
      return OrigS;
    auto CheckOperand = [&](ArrayRef<Value *> Ops) {
      if (allConstant(Ops) || isSplat(Ops))
        return true;
      // Non-instruction operands of a call (args, constants) are always a
      // trivial gather, same as the constant/splat cases above.
      if (MainOpcode == Instruction::Call && none_of(Ops, IsaPred<Instruction>))
        return true;
      // Check if it is "almost" splat, i.e. has >= 4 elements and only single
      // one is different.
      constexpr unsigned Limit = 4;
      if (Operands.front().size() >= Limit) {
        SmallDenseMap<const Value *, unsigned> Counters;
        for (Value *V : Ops) {
          if (isa<UndefValue>(V))
            continue;
          ++Counters[V];
        }
        if (Counters.size() == 2 &&
            any_of(Counters, [&](const std::pair<const Value *, unsigned> &C) {
              return C.second == 1;
            }))
          return true;
      }
      // First operand not a constant or splat? Last attempt - check for
      // potential vectorization.
      InstructionsCompatibilityAnalysis Analysis(DT, DL, TTI, TLI);
      InstructionsState OpS = Analysis.buildInstructionsState(Ops, R);
      if (!OpS || (OpS.getOpcode() == Instruction::PHI && !allSameBlock(Ops)))
        return false;
      unsigned CopyableNum =
          count_if(Ops, [&](Value *V) { return OpS.isCopyableElement(V); });
      return CopyableNum <= VL.size() / 2;
    };
    // Check the operand holding the copyable values.
    if (!CheckOperand(Operands[S.getCopyableOpIdx()])) {
      if (Operands.size() == 2)
        return OrigS;
      // Retry with the copyable modeled as the first multiplicand.
      S.setCopyableOpIdx(0);
      Operands = buildOperands(S, VL);
      if (!CheckOperand(Operands[S.getCopyableOpIdx()]))
        return OrigS;
    }
 
    return S;
  }
 
  SmallVector<BoUpSLP::ValueList> buildOperands(const InstructionsState &S,
                                                ArrayRef<Value *> VL) {
    assert(S && "Invalid state!");
    SmallVector<BoUpSLP::ValueList> Operands;
    if (S.areInstructionsWithCopyableElements()) {
      MainOp = S.getMainOp();
      MainOpcode = S.getOpcode();
      // ExtractElement copyable nodes carry a single operand (the shared
      // source vector); copyable lanes are inserted during codegen and do
      // not contribute an operand column.
      if (MainOpcode == Instruction::ExtractElement) {
        Operands.assign(1,
                        BoUpSLP::ValueList(VL.size(), MainOp->getOperand(0)));
        return Operands;
      }
      // Excludes the trailing callee operand (2 for min/max, 3 for fmuladd).
      // getNumberOfPotentiallyCommutativeOps collapses fmuladd to 2 and must
      // not be used here. Only the 2-operand case is commutative-normalized.
      auto *CI = dyn_cast<CallInst>(MainOp);
      const unsigned NumMainOpOperands =
          CI ? CI->arg_size() : MainOp->getNumOperands();
      const bool IsCommutative =
          isCommutative(MainOp) && NumMainOpOperands == 2;
      Operands.assign(NumMainOpOperands,
                      BoUpSLP::ValueList(VL.size(), nullptr));
      // Populate operands for every lane.
      for (auto [Idx, V] : enumerate(VL)) {
        SmallVector<Value *> OperandsForValue = getOperands(S, V);
        for (auto [OperandIdx, Operand] : enumerate(OperandsForValue))
          Operands[OperandIdx][Idx] = Operand;
      }
      // Operand-order normalization below swaps OpIdx 0 and OpIdx 1
      // of non-copyable lanes. That is only safe when the main op is
      // commutative (e.g. 0 - X is not X - 0, so `sub` must be
      // excluded). With absorbed fmul/fadd copyables the fmuladd
      // multiplicands are commutative per lane and get normalized too;
      // the 0/1 swaps never touch the addend column.
      if (IsCommutative || S.hasAbsorbedCopyableFMulOrFAdd()) {
        // IsCommutative can hold for MainOp (e.g. a Sub/FSub feeding only
        // fabs/icmp-eq-0) without every lane sharing that property, so
        // re-check the specific lane before swapping it. Absorbed fmul/fadd
        // lanes are always commutative.
        auto CanSwap = [&](Value *V) {
          if (S.hasAbsorbedCopyableFMulOrFAdd() &&
              isAbsorbableCopyableFMulOrFAdd(S, V))
            return true;
          return isCommutative(S.getMatchingMainOpOrAltOp(cast<Instruction>(V)),
                               V);
        };
        // Absorbed fmul/fadd copyables do not vote for the majority operand
        // pattern (their multiplicand order is arbitrary) but take part
        // in the swaps.
        auto SwappableLane = [&](Value *V) {
          return !isa<PoisonValue>(V) &&
                 (!S.isCopyableElement(V) ||
                  (S.hasAbsorbedCopyableFMulOrFAdd() &&
                   isAbsorbableCopyableFMulOrFAdd(S, V)));
        };
        // Count (ID0, ID1) pair frequencies for operand normalization.
        // Pairs and their inverses are tracked under a canonical key
        // so that (Load, Add) and (Add, Load) contribute to the same
        // bucket.
        struct PairInfo {
          unsigned FwdCount = 0;
          unsigned RevCount = 0;
        };
        SmallMapVector<std::pair<unsigned, unsigned>, PairInfo, 8> PairCounts;
        SmallMapVector<unsigned, unsigned, 4> AddendIDCounts;
        unsigned MajID0 = 0, MajID1 = 0;
        for (auto [Idx, V] : enumerate(VL)) {
          if (S.isCopyableElement(V) || isa<PoisonValue>(V))
            continue;
          unsigned ID0 = Operands[0][Idx]->getValueID();
          unsigned ID1 = Operands[1][Idx]->getValueID();
          if (S.hasAbsorbedCopyableFMulOrFAdd())
            ++AddendIDCounts[Operands[2][Idx]->getValueID()];
          if (ID0 == ID1)
            continue;
          unsigned MinID = std::min(ID0, ID1);
          unsigned MaxID = std::max(ID0, ID1);
          auto [It, Inserted] =
              PairCounts.try_emplace(std::make_pair(MinID, MaxID));
          PairInfo &Info = It->second;
          if (ID0 < ID1)
            ++Info.FwdCount;
          else
            ++Info.RevCount;
        }
        // Find the most frequent (ID0, ID1) pair across non-copyable
        // lanes. Select the orientation (original or inverse) that
        // has more votes as the majority pattern.
        unsigned BestCount = 0;
        for (const auto &P : PairCounts) {
          const PairInfo &Info = P.second;
          unsigned Total = Info.FwdCount + Info.RevCount;
          if (Total > BestCount) {
            BestCount = Total;
            if (Info.FwdCount >= Info.RevCount) {
              MajID0 = P.first.first;
              MajID1 = P.first.second;
            } else {
              MajID0 = P.first.second;
              MajID1 = P.first.first;
            }
          }
        }
        // Absorbed fadd copyables are fmuladd(1.0, a, b): pick the
        // addend/multiplicand assignment matching the majority operand
        // kinds of the non-copyable lanes in the multiplicand (1) and
        // addend (2) columns.
        if (S.hasAbsorbedCopyableFMulOrFAdd()) {
          unsigned MajID2 = 0, Best2 = 0;
          for (const auto &P : AddendIDCounts) {
            if (P.second > Best2) {
              Best2 = P.second;
              MajID2 = P.first;
            }
          }
          for (auto [Idx, V] : enumerate(VL)) {
            auto *I = dyn_cast<Instruction>(V);
            if (!I || I->getOpcode() != Instruction::FAdd ||
                !isAbsorbableCopyableFMulOrFAdd(S, I))
              continue;
            unsigned ID1 = Operands[1][Idx]->getValueID();
            unsigned ID2 = Operands[2][Idx]->getValueID();
            unsigned Cur = (ID1 == MajID1) + (ID2 == MajID2);
            unsigned Swapped = (ID2 == MajID1) + (ID1 == MajID2);
            if (Swapped > Cur)
              std::swap(Operands[1][Idx], Operands[2][Idx]);
          }
        }
        // Normalize swappable lanes in two steps:
        // 1) Swap lanes whose operand types are the exact inverse of
        //    the majority pattern, making the non-copyable lanes
        //    consistent.
        // 2) Independently, if a strict majority of swappable lanes
        //    have loads at OpIdx 1, swap those lanes to put loads at
        //    OpIdx 0 for better downstream vectorization.
        unsigned LAt0 = 0, LAt1 = 0, TotalNC = 0;
        for (auto [Idx, V] : enumerate(VL)) {
          if (!SwappableLane(V))
            continue;
          // Step 1: swap exact-inverse lanes.
          if (BestCount > 0) {
            unsigned ID0 = Operands[0][Idx]->getValueID();
            unsigned ID1 = Operands[1][Idx]->getValueID();
            if (ID0 == MajID1 && ID1 == MajID0 && CanSwap(V))
              std::swap(Operands[0][Idx], Operands[1][Idx]);
          }
          ++TotalNC;
          LAt0 += isa<LoadInst>(Operands[0][Idx]);
          LAt1 += isa<LoadInst>(Operands[1][Idx]);
        }
        // Step 2: if most non-copyable lanes have loads at OpIdx 1,
        // swap those lanes to put loads at OpIdx 0.
        if (TotalNC > 1 && LAt1 > LAt0 && LAt1 * 2 > TotalNC) {
          for (auto [Idx, V] : enumerate(VL)) {
            if (!SwappableLane(V))
              continue;
            if (!isa<LoadInst>(Operands[0][Idx]) &&
                isa<LoadInst>(Operands[1][Idx]) && CanSwap(V))
              std::swap(Operands[0][Idx], Operands[1][Idx]);
          }
        }
      }
    } else {
      buildOriginalOperands(S, VL, Operands);
    }
    return Operands;
  }
};
} // namespace
 
void BoUpSLP::tryToVectorizeSplatGatheredScalars() {
  auto LoadsSubkey = [](size_t /*Key*/, LoadInst *LI) {
    return hash_value(getUnderlyingObject(LI->getPointerOperand()));
  };
  // The key includes the value type: the opcode-based key does not always
  // distinguish types (e.g. all extractvalue instructions share one key),
  // while the tree requires same-typed scalars.
  SmallMapVector<std::tuple<size_t, size_t, Type *>, SmallSetVector<Value *, 4>,
                 4>
      Groups;
  for (const std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
    // Only gathers with vectorized (non-gather) users can reuse the broadcast.
    if (!TE->isGather() || !TE->UserTreeIndex ||
        TE->UserTreeIndex.UserTE->isGather() || !isSplat(TE->Scalars))
      continue;
    auto *I = dyn_cast<Instruction>(TE->Scalars.front());
    // Skip shuffle-like instructions: their splat gathers are already emitted
    // as cheap shuffles of the source vector.
    if (!I ||
        isa<ExtractElementInst, InsertElementInst, ShuffleVectorInst>(I) ||
        I->getType()->isVoidTy() || isVectorized(I) || isDeleted(I) ||
        (UserIgnoreList && UserIgnoreList->contains(I)))
      continue;
    // Scheduling new memory bundles in a to-be-versioned tree records extra
    // alias-check pairs and can push the region over the versioning limits.
    if (isTryingRuntimeAliasChecks() && I->mayReadOrWriteMemory())
      continue;
    auto [Key, SubKey] =
        generateKeySubkey(I, TLI, LoadsSubkey, /*AllowAlternate=*/true);
    Groups[std::make_tuple(Key, SubKey, I->getType())].insert(I);
  }
  // Values left in singleton groups cannot form a bundle on their own;
  // regroup them by the opcode-insensitive key so alternate/copyable
  // bundles still form.
  SmallMapVector<std::pair<size_t, Type *>, SmallSetVector<Value *, 4>, 4>
      FallbackGroups;
  for (auto &[Key, Group] : Groups) {
    if (Group.size() >= 2)
      continue;
    FallbackGroups[std::make_pair(std::get<0>(Key), Group.front()->getType())]
        .insert(Group.front());
  }
  InstructionsCompatibilityAnalysis Analysis(*DT, *DL, *TTI, *TLI);
  auto BuildSubtree = [&](const auto &GroupMap) {
    for (const auto &[_, Group] : GroupMap) {
      if (Group.size() < 2)
        continue;
      // Copyable-aware check so bundles with copyable lanes are not skipped.
      if (!Analysis.buildInstructionsState(Group.getArrayRef(), *this))
        continue;
      unsigned PrevSize = VectorizableTree.size();
      buildTreeRec(Group.getArrayRef(), 0, EdgeInfo());
      if (PrevSize == VectorizableTree.size())
        continue;
      TreeEntry *NewRoot = VectorizableTree[PrevSize].get();
      if (NewRoot->isGather()) {
        // Failed to vectorize the bundle: drop the added gather entry, it has
        // no users and only adds cost.
        for (Value *V : NewRoot->Scalars) {
          auto It = ValueToGatherNodes.find(V);
          if (It != ValueToGatherNodes.end())
            It->second.remove(NewRoot);
        }
        LoadEntriesToVectorize.remove(PrevSize);
        VectorizableTree.pop_back();
        continue;
      }
      SplatGatheredScalarsRoots.push_back(NewRoot);
    }
  };
  BuildSubtree(Groups);
  BuildSubtree(FallbackGroups);
}
 
BoUpSLP::ScalarsVectorizationLegality
BoUpSLP::getScalarsVectorizationLegality(ArrayRef<Value *> VL, unsigned Depth,
                                         const EdgeInfo &UserTreeIdx) const {
  assert((allConstant(VL) || allSameType(VL)) && "Invalid types!");
 
  InstructionsCompatibilityAnalysis Analysis(*DT, *DL, *TTI, *TLI);
  InstructionsState S = Analysis.buildInstructionsState(
      VL, *this, /*WithProfitabilityCheck=*/true);
 
  bool AreScatterAllGEPSameBlock = false;
  if (!S) {
    SmallVector<unsigned> SortedIndices;
    BasicBlock *BB = nullptr;
    bool IsScatterVectorizeUserTE =
        UserTreeIdx.UserTE &&
        UserTreeIdx.UserTE->State == TreeEntry::ScatterVectorize;
    AreScatterAllGEPSameBlock =
        (IsScatterVectorizeUserTE && VL.front()->getType()->isPointerTy() &&
         VL.size() > 2 &&
         all_of(VL,
                [&BB](Value *V) {
                  auto *I = dyn_cast<GetElementPtrInst>(V);
                  if (!I)
                    return doesNotNeedToBeScheduled(V);
                  if (!BB)
                    BB = I->getParent();
                  return BB == I->getParent() && I->getNumOperands() == 2;
                }) &&
         BB &&
         sortPtrAccesses(VL, UserTreeIdx.UserTE->getMainOp()->getType(), *DL,
                         *SE, SortedIndices));
    if (!AreScatterAllGEPSameBlock) {
      LLVM_DEBUG(dbgs() << "SLP: Try split and if failed, gathering due to "
                           "C,S,B,O, small shuffle. \n";
                 dbgs() << "[";
                 interleaveComma(VL, dbgs(), [&](Value *V) { dbgs() << *V; });
                 dbgs() << "]\n");
      return ScalarsVectorizationLegality(S, /*IsLegal=*/false,
                                          /*TryToFindDuplicates=*/true,
                                          /*TrySplitVectorize=*/true);
    }
    // Reset S to make it GetElementPtr kind of node.
    const auto *It = find_if(VL, IsaPred<GetElementPtrInst>);
    assert(It != VL.end() && "Expected at least one GEP.");
    S = getSameOpcode(*It, *TLI);
  }
  assert(S && "Must be valid.");
 
  // Gather very wide PHI bundles. Wide PHIs (e.g. produced by
  // jump threading) are not profitable to vectorize and make this analysis
  // explode, so gather them to keep the compile time bounded.
  if (S.getOpcode() == Instruction::PHI) {
    unsigned NumIncomingValues =
        cast<PHINode>(S.getMainOp())->getNumIncomingValues();
    if (static_cast<uint64_t>(VL.size()) * NumIncomingValues >
        PHINodeVectorizationBudget) {
      LLVM_DEBUG(dbgs() << "SLP: Gathering due to wide PHI operand fan-out ("
                        << VL.size() << " lanes x " << NumIncomingValues
                        << " incoming values).\n");
      return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
    }
  }
 
  // Don't handle vectors.
  if (!SLPReVec && getValueType(VL.front(), SLPReVec)->isVectorTy()) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to vector type.\n");
    // Do not try to pack to avoid extra instructions here.
    return ScalarsVectorizationLegality(S, /*IsLegal=*/false,
                                        /*TryToFindDuplicates=*/false);
  }
 
  // Check that all of the users of the scalars that we want to vectorize are
  // schedulable.
  BasicBlock *BB = S.getMainOp()->getParent();
 
  if (BB->isEHPad() || isa_and_nonnull<UnreachableInst>(BB->getTerminator()) ||
      !DT->isReachableFromEntry(BB)) {
    // Don't go into unreachable blocks. They may contain instructions with
    // dependency cycles which confuse the final scheduling.
    // Do not vectorize EH and non-returning blocks, not profitable in most
    // cases.
    LLVM_DEBUG(dbgs() << "SLP: bundle in unreachable block.\n");
    return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
  }
 
  // Don't go into catchswitch blocks, which can happen with PHIs.
  // Such blocks can only have PHIs and the catchswitch.  There is no
  // place to insert a shuffle if we need to, so just avoid that issue.
  if (isa<CatchSwitchInst>(BB->getTerminator())) {
    LLVM_DEBUG(dbgs() << "SLP: bundle in catchswitch block.\n");
    // Do not try to pack to avoid extra instructions here.
    return ScalarsVectorizationLegality(S, /*IsLegal=*/false,
                                        /*TryToFindDuplicates=*/false);
  }
 
  // Don't handle scalable vectors
  if (S.getOpcode() == Instruction::ExtractElement &&
      isa<ScalableVectorType>(
          cast<ExtractElementInst>(S.getMainOp())->getVectorOperandType())) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to scalable vector type.\n");
    return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
  }
 
  // Gather if we hit the RecursionMaxDepth, unless this is a load (or z/sext of
  // a load), in which case peek through to include it in the tree, without
  // ballooning over-budget.
  if (Depth >= RecursionMaxDepth &&
      (S.isAltShuffle() || VL.size() < 4 ||
       !(match(S.getMainOp(), m_Load(m_Value())) ||
         all_of(VL, [&S](const Value *I) {
           return match(I,
                        m_OneUse(m_ZExtOrSExt(m_OneUse(m_Load(m_Value()))))) &&
                  cast<Instruction>(I)->getOpcode() == S.getOpcode();
         })))) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to max recursion depth.\n");
    return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
  }
 
  // Check if this is a duplicate of another entry.
  LLVM_DEBUG(dbgs() << "SLP: \tChecking bundle: " << *S.getMainOp() << ".\n");
  // Cache invariants to avoid recomputing for every V in VL (and every E).
  const bool IsPHIWithLoop =
      S.getOpcode() == Instruction::PHI &&
      LI->getLoopFor(S.getMainOp()->getParent()) != nullptr;
  for (TreeEntry *E : getTreeEntries(S.getMainOp())) {
    if (E->isSame(VL)) {
      LLVM_DEBUG(dbgs() << "SLP: Perfect diamond merge at " << *S.getMainOp()
                        << ".\n");
      return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
    }
    SmallPtrSet<Value *, 8> Values(llvm::from_range, E->Scalars);
    if (all_of(VL, [&](Value *V) {
          return isa<PoisonValue>(V) || Values.contains(V) ||
                 (IsPHIWithLoop && isa<PHINode>(V) && isVectorized(V));
        })) {
      LLVM_DEBUG(dbgs() << "SLP: Gathering due to full overlap.\n");
      return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
    }
  }
 
  bool AreAllSameBlock = !AreScatterAllGEPSameBlock;
  bool AreAllSameInsts = AreAllSameBlock || AreScatterAllGEPSameBlock;
  if (!AreAllSameInsts || isSplat(VL) ||
      (isa<InsertElementInst, ExtractValueInst, ExtractElementInst>(
           S.getMainOp()) &&
       !all_of(VL, isVectorLikeInstWithConstOps))) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to C,S,B,O conditions. \n";
               dbgs() << "[";
               interleaveComma(VL, dbgs(), [&](Value *V) { dbgs() << *V; });
               dbgs() << "]\n");
    return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
  }
 
  // Don't vectorize ephemeral values.
  if (!EphValues.empty()) {
    for (Value *V : VL) {
      if (EphValues.count(V)) {
        LLVM_DEBUG(dbgs() << "SLP: The instruction (" << *V
                          << ") is ephemeral.\n");
        // Do not try to pack to avoid extra instructions here.
        return ScalarsVectorizationLegality(S, /*IsLegal=*/false,
                                            /*TryToFindDuplicates=*/false);
      }
    }
  }
 
  // We now know that this is a vector of instructions of the same type from
  // the same block.
 
  // Check that none of the instructions in the bundle are already in the tree
  // and the node may be not profitable for the vectorization as the small
  // alternate node.
  if (S.isAltShuffle()) {
    auto GetNumVectorizedExtracted = [&]() {
      APInt Extracted = APInt::getZero(VL.size());
      APInt Vectorized = APInt::getAllOnes(VL.size());
      for (auto [Idx, V] : enumerate(VL)) {
        auto *I = dyn_cast<Instruction>(V);
        if (!I || doesNotNeedToBeScheduled(I) ||
            all_of(I->operands(), [&](const Use &U) {
              return isa<ExtractElementInst>(U.get());
            }))
          continue;
        if (isVectorized(I))
          Vectorized.clearBit(Idx);
        else if (!I->hasOneUser() && !areAllUsersVectorized(I, UserIgnoreList))
          Extracted.setBit(Idx);
      }
      return std::make_pair(Vectorized, Extracted);
    };
    auto [Vectorized, Extracted] = GetNumVectorizedExtracted();
    bool PreferScalarize = !Vectorized.isAllOnes() && VL.size() == 2;
    if (!Vectorized.isAllOnes() && !PreferScalarize) {
      // Rough cost estimation, if the vector code (+ potential extracts) is
      // more profitable than the scalar + buildvector.
      Type *ScalarTy = VL.front()->getType();
      auto *VecTy = cast<VectorType>(getWidenedType(ScalarTy, VL.size()));
      InstructionCost VectorizeCostEstimate =
          getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc, VecTy, CostKind) +
          getScalarizationOverhead(*TTI, SLPReVec, ScalarTy, VecTy, Extracted,
                                   /*Insert=*/false, /*Extract=*/true,
                                   CostKind);
      InstructionCost ScalarizeCostEstimate = getScalarizationOverhead(
          *TTI, SLPReVec, ScalarTy, VecTy, Vectorized,
          /*Insert=*/true, /*Extract=*/false, CostKind, /*ForPoisonSrc=*/false);
      PreferScalarize = VectorizeCostEstimate > ScalarizeCostEstimate;
    }
    if (PreferScalarize) {
      LLVM_DEBUG(dbgs() << "SLP: The instructions are in tree and alternate "
                           "node is not profitable.\n");
      return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
    }
  }
 
  // The reduction nodes (stored in UserIgnoreList) also should stay scalar.
  if (UserIgnoreList && !UserIgnoreList->empty()) {
    for (Value *V : VL) {
      if (UserIgnoreList->contains(V)) {
        LLVM_DEBUG(dbgs() << "SLP: Gathering due to gathered scalar.\n");
        return ScalarsVectorizationLegality(S, /*IsLegal=*/false);
      }
    }
  }
 
  return ScalarsVectorizationLegality(S, /*IsLegal=*/true);
}
 
/// Peel associative single-use binary chains into operand columns. Peeled
/// instructions go to \p ReassocScalars; sub-operands are inserted after the
/// peeled column so real pairs stay adjacent for flag reuse.
/// Poison lanes are allowed through; a lane that is not itself a matching
/// real instruction can still be peeled as a copyable identity leaf, as
/// long as some other lane anchors the opcode.
/// Add/sub (and fadd/fsub) chains peel together: a peeled subtract keeps
/// the column sign for its first operand and flips it for the second one,
/// recorded per column in \p NegatedColumns, so subtracted leaves stay out
/// of the positive total.
static void scanAssociativeOperands(
    const InstructionsState &S, DominatorTree &DT, const DataLayout &DL,
    const TargetTransformInfo &TTI, const TargetLibraryInfo &TLI,
    const BoUpSLP &R, SmallVectorImpl<BoUpSLP::ValueList> &Operands,
    SmallBitVector &NegatedColumns, SmallVectorImpl<Value *> &ReassocScalars) {
  assert(Operands.size() == 2 && NegatedColumns.empty() &&
         "Expected the initial 2 operand columns.");
  // The subtract family flips the sign of its second operand column.
  auto IsSub = [](unsigned Opcode) {
    return Opcode == Instruction::Sub || Opcode == Instruction::FSub;
  };
  // Signs ride with the columns so the parallel lists cannot drift apart.
  struct SignedColumn {
    BoUpSLP::ValueList Col;
    bool Negated;
  };
  SmallVector<SignedColumn, 4> Columns = {
      {std::move(Operands[0]), false},
      {std::move(Operands[1]), IsSub(S.getOpcode())}};
  InstructionsCompatibilityAnalysis Analysis(DT, DL, TTI, TLI);
  // The opcode family a column may peel into: subtracts flatten as adds of
  // a negated leaf, other opcodes peel only into their own kind.
  const unsigned RootFamily = getReassocCombineOpcode(S.getOpcode());
  // A lane whose value is itself a single-use chain link; block placement
  // does not matter here (buildTreeRec re-checks that later for whatever
  // columns end up as real leaves).
  auto IsRealLane = [&](Value *V) {
    auto *I = dyn_cast<Instruction>(V);
    return I && I->hasOneUse() &&
           getReassocCombineOpcode(I->getOpcode()) == RootFamily &&
           isReassocChainLink(I);
  };
  // Builds on InstructionsCompatibilityAnalysis's own opcode matching
  // (rather than the more permissive getSameOpcode()) so a column that
  // falls back to a copyable identity leaf only ever anchors on a MainOp
  // that InstructionsCompatibilityAnalysis::isSupportedOpcode() accepts.
  auto CanPeel = [&](ArrayRef<Value *> Column) {
    InstructionsState ColS = Analysis.buildInstructionsState(Column, R);
    if (!ColS || getReassocCombineOpcode(ColS.getOpcode()) != RootFamily ||
        ColS.isAltShuffle())
      return InstructionsState::invalid();
    // Every lane must be poison, a genuine matching real lane, or (only for
    // a copyable ColS) stand in as a copyable identity leaf
    // (Opcode(V, identity) == V).
    if (!all_of(Column, [&](Value *V) {
          return isa<PoisonValue>(V) || IsRealLane(V) ||
                 (ColS.areInstructionsWithCopyableElements() &&
                  ColS.isCopyableElement(V));
        }))
      return InstructionsState::invalid();
    return ColS;
  };
  for (unsigned Idx = 0; Idx != Columns.size();) {
    InstructionsState ColS = CanPeel(Columns[Idx].Col);
    if (!ColS) {
      ++Idx;
      continue;
    }
    BoUpSLP::ValueList Column = std::move(Columns[Idx].Col);
    SmallVector<BoUpSLP::ValueList> SubOperands =
        Analysis.buildOperands(ColS, Column);
    assert(SubOperands.size() == 2 && "Expected 2 operand columns.");
    // Poison and copyable lanes have no real instruction left to erase
    // later: a copyable V is used as-is, not subsumed by the flattened
    // combine.
    for (Value *V : Column)
      if (!isa<PoisonValue>(V) && !ColS.isCopyableElement(V))
        ReassocScalars.push_back(V);
    // A peeled subtract negates its second operand column.
    const bool Negated = Columns[Idx].Negated;
    Columns[Idx].Col = std::move(SubOperands.front());
    Columns.insert(
        std::next(Columns.begin(), Idx + 1),
        {std::move(SubOperands.back()), IsSub(ColS.getOpcode()) != Negated});
    // Do not advance Idx: re-examine the column that was just placed here.
  }
  Operands.clear();
  for (auto &[Col, Negated] : Columns) {
    Operands.push_back(std::move(Col));
    NegatedColumns.push_back(Negated);
  }
}
 
/// Realign flattened operand columns so matching value families share a
/// column across lanes. Lane 0 is canonical; other lanes match by key equality
/// only (never compare hash values). Columns sharing a key are paired by the
/// family of their first operand where available, so e.g. shifts fed by the
/// same load family land in one column instead of pairing by encounter order.
/// Values move between columns only within the same sign: a subtracted leaf
/// never lands in an added column. The sign is queried per lane and column
/// with \p IsNegated: alternate add/sub nodes negate only the non-leading
/// columns of their subtract lanes.
static SmallVector<BoUpSLP::ValueList> alignReassociatedOperandsByKey(
    ArrayRef<BoUpSLP::ValueList> Operands,
    function_ref<bool(unsigned Lane, unsigned Col)> IsNegated,
    const TargetLibraryInfo &TLI) {
  const unsigned NumCols = Operands.size();
  const unsigned NumLanes = Operands.front().size();
  auto LoadsSubkey = [](size_t /*Key*/, LoadInst *LI) {
    return hash_value(getUnderlyingObject(LI->getPointerOperand()));
  };
  auto GetKey = [&](Value *V) {
    return generateKeySubkey(V, &TLI, LoadsSubkey, /*AllowAlternate=*/false);
  };
  const std::pair<size_t, size_t> NoSecondaryKey{0, 0};
  // Family of the value's first operand; loads already carry their family in
  // the primary key.
  auto GetSecondaryKey = [&](Value *V) -> std::pair<size_t, size_t> {
    auto *I = dyn_cast<Instruction>(V);
    if (!I || isa<LoadInst>(I) || I->getNumOperands() == 0)
      return NoSecondaryKey;
    return GetKey(I->getOperand(0));
  };
  SmallVector<std::pair<size_t, size_t>> Lane0Keys = map_to_vector(
      Operands, [&](const BoUpSLP::ValueList &Col) { return GetKey(Col[0]); });
  SmallVector<std::pair<size_t, size_t>> Lane0SecondaryKeys =
      map_to_vector(Operands, [&](const BoUpSLP::ValueList &Col) {
        return GetSecondaryKey(Col[0]);
      });
 
  SmallVector<BoUpSLP::ValueList> Aligned(NumCols,
                                          BoUpSLP::ValueList(NumLanes));
  for (unsigned Col : seq<unsigned>(NumCols))
    Aligned[Col][0] = Operands[Col][0];
 
  for (unsigned Lane : seq<unsigned>(1, NumLanes)) {
    // Buckets are keyed by the value key and the column sign.
    using Key = std::pair<std::pair<size_t, size_t>, unsigned>;
    auto Sign = [&](unsigned Col) { return IsNegated(Lane, Col) ? 1U : 0U; };
    SmallDenseMap<Key, SmallVector<unsigned, 2>, 8> Buckets;
    for (unsigned Col : seq<unsigned>(NumCols))
      Buckets[{GetKey(Operands[Col][Lane]), Sign(Col)}].push_back(Col);
    SmallDenseMap<Key, unsigned, 8> NextInBucket;
    SmallVector<unsigned> SlotSrcCol(NumCols, NumCols);
    SmallBitVector ColClaimed(NumCols, false);
    for (unsigned Slot : seq<unsigned>(NumCols)) {
      const auto SlotKey = std::make_pair(Lane0Keys[Slot], Sign(Slot));
      auto BucketIt = Buckets.find(SlotKey);
      if (BucketIt == Buckets.end())
        continue;
      unsigned &Cursor = NextInBucket[SlotKey];
      if (Cursor >= BucketIt->second.size())
        continue;
      // Among the remaining same-key columns prefer the one whose operand
      // family matches lane 0's; no match keeps the bucket order.
      unsigned PickPos = Cursor;
      if (Lane0SecondaryKeys[Slot] != NoSecondaryKey)
        for (unsigned P = Cursor, E = BucketIt->second.size(); P != E; ++P)
          if (GetSecondaryKey(Operands[BucketIt->second[P]][Lane]) ==
              Lane0SecondaryKeys[Slot]) {
            PickPos = P;
            break;
          }
      std::swap(BucketIt->second[Cursor], BucketIt->second[PickPos]);
      unsigned SrcCol = BucketIt->second[Cursor++];
      SlotSrcCol[Slot] = SrcCol;
      ColClaimed[SrcCol] = true;
    }
    // Leftover slots take leftover columns of the same sign; per-lane sign
    // counts match, so every slot finds one.
    for (unsigned Slot : seq<unsigned>(NumCols)) {
      if (SlotSrcCol[Slot] != NumCols)
        continue;
      for (unsigned Col : seq<unsigned>(NumCols)) {
        if (!ColClaimed[Col] && IsNegated(Lane, Col) == IsNegated(Lane, Slot)) {
          SlotSrcCol[Slot] = Col;
          ColClaimed[Col] = true;
          break;
        }
      }
    }
    for (unsigned Slot : seq<unsigned>(NumCols))
      Aligned[Slot][Lane] = Operands[SlotSrcCol[Slot]][Lane];
  }
  return Aligned;
}
 
/// Rank reassociated operand layouts by vectorizable load columns, other load
/// patterns, broadcast/constant columns, then fewer unique values per column.
/// Identity columns are scored by the opcode of the emitted combines.
static std::tuple<unsigned, unsigned, unsigned, int>
getReassocColumnsQuality(ArrayRef<BoUpSLP::ValueList> Columns, const BoUpSLP &R,
                         unsigned Opcode) {
  unsigned NumConsecutiveLoadCols = 0;
  unsigned NumOtherVecLoadCols = 0;
  unsigned NumBroadcastOrConstCols = 0;
  int NumUniqueValues = 0;
  const unsigned CombineOpcode = getReassocCombineOpcode(Opcode);
  for (ArrayRef<Value *> Col : Columns) {
    if (all_of(Col, IsaPred<Constant>)) {
      ++NumBroadcastOrConstCols;
      continue;
    }
    // Identity constants are free beyond the base they are inserted into.
    SmallPtrSet<Value *, 8> UniqueValues;
    for (Value *V : Col)
      if (!isBinOpIdentityConstant(V, CombineOpcode))
        UniqueValues.insert(V);
    NumUniqueValues += UniqueValues.size();
    if (UniqueValues.size() <= 1) {
      ++NumBroadcastOrConstCols;
      continue;
    }
    if (all_of(Col, IsaPred<LoadInst>)) {
      BoUpSLP::OrdersType Order;
      SmallVector<Value *> PointerOps;
      BoUpSLP::StridedPtrInfo SPtrInfo;
      BoUpSLP::LoadsState LS =
          R.canVectorizeLoads(Col, Col.front(), Order, PointerOps, SPtrInfo);
      if (LS == BoUpSLP::LoadsState::Vectorize)
        ++NumConsecutiveLoadCols;
      else if (LS != BoUpSLP::LoadsState::Gather)
        ++NumOtherVecLoadCols;
    }
  }
  // Greater is better; negate unique count as a tie-breaker.
  return std::make_tuple(NumConsecutiveLoadCols, NumOtherVecLoadCols,
                         NumBroadcastOrConstCols, -NumUniqueValues);
}
 
void BoUpSLP::buildTreeRec(ArrayRef<Value *> VLRef, unsigned Depth,
                           const EdgeInfo &UserTreeIdx,
                           unsigned InterleaveFactor) {
  assert((allConstant(VLRef) || allSameType(VLRef)) && "Invalid types!");
 
  SmallVector<int> ReuseShuffleIndices;
  SmallVector<Value *> VL(VLRef);
 
  // Tries to build split node.
  auto TrySplitNode = [&](const InstructionsState &LocalState) {
    SmallVector<Value *> Op1, Op2;
    OrdersType ReorderIndices;
    if (!canBuildSplitNode(VL, LocalState, Op1, Op2, ReorderIndices))
      return false;
 
    auto Invalid = ScheduleBundle::invalid();
    auto *TE = newTreeEntry(VL, TreeEntry::SplitVectorize, Invalid, LocalState,
                            UserTreeIdx, {}, ReorderIndices);
    LLVM_DEBUG(dbgs() << "SLP: split alternate node.\n"; TE->dump());
    auto AddNode = [&](ArrayRef<Value *> Op, unsigned Idx) {
      InstructionsState S = getSameOpcode(Op, *TLI);
      if (S && (isa<LoadInst>(S.getMainOp()) ||
                getSameValuesTreeEntry(S.getMainOp(), Op, /*SameVF=*/true))) {
        // Build gather node for loads, they will be gathered later.
        TE->CombinedEntriesWithIndices.emplace_back(VectorizableTree.size(),
                                                    Idx == 0 ? 0 : Op1.size());
        (void)newTreeEntry(Op, TreeEntry::NeedToGather, Invalid, S, {TE, Idx});
      } else {
        TE->CombinedEntriesWithIndices.emplace_back(VectorizableTree.size(),
                                                    Idx == 0 ? 0 : Op1.size());
        buildTreeRec(Op, Depth, {TE, Idx});
      }
    };
    AddNode(Op1, 0);
    AddNode(Op2, 1);
    return true;
  };
 
  auto AreOnlyConstsWithPHIs = [](ArrayRef<Value *> VL) {
    bool AreConsts = false;
    for (Value *V : VL) {
      if (isa<PoisonValue>(V))
        continue;
      if (isa<Constant>(V)) {
        AreConsts = true;
        continue;
      }
      if (!isa<PHINode>(V))
        return false;
    }
    return AreConsts;
  };
  if (AreOnlyConstsWithPHIs(VL)) {
    LLVM_DEBUG(dbgs() << "SLP: Gathering due to all constants and PHIs.\n");
    newGatherTreeEntry(VL, InstructionsState::invalid(), UserTreeIdx);
    return;
  }
 
  ScalarsVectorizationLegality Legality =
      getScalarsVectorizationLegality(VL, Depth, UserTreeIdx);
  InstructionsState S = Legality.getInstructionsState();
  if (!Legality.isLegal()) {
    if (Legality.trySplitVectorize()) {
      auto [MainOp, AltOp] = getMainAltOpsNoStateVL(VL);
      // Last chance to try to vectorize alternate node.
      if (MainOp && AltOp && TrySplitNode(InstructionsState(MainOp, AltOp)))
        return;
    }
    if (Legality.tryToFindDuplicates())
      (void)tryToFindDuplicates(VL, ReuseShuffleIndices, *TTI, *TLI, S,
                                UserTreeIdx, *this);
 
    newGatherTreeEntry(VL, S, UserTreeIdx, ReuseShuffleIndices);
    return;
  }
 
  // FIXME: investigate if there are profitable cases for VL.size() <= 4.
  if (S.isAltShuffle() && TrySplitNode(S))
    return;
 
  // Cmp nodes with interchangeable lanes (e.g. x == 0 mixed with x <u C)
  // try the split by the original predicates, like the alternate nodes.
  if (S.getOpcode() == Instruction::ICmp && !S.isAltShuffle()) {
    auto SameOrSwapped = [](const ICmpInst *CI, CmpInst::Predicate P) {
      return CI->getPredicate() == P ||
             CI->getPredicate() == CmpInst::getSwappedPredicate(P);
    };
    ICmpInst *MainI = cast<ICmpInst>(*find_if(VL, IsaPred<ICmpInst>));
    auto *AltIt = find_if(VL, [&](Value *V) {
      auto *CI = dyn_cast<ICmpInst>(V);
      return CI && !SameOrSwapped(CI, MainI->getPredicate());
    });
    ICmpInst *AltI = AltIt == VL.end() ? nullptr : cast<ICmpInst>(*AltIt);
    if (AltI &&
        all_of(VL,
               [&](Value *V) {
                 auto *CI = dyn_cast<ICmpInst>(V);
                 return !CI || SameOrSwapped(CI, MainI->getPredicate()) ||
                        SameOrSwapped(CI, AltI->getPredicate());
               }) &&
        TrySplitNode(InstructionsState(MainI, AltI)))
      return;
  }
 
  // Check that every instruction appears once in this bundle.
  if (!tryToFindDuplicates(VL, ReuseShuffleIndices, *TTI, *TLI, S, UserTreeIdx,
                           *this, /*BuildGatherOnly=*/false)) {
    newGatherTreeEntry(VL, S, UserTreeIdx, ReuseShuffleIndices);
    return;
  }
 
  // Perform specific checks for each particular instruction kind.
  bool IsScatterVectorizeUserTE =
      UserTreeIdx.UserTE &&
      UserTreeIdx.UserTE->State == TreeEntry::ScatterVectorize;
  OrdersType CurrentOrder;
  SmallVector<Value *> PointerOps;
  StridedPtrInfo SPtrInfo;
  SmallVector<int> ExpandShuffleMask;
  TreeEntry::EntryState State = getScalarsVectorizationState(
      S, VL, IsScatterVectorizeUserTE, CurrentOrder, PointerOps, SPtrInfo,
      ExpandShuffleMask);
  if (State == TreeEntry::NeedToGather) {
    newGatherTreeEntry(VL, S, UserTreeIdx, ReuseShuffleIndices);
    return;
  }
 
  // Check the loop nest. We need to be sure we handle a single loop nest at a
  // time to avoid incorrect cost estimation because of the loop aware cost
  // model.
  if (VectorizableTree.empty()) {
    assert(CurrentLoopNest.empty() && "Expected empty loop nest");
    // Process the first node? Initial fill of the loop nest.
    BasicBlock *Parent = S.getMainOp()->getParent();
    if (const Loop *L = LI->getLoopFor(Parent)) {
      L = findInnermostNonInvariantLoop(L, VL);
      if (L)
        CurrentLoopNest.assign(getLoopNest(L));
    }
  } else if (!UserTreeIdx ||
             UserTreeIdx.UserTE->State == TreeEntry::SplitVectorize ||
             UserTreeIdx.UserTE->isGather() ||
             UserTreeIdx.UserTE->getMainOp()->getParent() !=
                 S.getMainOp()->getParent()) {
    BasicBlock *Parent = S.getMainOp()->getParent();
    if (const Loop *L = LI->getLoopFor(Parent)) {
      // Check that the new loop nest shares the same outer structure as the
      // tree's current loop nest. Completely disjoint nests (different
      // outermost loops) are forced to gather because their scales cannot be
      // meaningfully combined. Sibling inner loops (inside a common outer
      // loop or outside any loops at all) are allowed: the cost model scales
      // each entry by its own loop via getScaleToLoopIterations(), so a tree
      // that spans sibling inner loops (e.g. a PHI at their merge block) can
      // still be costed correctly. Contract CurrentLoopNest to the longest
      // common prefix with the new entry's nest so subsequent entries in yet
      // another sibling can also be admitted.
      L = findInnermostNonInvariantLoop(L, VL);
      if (L) {
        SmallVector<const Loop *> NewLoopNest(getLoopNest(L));
        unsigned CommonLen = 0;
        for (const auto [L1, L2] : zip(CurrentLoopNest, NewLoopNest)) {
          if (L1 != L2)
            break;
          ++CommonLen;
        }
        auto ValidateMergedBTCs = [&](unsigned StartDepth) -> bool {
          unsigned EndDepth =
              std::min<unsigned>(NewLoopNest.size(), MergedLoopBTCs.size());
          for (unsigned D = StartDepth; D < EndDepth; ++D) {
            const SCEV *Constraint = MergedLoopBTCs[D];
            if (!Constraint)
              continue;
            const SCEV *NewBTC = SE->getBackedgeTakenCount(NewLoopNest[D]);
            if (isa<SCEVCouldNotCompute>(NewBTC) || NewBTC != Constraint)
              return false;
          }
          return true;
        };
        auto BailOutToGather = [&]() {
          LLVM_DEBUG(dbgs()
                     << "SLP: Sibling loops have different trip counts.\n");
          newGatherTreeEntry(VL, S, UserTreeIdx, ReuseShuffleIndices);
        };
        if (CurrentLoopNest.empty()) {
          if (!ValidateMergedBTCs(0)) {
            BailOutToGather();
            return;
          }
          CurrentLoopNest.assign(NewLoopNest);
        } else if (CommonLen < CurrentLoopNest.size() &&
                   CommonLen < NewLoopNest.size()) {
          // Divergence below the common prefix: the tree now spans sibling
          // loops at depth CommonLen. Admitting them into one tree makes
          // the profitability decision JOINT across both siblings, so a
          // very hot sibling could otherwise let an unprofitable cold
          // sibling ride along "for free" (per-entry scaling of the cold
          // sibling's entries would be dwarfed by the hot one). Require
          // SCEV-proven equal backedge-taken counts for the diverging
          // siblings before joining; otherwise force gather.
          const Loop *SibA = CurrentLoopNest[CommonLen];
          const Loop *SibB = NewLoopNest[CommonLen];
          const SCEV *BecA = SE->getBackedgeTakenCount(SibA);
          const SCEV *BecB = SE->getBackedgeTakenCount(SibB);
          if (isa<SCEVCouldNotCompute>(BecA) || BecA != BecB) {
            BailOutToGather();
            return;
          }
          if (!ValidateMergedBTCs(CommonLen + 1)) {
            BailOutToGather();
            return;
          }
          if (MergedLoopBTCs.size() <= CommonLen)
            MergedLoopBTCs.resize(CommonLen + 1, nullptr);
          MergedLoopBTCs[CommonLen] = BecA;
          CurrentLoopNest.truncate(CommonLen);
        } else if (NewLoopNest.size() > CurrentLoopNest.size()) {
          if (!ValidateMergedBTCs(CurrentLoopNest.size())) {
            BailOutToGather();
            return;
          }
          CurrentLoopNest.append(
              std::next(NewLoopNest.begin(), CurrentLoopNest.size()),
              NewLoopNest.end());
        }
        // Otherwise NewLoopNest is a prefix of CurrentLoopNest: keep as-is.
      }
    }
  }
 
  Instruction *VL0 = S.getMainOp();
  BasicBlock *BB = VL0->getParent();
  auto &BSRef = BlocksSchedules[BB];
  if (!BSRef)
    BSRef = std::make_unique<BlockScheduling>(BB);
 
  BlockScheduling &BS = *BSRef;
 
  SetVector<Value *> UniqueValues(llvm::from_range, VL);
  std::optional<ScheduleBundle *> BundlePtr =
      BS.tryScheduleBundle(UniqueValues.getArrayRef(), this, S, UserTreeIdx);
#ifdef EXPENSIVE_CHECKS
  // Make sure we didn't break any internal invariants
  BS.verify();
#endif
  if (!BundlePtr || (*BundlePtr && !*BundlePtr.value())) {
    LLVM_DEBUG(dbgs() << "SLP: We are not able to schedule this bundle!\n");
    // Last chance to try to vectorize alternate node.
    if (S.isAltShuffle() && ReuseShuffleIndices.empty() && TrySplitNode(S))
      return;
    // Last chance to try to vectorize copyable node.
    if (S.areInstructionsWithCopyableElements() &&
        ReuseShuffleIndices.empty()) {
      InstructionsState AltS = getSameOpcode(VL, *TLI);
      if (AltS && AltS.isAltShuffle() && TrySplitNode(AltS))
        return;
    }
    newGatherTreeEntry(VL, S, UserTreeIdx, ReuseShuffleIndices);
    NonScheduledFirst.insert(VL.front());
    if (S.getOpcode() == Instruction::Load &&
        BS.ScheduleRegionSize < BS.ScheduleRegionSizeLimit)
      registerNonVectorizableLoads(ArrayRef(VL));
    return;
  }
  InstructionsCompatibilityAnalysis Analysis(*DT, *DL, *TTI, *TLI);
  SmallVector<ValueList> Operands = Analysis.buildOperands(S, VL);
  // Flatten associative binary chains into operand columns. Only the peeled
  // chain links are required to be single-use (they are erased); the root
  // being flattened may have other uses. Skip lanes that are neither chain
  // links nor copyable identity leaves. Alternate nodes flatten too, but
  // each lane keeps its own opcode on every combine level. Restricted to
  // BinaryOperator: isAssociative() is also true for associative intrinsics
  // (e.g. smax/smin/umax/umin), which are CallInst, not BinaryOperator, and
  // are not supported by the copyable-identity machinery used below
  // (ConstantExpr::getBinOpIdentity, isSupportedOpcode,
  // isBinOpIdentityConstant).
  SmallVector<Value *> ReassocScalars;
  // Sign of each flattened operand column (a subtracted leaf is negated).
  SmallBitVector NegatedColumns;
  // Per-lane subtract markers for flattened alternate nodes.
  SmallBitVector SubLanes;
  // Cached below (when the peel is kept) so the reorder step further down
  // does not need to redo the aligning/scoring work.
  SmallVector<ValueList> ReassocAlignedOperands;
  // Snapshot of the pre-flatten operand columns, used by both revert points.
  SmallVector<ValueList> NaturalTwoColumns;
  std::tuple<unsigned, unsigned, unsigned, int> ReassocPeeledQuality;
  if (VectorizeReassociatedOps && Operands.size() == 2 &&
      all_of(VL, [&](Value *V) {
        if (!S.isAltShuffle() && S.isCopyableElement(V))
          return true;
        auto *I = dyn_cast<BinaryOperator>(V);
        return I && isReassocChainLink(I);
      })) {
    NaturalTwoColumns = Operands;
    if (S.isAltShuffle()) {
      if (SmallVector<SmallVector<Value *>> Flattened =
              scanAltAssociativeOperands(S, *TLI, VL, Operands[0], Operands[1],
                                         ReassocScalars, SubLanes);
          !Flattened.empty()) {
        Operands.clear();
        for (auto &Col : Flattened)
          Operands.emplace_back(std::move(Col));
      }
    } else {
      scanAssociativeOperands(S, *DT, *DL, *TTI, *TLI, *this, Operands,
                              NegatedColumns, ReassocScalars);
    }
    // Drop flattening unless realigning improves load or broadcast column
    // structure; an unimproved peel ties and reverts to natural columns.
    if (!ReassocScalars.empty()) {
      auto IsNegated = [&](unsigned Lane, unsigned Col) {
        return S.isAltShuffle() ? Col > 0 && SubLanes.test(Lane)
                                : NegatedColumns.test(Col);
      };
      ReassocAlignedOperands =
          alignReassociatedOperandsByKey(Operands, IsNegated, *TLI);
      // Alternate nodes compare against the natural two-column form rather
      // than the raw peel: the natural form pays one lane-select shuffle per
      // level, so the flatten is worth keeping whenever the realigned
      // columns improve on the natural ones.
      ReassocPeeledQuality = getReassocColumnsQuality(
          S.isAltShuffle() ? NaturalTwoColumns : Operands, *this,
          S.getOpcode());
      // The unique-value count (4th field) is only a tie-break for the
      // later reorder-or-not decision, not for this one.
      auto DropUniqueCount = [](const auto &Quality) {
        return std::make_tuple(std::get<0>(Quality), std::get<1>(Quality),
                               std::get<2>(Quality));
      };
      if (DropUniqueCount(getReassocColumnsQuality(ReassocAlignedOperands,
                                                   *this, S.getOpcode())) <=
          DropUniqueCount(ReassocPeeledQuality)) {
        Operands = std::move(NaturalTwoColumns);
        ReassocScalars.clear();
      }
    }
  }
  // Registers peeled chain links on the node so they are erased with it and
  // their scheduling deps are released as reassociated operands.
  auto RegisterReassocScalars = [&](TreeEntry *TE) {
    for (Value *V : ReassocScalars) {
      TE->addReassocScalar(V);
      SmallVectorImpl<const TreeEntry *> &Owners =
          ReassocScalarToTreeEntries.try_emplace(V).first->second;
      if (!is_contained(Owners, TE))
        Owners.push_back(TE);
    }
  };
  // A value routed into several columns is consumed by several combines;
  // that distorts both the cost and the vector structure (duplicated loads,
  // partial masked columns), so prefer the natural two-column shape.
  auto HasDupColumnValues = [&]() {
    SmallPtrSet<const Value *, 16> ColumnValues;
    return any_of(Operands, [&](const ValueList &Col) {
      return any_of(Col, [&](const Value *V) {
        return !isa<Constant>(V) && !ColumnValues.insert(V).second;
      });
    });
  };
  ScheduleBundle Empty;
  ScheduleBundle &Bundle = BundlePtr.value() ? *BundlePtr.value() : Empty;
  LLVM_DEBUG(dbgs() << "SLP: We are able to schedule this bundle.\n");
 
  unsigned ShuffleOrOp =
      S.isAltShuffle() ? (unsigned)Instruction::ShuffleVector : S.getOpcode();
  auto CreateOperandNodes = [&](TreeEntry *TE, const auto &Operands) {
    // Postpone PHI nodes creation
    SmallVector<unsigned> PHIOps;
    for (unsigned I : seq<unsigned>(Operands.size())) {
      ArrayRef<Value *> Op = Operands[I];
      if (Op.empty())
        continue;
      InstructionsState S = getSameOpcode(Op, *TLI);
      if ((!S || S.getOpcode() != Instruction::PHI) || S.isAltShuffle())
        buildTreeRec(Op, Depth + 1, {TE, I});
      else
        PHIOps.push_back(I);
    }
    for (unsigned I : PHIOps)
      buildTreeRec(Operands[I], Depth + 1, {TE, I});
  };
  switch (ShuffleOrOp) {
    case Instruction::PHI: {
      TreeEntry *TE =
          newTreeEntry(VL, Bundle, S, UserTreeIdx, ReuseShuffleIndices);
      LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (PHINode).\n";
                 TE->dump());
 
      TE->setOperands(Operands);
      CreateOperandNodes(TE, Operands);
      return;
    }
    case Instruction::ExtractValue:
    case Instruction::ExtractElement: {
      if (CurrentOrder.empty()) {
        LLVM_DEBUG(dbgs() << "SLP: Reusing or shuffling extract sequence.\n");
      } else {
        LLVM_DEBUG({
          dbgs() << "SLP: Reusing or shuffling of reordered extract sequence "
                    "with order";
          for (unsigned Idx : CurrentOrder)
            dbgs() << " " << Idx;
          dbgs() << "\n";
        });
        fixupOrderingIndices(CurrentOrder);
      }
      // Insert new order with initial value 0, if it does not exist,
      // otherwise return the iterator to the existing one.
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndices, CurrentOrder);
      LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry "
                           "(ExtractValueInst/ExtractElementInst).\n";
                 TE->dump());
      // This is a special case, as it does not gather, but at the same time
      // we are not extending buildTreeRec() towards the operands.
      TE->setOperands(Operands);
      if (ShuffleOrOp == Instruction::ExtractValue) {
        SmallVector<unsigned> Indices;
        SmallVector<Value *> Calls;
        if (checkEVsForVecCalls(VL, S, *TLI, Indices, Calls)) {
          TE->StructEVIndices = std::move(Indices);
          buildTreeRec(Operands.front(), Depth + 1, {TE, 0});
        }
      }
      return;
    }
    case Instruction::InsertValue:
    case Instruction::InsertElement: {
      assert(ReuseShuffleIndices.empty() && "All inserts should be unique");
 
      auto OrdCompare = [](const std::pair<int, int> &P1,
                           const std::pair<int, int> &P2) {
        return P1.first > P2.first;
      };
      PriorityQueue<std::pair<int, int>, SmallVector<std::pair<int, int>>,
                    decltype(OrdCompare)>
          Indices(OrdCompare);
      for (int I = 0, E = VL.size(); I < E; ++I) {
        unsigned Idx = *getElementIndex(VL[I]);
        Indices.emplace(Idx, I);
      }
      OrdersType CurrentOrder(VL.size(), VL.size());
      bool IsIdentity = true;
      for (int I = 0, E = VL.size(); I < E; ++I) {
        CurrentOrder[Indices.top().second] = I;
        IsIdentity &= Indices.top().second == I;
        Indices.pop();
      }
      if (IsIdentity)
        CurrentOrder.clear();
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   {}, CurrentOrder);
      LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (InsertElementInst).\n";
                 TE->dump());
 
      TE->setOperands(Operands);
      buildTreeRec(TE->getOperand(1), Depth + 1, {TE, 1});
      return;
    }
    case Instruction::Load: {
      // Check that a vectorized load would load the same memory as a scalar
      // load. For example, we don't want to vectorize loads that are smaller
      // than 8-bit. Even though we have a packed struct {<i2, i2, i2, i2>} LLVM
      // treats loading/storing it as an i8 struct. If we vectorize loads/stores
      // from such a struct, we read/write packed bits disagreeing with the
      // unvectorized version.
      TreeEntry *TE = nullptr;
      fixupOrderingIndices(CurrentOrder);
      switch (State) {
      case TreeEntry::Vectorize:
        TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                          ReuseShuffleIndices, CurrentOrder, InterleaveFactor);
        if (CurrentOrder.empty())
          LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (LoadInst).\n";
                     TE->dump());
        else
          LLVM_DEBUG(dbgs()
                         << "SLP: added a new TreeEntry (jumbled LoadInst).\n";
                     TE->dump());
        break;
      case TreeEntry::CompressVectorize:
        // Vectorizing non-consecutive loads with (masked)load + compress.
        TE = newTreeEntry(VL, TreeEntry::CompressVectorize, Bundle, S,
                          UserTreeIdx, ReuseShuffleIndices, CurrentOrder);
        LLVM_DEBUG(
            dbgs()
                << "SLP: added a new TreeEntry (masked LoadInst + compress).\n";
            TE->dump());
        break;
      case TreeEntry::StridedVectorize:
        // Vectorizing non-consecutive loads with `llvm.masked.gather`.
        TE = newTreeEntry(VL, TreeEntry::StridedVectorize, Bundle, S,
                          UserTreeIdx, ReuseShuffleIndices, CurrentOrder);
        TreeEntryToStridedPtrInfoMap[TE] = SPtrInfo;
        LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (strided LoadInst).\n";
                   TE->dump());
        break;
      case TreeEntry::ScatterVectorize:
        // Vectorizing non-consecutive loads with `llvm.masked.gather`.
        TE = newTreeEntry(VL, TreeEntry::ScatterVectorize, Bundle, S,
                          UserTreeIdx, ReuseShuffleIndices);
        LLVM_DEBUG(
            dbgs()
                << "SLP: added a new TreeEntry (non-consecutive LoadInst).\n";
            TE->dump());
        break;
      case TreeEntry::BlendedLoadVectorize: {
        // Two masked loads from the candidate bases, blended by the per-lane
        // select conditions. The conditions are added as the operand node, so
        // the blend mask is vectorized through the tree.
        TE = newTreeEntry(VL, TreeEntry::BlendedLoadVectorize, Bundle, S,
                          UserTreeIdx, ReuseShuffleIndices);
        LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (blended LoadInst).\n";
                   TE->dump());
        Value *TrueBase = nullptr;
        Value *FalseBase = nullptr;
        SmallVector<Value *> Conditions;
        [[maybe_unused]] bool Found = isSelectedBaseLoad(
            VL0->getType(), PointerOps, *DL, TrueBase, FalseBase, Conditions);
        assert(Found && "Expected a valid blended-load pattern.");
        Operands.assign(1, ValueList(Conditions.begin(), Conditions.end()));
        TE->setOperands(Operands);
        buildTreeRec(TE->getOperand(0), Depth + 1, {TE, 0});
        return;
      }
      case TreeEntry::ExpandVectorize:
      case TreeEntry::CombinedVectorize:
      case TreeEntry::SplitVectorize:
      case TreeEntry::NeedToGather:
        llvm_unreachable("Unexpected loads state.");
      }
      if (!CurrentOrder.empty() && State != TreeEntry::ScatterVectorize) {
        assert(Operands.size() == 1 && "Expected a single operand only");
        SmallVector<int> Mask;
        inversePermutation(CurrentOrder, Mask);
        reorderScalars(Operands.front(), Mask);
      }
      TE->setOperands(Operands);
      if (State == TreeEntry::ScatterVectorize)
        buildTreeRec(PointerOps, Depth + 1, {TE, 0});
      return;
    }
    case Instruction::ZExt:
    case Instruction::SExt:
    case Instruction::FPToUI:
    case Instruction::FPToSI:
    case Instruction::FPExt:
    case Instruction::PtrToInt:
    case Instruction::PtrToAddr:
    case Instruction::IntToPtr:
    case Instruction::SIToFP:
    case Instruction::UIToFP:
    case Instruction::Trunc:
    case Instruction::FPTrunc:
    case Instruction::BitCast: {
      auto [PrevMaxBW, PrevMinBW] = CastMaxMinBWSizes.value_or(
          std::make_pair(std::numeric_limits<unsigned>::min(),
                         std::numeric_limits<unsigned>::max()));
      if (ShuffleOrOp == Instruction::ZExt ||
          ShuffleOrOp == Instruction::SExt) {
        CastMaxMinBWSizes = std::make_pair(
            std::max<unsigned>(DL->getTypeSizeInBits(VL0->getType()),
                               PrevMaxBW),
            std::min<unsigned>(
                DL->getTypeSizeInBits(VL0->getOperand(0)->getType()),
                PrevMinBW));
      } else if (ShuffleOrOp == Instruction::Trunc) {
        CastMaxMinBWSizes = std::make_pair(
            std::max<unsigned>(
                DL->getTypeSizeInBits(VL0->getOperand(0)->getType()),
                PrevMaxBW),
            std::min<unsigned>(DL->getTypeSizeInBits(VL0->getType()),
                               PrevMinBW));
      }
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndices);
      LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (CastInst).\n";
                 TE->dump());
 
      TE->setOperands(Operands);
      for (unsigned I : seq<unsigned>(VL0->getNumOperands()))
        buildTreeRec(TE->getOperand(I), Depth, {TE, I});
      if (ShuffleOrOp == Instruction::Trunc) {
        ExtraBitWidthNodes.insert(getOperandEntry(TE, 0)->Idx);
      } else if (ShuffleOrOp == Instruction::SIToFP ||
                 ShuffleOrOp == Instruction::UIToFP) {
        unsigned NumSignBits =
            ComputeNumSignBits(VL0->getOperand(0), *DL, AC, nullptr, DT);
        if (auto *OpI = dyn_cast<Instruction>(VL0->getOperand(0))) {
          APInt Mask = DB->getDemandedBits(OpI);
          NumSignBits = std::max(NumSignBits, Mask.countl_zero());
        }
        if (NumSignBits * 2 >=
            DL->getTypeSizeInBits(VL0->getOperand(0)->getType()))
          ExtraBitWidthNodes.insert(getOperandEntry(TE, 0)->Idx);
      }
      return;
    }
    case Instruction::ICmp:
    case Instruction::FCmp: {
      // Check that all of the compares have the same predicate.
      CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate();
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndices);
      LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (CmpInst).\n";
                 TE->dump());
 
      VLOperands Ops(VL, Operands, S, *this);
      if (cast<CmpInst>(VL0)->isCommutative()) {
        // Commutative predicate - collect + sort operands of the instructions
        // so that each side is more likely to have the same opcode.
        assert(P0 == CmpInst::getSwappedPredicate(P0) &&
               "Commutative Predicate mismatch");
        Ops.reorder();
        Operands.front() = Ops.getVL(0);
        Operands.back() = Ops.getVL(1);
      } else {
        // Collect operands - commute if it uses the swapped predicate.
        // Lanes interchangeable with P0 (e.g. x == 0 in an x <u C bundle)
        // already have their operands adjusted, no need to commute them.
        for (auto [Idx, V] : enumerate(VL)) {
          if (isa<PoisonValue>(V))
            continue;
          auto *Cmp = cast<CmpInst>(V);
          if (Cmp->getPredicate() != P0 &&
              !CmpSamePredicateHelper::canConvertTo(Cmp, P0))
            std::swap(Operands.front()[Idx], Operands.back()[Idx]);
        }
      }
      TE->setOperands(Operands);
      buildTreeRec(Operands.front(), Depth, {TE, 0});
      buildTreeRec(Operands.back(), Depth, {TE, 1});
      if (ShuffleOrOp == Instruction::ICmp) {
        unsigned NumSignBits0 =
            ComputeNumSignBits(VL0->getOperand(0), *DL, AC, nullptr, DT);
        if (NumSignBits0 * 2 >=
            DL->getTypeSizeInBits(VL0->getOperand(0)->getType()))
          ExtraBitWidthNodes.insert(getOperandEntry(TE, 0)->Idx);
        unsigned NumSignBits1 =
            ComputeNumSignBits(VL0->getOperand(1), *DL, AC, nullptr, DT);
        if (NumSignBits1 * 2 >=
            DL->getTypeSizeInBits(VL0->getOperand(1)->getType()))
          ExtraBitWidthNodes.insert(getOperandEntry(TE, 1)->Idx);
      }
      return;
    }
    case Instruction::Select:
    case Instruction::FNeg:
    case Instruction::Add:
    case Instruction::FAdd:
    case Instruction::Sub:
    case Instruction::FSub:
    case Instruction::Mul:
    case Instruction::FMul:
    case Instruction::UDiv:
    case Instruction::SDiv:
    case Instruction::FDiv:
    case Instruction::URem:
    case Instruction::SRem:
    case Instruction::FRem:
    case Instruction::Shl:
    case Instruction::LShr:
    case Instruction::AShr:
    case Instruction::And:
    case Instruction::Or:
    case Instruction::Xor:
    case Instruction::Freeze: {
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndices);
      LLVM_DEBUG(
          dbgs() << "SLP: added a new TreeEntry "
                    "(SelectInst/UnaryOperator/BinaryOperator/FreezeInst).\n";
          TE->dump());
 
      if (!ReassocScalars.empty()) {
        // Realign columns by value family (cached above as
        // ReassocAlignedOperands), then let VLOperands polish the order,
        // keeping the polish only if the column-quality score does not regress
        // vs the peeled order. That polish is quadratic in the column count, so
        // past ReassocReorderColumnLimit keep the realigned seed, which already
        // groups the vectorizable columns. The polish swaps values between
        // columns per lane, which must not mix added and subtracted leaves in
        // one column, so signed nodes polish each sign group separately.
        if (Operands.size() <= ReassocReorderColumnLimit) {
          SmallVector<ValueList> Reordered = ReassocAlignedOperands;
          SmallVector<unsigned> GroupIds[2];
          for (unsigned I : seq<unsigned>(NegatedColumns.size()))
            GroupIds[NegatedColumns[I]].push_back(I);
          for (ArrayRef<unsigned> Group : GroupIds) {
            if (Group.size() <= 1)
              continue;
            SmallVector<ValueList> GroupCols = map_to_vector(
                Group, [&](unsigned I) { return ReassocAlignedOperands[I]; });
            VLOperands Ops(GroupCols, VL0->getParent(), *this);
            Ops.reorder();
            for (unsigned Pos : seq<unsigned>(Group.size()))
              Reordered[Group[Pos]] = Ops.getVL(Pos);
          }
          if (getReassocColumnsQuality(Reordered, *this, S.getOpcode()) >=
              ReassocPeeledQuality)
            Operands = std::move(Reordered);
          else
            LLVM_DEBUG(dbgs() << "SLP: reassociation reorder loses column "
                                 "structure, keeping the natural operand "
                                 "order.\n");
        } else {
          Operands = std::move(ReassocAlignedOperands);
        }
        if (HasDupColumnValues()) {
          Operands = std::move(NaturalTwoColumns);
          ReassocScalars.clear();
        }
      }
      if (!ReassocScalars.empty())
        RegisterReassocScalars(TE);
      else if (isa<BinaryOperator>(VL0) && isCommutative(VL0)) {
        VLOperands Ops(VL, Operands, S, *this);
        Ops.reorder();
        Operands[0] = Ops.getVL(0);
        Operands[1] = Ops.getVL(1);
      }
      TE->setOperands(Operands);
      if (!ReassocScalars.empty() && NegatedColumns.any())
        TE->setReassocNegatedOps(NegatedColumns);
      for (unsigned I : seq<unsigned>(TE->getNumOperands()))
        buildTreeRec(TE->getOperand(I), Depth + 1, {TE, I});
      return;
    }
    case Instruction::GetElementPtr: {
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndices);
      LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (GetElementPtrInst).\n";
                 TE->dump());
      TE->setOperands(Operands);
 
      for (unsigned I = 0, Ops = Operands.size(); I < Ops; ++I)
        buildTreeRec(Operands[I], Depth + 1, {TE, I});
      return;
    }
    case Instruction::Store: {
      assert(CurrentOrder.empty() &&
             "Expected ordered store during tree building");
      if (State == TreeEntry::StridedVectorize) {
        TreeEntry *TE =
            newTreeEntry(VL, TreeEntry::StridedVectorize, Bundle, S,
                         UserTreeIdx, ReuseShuffleIndices, CurrentOrder);
        TreeEntryToStridedPtrInfoMap[TE] = SPtrInfo;
        LLVM_DEBUG(
            dbgs() << "SLP: added a new TreeEntry (strided StoreInst).\n";
            TE->dump());
        TE->setOperands(Operands);
        buildTreeRec(TE->getOperand(0), Depth + 1, {TE, 0});
        return;
      }
      if (State == TreeEntry::ExpandVectorize) {
        assert(ReuseShuffleIndices.empty() &&
               "Expected no reuse shuffle for an expanded masked store.");
        TreeEntry *TE =
            newTreeEntry(VL, TreeEntry::ExpandVectorize, Bundle, S, UserTreeIdx,
                         ExpandShuffleMask, CurrentOrder);
        TreeEntryToStridedPtrInfoMap[TE] = SPtrInfo;
        LLVM_DEBUG(
            dbgs()
                << "SLP: added a new TreeEntry (expanded masked StoreInst).\n";
            TE->dump());
        TE->setOperands(Operands);
        buildTreeRec(TE->getOperand(0), Depth + 1, {TE, 0});
        return;
      }
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndices, CurrentOrder);
      LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (StoreInst).\n";
                 TE->dump());
      TE->setOperands(Operands);
      buildTreeRec(TE->getOperand(0), Depth + 1, {TE, 0});
      return;
    }
    case Instruction::Call: {
      // Check if the calls are all to the same vectorizable intrinsic or
      // library function.
      CallInst *CI = cast<CallInst>(VL0);
      Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
 
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndices);
      LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (CallInst).\n";
                 TE->dump());
      if (isCommutative(VL0)) {
        VLOperands Ops(VL, Operands, S, *this);
        Ops.reorder();
        Operands[0] = Ops.getVL(0);
        Operands[1] = Ops.getVL(1);
      }
      TE->setOperands(Operands);
      for (unsigned I : seq<unsigned>(CI->arg_size())) {
        // For scalar operands no need to create an entry since no need to
        // vectorize it.
        if (isVectorIntrinsicWithScalarOpAtArg(ID, I, TTI))
          continue;
        buildTreeRec(TE->getOperand(I), Depth + 1, {TE, I});
      }
      return;
    }
    case Instruction::ShuffleVector: {
      TreeEntry *TE = newTreeEntry(VL, Bundle /*vectorized*/, S, UserTreeIdx,
                                   ReuseShuffleIndices);
      if (S.isAltShuffle()) {
        LLVM_DEBUG(dbgs() << "SLP: added a new TreeEntry (isAltShuffle).\n";
                   TE->dump());
      } else {
        assert(SLPReVec && "Only supported by REVEC.");
        LLVM_DEBUG(
            dbgs() << "SLP: added a new TreeEntry (ShuffleVectorInst).\n";
            TE->dump());
      }
 
      // Reorder operands if reordering would enable vectorization.
      auto *CI = dyn_cast<CmpInst>(VL0);
      if (CI && any_of(VL, [](Value *V) {
            return !isa<PoisonValue>(V) && !cast<CmpInst>(V)->isCommutative();
          })) {
        auto *MainCI = cast<CmpInst>(S.getMainOp());
        auto *AltCI = cast<CmpInst>(S.getAltOp());
        CmpInst::Predicate MainP = MainCI->getPredicate();
        CmpInst::Predicate AltP = AltCI->getPredicate();
        assert(MainP != AltP &&
               "Expected different main/alternate predicates.");
        // Collect operands - commute if it uses the swapped predicate or
        // alternate operation.
        for (auto [Idx, V] : enumerate(VL)) {
          if (isa<PoisonValue>(V))
            continue;
          auto *Cmp = cast<CmpInst>(V);
 
          if (isAlternateInstruction(Cmp, MainCI, AltCI, *TLI)) {
            if (AltP == CmpInst::getSwappedPredicate(Cmp->getPredicate()))
              std::swap(Operands.front()[Idx], Operands.back()[Idx]);
          } else {
            if (MainP == CmpInst::getSwappedPredicate(Cmp->getPredicate()))
              std::swap(Operands.front()[Idx], Operands.back()[Idx]);
          }
        }
        TE->setOperands(Operands);
        buildTreeRec(Operands.front(), Depth + 1, {TE, 0});
        buildTreeRec(Operands.back(), Depth + 1, {TE, 1});
        return;
      }
 
      if (!ReassocScalars.empty()) {
        // Peeled alternate chains take the realigned columns; the operand
        // reorder below does not preserve the per-lane sign constraints.
        Operands = std::move(ReassocAlignedOperands);
        if (HasDupColumnValues()) {
          Operands = std::move(NaturalTwoColumns);
          ReassocScalars.clear();
        }
      }
      if (ReassocScalars.empty() && (isa<BinaryOperator>(VL0) || CI)) {
        VLOperands Ops(VL, Operands, S, *this);
        Ops.reorder();
        Operands[0] = Ops.getVL(0);
        Operands[1] = Ops.getVL(1);
      }
      if (!ReassocScalars.empty())
        RegisterReassocScalars(TE);
      TE->setOperands(Operands);
      for (unsigned I : seq<unsigned>(TE->getNumOperands()))
        buildTreeRec(TE->getOperand(I), Depth + 1, {TE, I});
      return;
    }
    default:
      break;
  }
  llvm_unreachable("Unexpected vectorization of the instructions.");
}
 
unsigned BoUpSLP::canMapToVector(Type *T) const {
  unsigned N = 1;
  Type *EltTy = T;
 
  while (isa<StructType, ArrayType, FixedVectorType>(EltTy)) {
    if (EltTy->isEmptyTy())
      return 0;
    if (auto *ST = dyn_cast<StructType>(EltTy)) {
      // Check that struct is homogeneous.
      for (const auto *Ty : ST->elements())
        if (Ty != *ST->element_begin())
          return 0;
      N *= ST->getNumElements();
      EltTy = *ST->element_begin();
    } else if (auto *AT = dyn_cast<ArrayType>(EltTy)) {
      N *= AT->getNumElements();
      EltTy = AT->getElementType();
    } else {
      auto *VT = cast<FixedVectorType>(EltTy);
      N *= VT->getNumElements();
      EltTy = VT->getElementType();
    }
  }
 
  if (!isValidElementType(EltTy, SLPReVec))
    return 0;
  size_t VTSize = DL->getTypeStoreSizeInBits(getWidenedType(EltTy, N));
  if (VTSize < MinVecRegSize || VTSize > MaxVecRegSize ||
      VTSize != DL->getTypeStoreSizeInBits(T))
    return 0;
  return N;
}
 
bool BoUpSLP::canVectorStoreInsertValue(const TreeEntry *E) const {
  if (E->getOpcode() != Instruction::InsertValue ||
      E->State != TreeEntry::Vectorize)
    return false;
 
  bool HasStore = false;
  for (Value *V : E->Scalars) {
    auto *IV = dyn_cast<InsertValueInst>(V);
    if (!IV || isDeleted(IV))
      return false;
 
    bool HasExternalUser = false;
    for (User *U : IV->users()) {
      if (auto *UI = dyn_cast<Instruction>(U); UI && isDeleted(UI))
        continue;
      if (is_contained(E->Scalars, U))
        continue;
      HasExternalUser = true;
      // Only a plain (non-volatile, non-atomic) store may consume the vector
      // directly. Volatile/atomic stores must keep their original aggregate
      // type and access semantics, so fall back to the store + load roundtrip.
      auto *SI = dyn_cast<StoreInst>(U);
      if (!SI || !SI->isSimple())
        return false;
      HasStore = true;
    }
    if (HasExternalUser && !IV->hasOneUse())
      return false;
  }
  return HasStore;
}
 
FixedVectorType *BoUpSLP::getInsertBuildVectorSrcTy(const TreeEntry *E) const {
  Value *VL0 = E->getMainOp();
  if (E->getOpcode() == Instruction::InsertElement)
    return cast<FixedVectorType>(VL0->getType());
  assert(E->getOpcode() == Instruction::InsertValue &&
         "Expected InsertElement or InsertValue node.");
  unsigned MaxIdx = E->Scalars.size() - 1;
  for (Value *V : E->Scalars) {
    auto *I = dyn_cast<InsertValueInst>(V);
    if (!I)
      continue;
    MaxIdx = std::max(MaxIdx, I->getIndices().front());
  }
  return cast<FixedVectorType>(
      getWidenedType(getValueType(VL0, SLPReVec), MaxIdx + 1));
}
 
bool BoUpSLP::canReuseExtract(ArrayRef<Value *> VL,
                              SmallVectorImpl<unsigned> &CurrentOrder,
                              bool ResizeAllowed) const {
  const auto *It = find_if(VL, IsaPred<ExtractElementInst, ExtractValueInst>);
  assert(It != VL.end() && "Expected at least one extract instruction.");
  auto *E0 = cast<Instruction>(*It);
  // Check if all of the extracts come from the same vector and from the
  // correct offset.
  Value *Vec = E0->getOperand(0);
 
  CurrentOrder.clear();
 
  // We have to extract from a vector/aggregate with the same number of elements.
  unsigned NElts;
  if (E0->getOpcode() == Instruction::ExtractValue) {
    NElts = canMapToVector(Vec->getType());
    if (!NElts)
      return false;
    // Check if load can be rewritten as load of vector.
    LoadInst *LI = dyn_cast<LoadInst>(Vec);
    if (!LI || !LI->isSimple() || !LI->hasNUses(VL.size()))
      return false;
  } else {
    NElts = cast<FixedVectorType>(Vec->getType())->getNumElements();
  }
 
  unsigned E = VL.size();
  if (!ResizeAllowed && NElts != E)
    return false;
  SmallVector<int> Indices(E, PoisonMaskElem);
  unsigned MinIdx = NElts, MaxIdx = 0;
  for (auto [I, V] : enumerate(VL)) {
    // Non-extract lanes (copyable elements modeled on an extract main op, or
    // undefs) are treated as holes.
    if (!isa<ExtractElementInst, ExtractValueInst>(V))
      continue;
    auto *Inst = cast<Instruction>(V);
    if (Inst->getOperand(0) != Vec)
      return false;
    if (auto *EE = dyn_cast<ExtractElementInst>(Inst))
      if (isa<UndefValue>(EE->getIndexOperand()))
        continue;
    std::optional<unsigned> Idx = getExtractIndex(Inst);
    if (!Idx)
      return false;
    const unsigned ExtIdx = *Idx;
    if (ExtIdx >= NElts)
      continue;
    Indices[I] = ExtIdx;
    if (MinIdx > ExtIdx)
      MinIdx = ExtIdx;
    if (MaxIdx < ExtIdx)
      MaxIdx = ExtIdx;
  }
  if (MaxIdx - MinIdx + 1 > E)
    return false;
  if (MaxIdx + 1 <= E)
    MinIdx = 0;
 
  // Check that all of the indices extract from the correct offset.
  bool ShouldKeepOrder = true;
  // Assign to all items the initial value E + 1 so we can check if the extract
  // instruction index was used already.
  // Also, later we can check that all the indices are used and we have a
  // consecutive access in the extract instructions, by checking that no
  // element of CurrentOrder still has value E + 1.
  CurrentOrder.assign(E, E);
  for (unsigned I = 0; I < E; ++I) {
    if (Indices[I] == PoisonMaskElem)
      continue;
    const unsigned ExtIdx = Indices[I] - MinIdx;
    if (CurrentOrder[ExtIdx] != E) {
      CurrentOrder.clear();
      return false;
    }
    ShouldKeepOrder &= ExtIdx == I;
    CurrentOrder[ExtIdx] = I;
  }
  if (ShouldKeepOrder)
    CurrentOrder.clear();
 
  return ShouldKeepOrder;
}
 
bool BoUpSLP::areAllUsersVectorized(
    Instruction *I, const SmallDenseSet<Value *> *VectorizedVals) const {
  return (I->hasOneUse() && (!VectorizedVals || VectorizedVals->contains(I))) ||
         all_of(I->users(), [this](User *U) {
           return isVectorized(U) || isVectorLikeInstWithConstOps(U) ||
                  (isa<ExtractElementInst>(U) && MustGather.contains(U));
         });
}
 
static InstructionCost canConvertToFMA(ArrayRef<Value *> VL,
                                       const InstructionsState &S,
                                       DominatorTree &DT, const DataLayout &DL,
                                       TargetTransformInfo &TTI,
                                       const TargetLibraryInfo &TLI,
                                       const TTI::TargetCostKind CostKind);
 
uint64_t BoUpSLP::getNumScalarInsts(bool HasTreeLoop) {
  uint64_t Total = 0;
  for (const std::unique_ptr<TreeEntry> &Ptr : VectorizableTree) {
    const TreeEntry &TE = *Ptr;
    if (DeletedNodes.contains(&TE))
      continue;
    // CombinedVectorize entries (e.g. the fmul child of an FMulAdd, or the
    // cmp child of a MinMax select) are absorbed into the parent on both
    // scalar and vector sides. The backend fuses fadd+fmul -> fma and
    // select+cmp -> smin/smax even for scalar code, so skip to avoid
    // double-counting. Checked before the scale lookup below: some of
    // these entries started out as constant-only gather nodes and have no
    // valid instruction state to key the loop-nest scale off of.
    if (TE.State == TreeEntry::CombinedVectorize)
      continue;
    uint64_t Scale = getEntryEffectiveScale(TE);
    if (HasTreeLoop && Scale <= 1)
      continue;
    unsigned Count = 0;
    if (TE.isGather() || TransformedToGatherNodes.contains(&TE)) {
      // Count extractelement scalars in gathers - they exist in the scalar
      // code regardless of vectorization. ExtractElement instructions
      // become free when the vector input is used directly.
      for (Value *V : TE.Scalars)
        if (isa<ExtractElementInst>(V))
          ++Count;
      Total = SaturatingMultiplyAdd<uint64_t>(Count, Scale, Total);
      continue;
    }
    // Each vectorize entry represents a bundle of scalar instructions.
    // Count per-entry without cross-entry deduplication, since shared
    // scalars across entries still represent separate work in scalar code.
    for (Value *V : TE.Scalars) {
      if (!isa<Instruction>(V) ||
          (TE.hasCopyableElements() && TE.isCopyableElement(V)))
        continue;
      ++Count;
      // Calculate calls/divs/rems twice, they may cost higher, so better to
      // include their count twice to mimic slightly real cost here.
      auto *I = dyn_cast<Instruction>(V);
      if (I && (I->isIntDivRem() || I->isFPDivRem()))
        ++Count;
      if (auto *CI = dyn_cast<CallInst>(V)) {
        Intrinsic::ID BaseID = getVectorIntrinsicIDForCall(CI, TLI);
        if (!isTriviallyVectorizable(BaseID))
          ++Count;
      }
    }
    // Count peeled intermediates in the scalar baseline (not in Scalars).
    Count += TE.getReassocScalars().size();
    // Even when the whole node is not combined, individual scalar
    // instructions may be fused by the backend. Each fused pair (e.g.
    // fadd+fmul -> fma, select+cmp -> smin/smax) becomes a single scalar
    // instruction, absorbing the operand instruction. Subtract 1 for each
    // such match to avoid over-counting the scalar side.
    if (TE.CombinedOp == TreeEntry::NotCombinedOp && TE.hasState()) {
      unsigned Opcode = TE.getOpcode();
      if (Opcode == Instruction::Select) {
        for (Value *V : TE.Scalars) {
          if (TE.hasCopyableElements() && TE.isCopyableElement(V))
            continue;
          auto *SI = dyn_cast<SelectInst>(V);
          if (!SI)
            continue;
          auto [ID, _] = canConvertToMinOrMaxIntrinsic({V});
          if (ID != Intrinsic::not_intrinsic) {
            assert(Count > 0 && "Underflow in scalar inst count (minmax)");
            --Count;
          }
        }
      } else if (Opcode == Instruction::FAdd || Opcode == Instruction::FSub) {
        for (Value *V : TE.Scalars) {
          if (TE.hasCopyableElements() && TE.isCopyableElement(V))
            continue;
          auto *I = dyn_cast<Instruction>(V);
          if (!I || (TE.isAltShuffle() && I->getOpcode() != Instruction::FAdd &&
                     I->getOpcode() != Instruction::FSub))
            continue;
          if (canConvertToFMA(I, InstructionsState(I, I), *DT, *DL, *TTI, *TLI,
                              CostKind)
                  .isValid()) {
            assert(Count > 0 && "Underflow in scalar inst count (fma)");
            --Count;
          }
        }
      }
    }
    Total = SaturatingMultiplyAdd<uint64_t>(Count, Scale, Total);
  }
  return Total;
}
 
uint64_t BoUpSLP::getNumVectorInsts(bool HasTreeLoop) {
  uint64_t Total = 0;
  // Source vector -> max scale among the gather entries sharing it, so the
  // combined shufflevector is still weighted like an in-loop entry below.
  SmallDenseMap<Value *, uint64_t, 4> GatherExtractSourceVecs;
  for (const std::unique_ptr<TreeEntry> &Ptr : VectorizableTree) {
    const TreeEntry &TE = *Ptr;
    if (DeletedNodes.contains(&TE))
      continue;
    if (TE.State == TreeEntry::CombinedVectorize)
      continue;
    uint64_t Scale = getEntryEffectiveScale(TE);
    if (HasTreeLoop && Scale <= 1)
      continue;
    unsigned Count = 0;
    if (TE.isGather() || TransformedToGatherNodes.contains(&TE)) {
      if (TE.hasState()) {
        if (const TreeEntry *E =
                getSameValuesTreeEntry(TE.getMainOp(), TE.Scalars);
            E && E != &TE && E->getVectorFactor() == TE.getVectorFactor())
          continue;
        SmallVector<Value *> RevScalars(TE.Scalars.rbegin(), TE.Scalars.rend());
        if (const TreeEntry *E =
                getSameValuesTreeEntry(TE.getMainOp(), RevScalars);
            E && E->getVectorFactor() == TE.getVectorFactor()) {
          Total = SaturatingAdd(Total, Scale);
          continue;
        }
      }
      // ExtractElement gathers from the same source vector become a single
      // shufflevector. Collect source vectors globally across all gather
      // entries and count once at the end.
      if (all_of(TE.Scalars,
                 IsaPred<ExtractElementInst, UndefValue, Constant>)) {
        for (Value *V : TE.Scalars)
          if (auto *EE = dyn_cast<ExtractElementInst>(V)) {
            uint64_t &VecScale =
                GatherExtractSourceVecs.try_emplace(EE->getVectorOperand(), 0)
                    .first->second;
            VecScale = std::max(VecScale, Scale);
          }
      } else {
        for (Value *V : TE.Scalars) {
          if (!isConstant(V))
            ++Count;
        }
      }
      Total = SaturatingMultiplyAdd<uint64_t>(Count, Scale, Total);
      continue;
    }
    // InsertElement/ExtractElement vectorize entries don't produce real
    // vector instructions - InsertElement at root IS the result, and
    // ExtractElement entries reference the input vector directly.
    if (TE.getOpcode() == Instruction::InsertElement ||
        TE.getOpcode() == Instruction::ExtractElement)
      continue;
    if (TE.getOpcode() == Instruction::InsertValue) {
      // InsertValue materializes via store + load unless it is stored directly
      // as a vector.
      if (!canVectorStoreInsertValue(&TE))
        Count += 2;
      if (!TE.ReorderIndices.empty() || !TE.ReuseShuffleIndices.empty())
        ++Count;
      Total = SaturatingMultiplyAdd<uint64_t>(Count, Scale, Total);
      continue;
    }
    if (TE.State == TreeEntry::SplitVectorize)
      Count += 2;
    else if (TE.hasReassocScalars())
      // N operand columns need N-1 binary combines.
      Count += TE.getNumOperands() - 1;
    else
      ++Count;
    if (!TE.ReorderIndices.empty() || !TE.ReuseShuffleIndices.empty())
      ++Count;
    Total = SaturatingMultiplyAdd<uint64_t>(Count, Scale, Total);
  }
  for (const auto &VecAndScale : GatherExtractSourceVecs)
    Total = SaturatingAdd(Total, VecAndScale.second);
  // Count extract instructions from ExternalUses, skipping insertelements
  // (those get folded into shuffles, not real extracts).
  SmallPtrSet<Value *, 8> CountedExtracts;
  for (const ExternalUser &EU : ExternalUses) {
    if (isa_and_nonnull<InsertElementInst>(EU.User))
      continue;
    if (EU.User && EphValues.count(EU.User))
      continue;
    // Vector-typed scalars are consumed as whole vectors, no extract
    // instruction is emitted for them.
    if (isVectorizedTy(EU.Scalar->getType()) &&
        (!SLPReVec ||
         (EU.E.hasState() && EU.E.getOpcode() == Instruction::InsertElement)))
      continue;
    if (ExternalUsesAsOriginalScalar.contains(EU.Scalar))
      continue;
    if (!CountedExtracts.insert(EU.Scalar).second)
      continue;
    ++Total;
  }
  return Total;
}
 
void BoUpSLP::TreeEntry::buildAltOpShuffleMask(
    const function_ref<bool(Instruction *)> IsAltOp, SmallVectorImpl<int> &Mask,
    SmallVectorImpl<Value *> *OpScalars,
    SmallVectorImpl<Value *> *AltScalars) const {
  unsigned Sz = Scalars.size();
  Mask.assign(Sz, PoisonMaskElem);
  SmallVector<int> OrderMask;
  if (!ReorderIndices.empty())
    inversePermutation(ReorderIndices, OrderMask);
  for (unsigned I = 0; I < Sz; ++I) {
    unsigned Idx = I;
    if (!ReorderIndices.empty())
      Idx = OrderMask[I];
    if (isa<PoisonValue>(Scalars[Idx]))
      continue;
    auto *OpInst = cast<Instruction>(Scalars[Idx]);
    if (IsAltOp(OpInst)) {
      Mask[I] = Sz + Idx;
      if (AltScalars)
        AltScalars->push_back(OpInst);
    } else {
      Mask[I] = Idx;
      if (OpScalars)
        OpScalars->push_back(OpInst);
    }
  }
  if (!ReuseShuffleIndices.empty()) {
    SmallVector<int> NewMask(ReuseShuffleIndices.size(), PoisonMaskElem);
    transform(ReuseShuffleIndices, NewMask.begin(), [&Mask](int Idx) {
      return Idx != PoisonMaskElem ? Mask[Idx] : PoisonMaskElem;
    });
    Mask.swap(NewMask);
  }
}
 
static bool isMainInstruction(Instruction *I, Instruction *MainOp,
                              Instruction *AltOp,
                              const TargetLibraryInfo &TLI) {
  return InstructionsState(MainOp, AltOp).getMatchingMainOpOrAltOp(I) == MainOp;
}
 
TTI::OperandValueInfo BoUpSLP::getOperandInfo(ArrayRef<Value *> Ops) const {
  assert(!Ops.empty());
  const auto *Op0 = Ops.front();
 
  const bool IsConstant = all_of(Ops, [](Value *V) {
    // TODO: We should allow undef elements here
    return isConstant(V) && !isa<UndefValue>(V);
  });
  const bool IsUniform = all_of(Ops, [=](Value *V) {
    // TODO: We should allow undef elements here
    return V == Op0;
  });
  const bool IsPowerOfTwo = all_of(Ops, [](Value *V) {
    // TODO: We should allow undef elements here
    if (auto *CI = dyn_cast<ConstantInt>(V))
      return CI->getValue().isPowerOf2();
    return false;
  });
  const bool IsNegatedPowerOfTwo = all_of(Ops, [](Value *V) {
    // TODO: We should allow undef elements here
    if (auto *CI = dyn_cast<ConstantInt>(V))
      return CI->getValue().isNegatedPowerOf2();
    return false;
  });
 
  TTI::OperandValueKind VK = TTI::OK_AnyValue;
  if (IsConstant && IsUniform)
    VK = TTI::OK_UniformConstantValue;
  else if (IsConstant)
    VK = TTI::OK_NonUniformConstantValue;
  else if (IsUniform)
    VK = TTI::OK_UniformValue;
 
  TTI::OperandValueProperties VP = TTI::OP_None;
  VP = IsPowerOfTwo ? TTI::OP_PowerOf2 : VP;
  VP = IsNegatedPowerOfTwo ? TTI::OP_NegatedPowerOf2 : VP;
 
  return {VK, VP};
}
 
namespace {
/// The base class for shuffle instruction emission and shuffle cost estimation.
class BaseShuffleAnalysis {
protected:
  Type *ScalarTy = nullptr;
 
  BaseShuffleAnalysis(Type *ScalarTy) : ScalarTy(ScalarTy) {}
 
  /// V is expected to be a vectorized value.
  /// When REVEC is disabled, there is no difference between VF and
  /// VNumElements.
  /// When REVEC is enabled, VF is VNumElements / ScalarTyNumElements.
  /// e.g., if ScalarTy is <4 x Ty> and V1 is <8 x Ty>, 2 is returned instead
  /// of 8.
  unsigned getVF(Value *V) const {
    assert(V && "V cannot be nullptr");
    assert(isa<FixedVectorType>(V->getType()) &&
           "V does not have FixedVectorType");
    assert(ScalarTy && "ScalarTy cannot be nullptr");
    unsigned ScalarTyNumElements = getNumElements(ScalarTy);
    unsigned VNumElements =
        cast<FixedVectorType>(V->getType())->getNumElements();
    assert(VNumElements > ScalarTyNumElements &&
           "the number of elements of V is not large enough");
    assert(VNumElements % ScalarTyNumElements == 0 &&
           "the number of elements of V is not a vectorized value");
    return VNumElements / ScalarTyNumElements;
  }
 
  /// Checks if the mask is an identity mask.
  /// \param IsStrict if is true the function returns false if mask size does
  /// not match vector size.
  static bool isIdentityMask(ArrayRef<int> Mask, const FixedVectorType *VecTy,
                             bool IsStrict) {
    int Limit = Mask.size();
    int VF = VecTy->getNumElements();
    int Index = -1;
    if (VF == Limit && ShuffleVectorInst::isIdentityMask(Mask, Limit))
      return true;
    if (!IsStrict) {
      // Consider extract subvector starting from index 0.
      if (ShuffleVectorInst::isExtractSubvectorMask(Mask, VF, Index) &&
          Index == 0)
        return true;
      // All VF-size submasks are identity (e.g.
      // <poison,poison,poison,poison,0,1,2,poison,poison,1,2,3> etc. for VF 4).
      if (Limit % VF == 0 && all_of(seq<int>(0, Limit / VF), [=](int Idx) {
            ArrayRef<int> Slice = Mask.slice(Idx * VF, VF);
            return all_of(Slice, equal_to(PoisonMaskElem)) ||
                   ShuffleVectorInst::isIdentityMask(Slice, VF);
          }))
        return true;
    }
    return false;
  }
 
  /// Tries to combine 2 different masks into single one.
  /// \param LocalVF Vector length of the permuted input vector. \p Mask may
  /// change the size of the vector, \p LocalVF is the original size of the
  /// shuffled vector.
  static void combineMasks(unsigned LocalVF, SmallVectorImpl<int> &Mask,
                           ArrayRef<int> ExtMask) {
    unsigned VF = Mask.size();
    SmallVector<int> NewMask(ExtMask.size(), PoisonMaskElem);
    for (int I = 0, Sz = ExtMask.size(); I < Sz; ++I) {
      if (ExtMask[I] == PoisonMaskElem)
        continue;
      int MaskedIdx = Mask[ExtMask[I] % VF];
      NewMask[I] =
          MaskedIdx == PoisonMaskElem ? PoisonMaskElem : MaskedIdx % LocalVF;
    }
    Mask.swap(NewMask);
  }
 
  /// Looks through shuffles trying to reduce final number of shuffles in the
  /// code. The function looks through the previously emitted shuffle
  /// instructions and properly mark indices in mask as undef.
  /// For example, given the code
  /// \code
  /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0>
  /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0>
  /// \endcode
  /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 3, 2>, it will
  /// look through %s1 and %s2 and select vectors %0 and %1 with mask
  /// <0, 1, 2, 3> for the shuffle.
  /// If 2 operands are of different size, the smallest one will be resized and
  /// the mask recalculated properly.
  /// For example, given the code
  /// \code
  /// %s1 = shufflevector <2 x ty> %0, poison, <1, 0, 1, 0>
  /// %s2 = shufflevector <2 x ty> %1, poison, <1, 0, 1, 0>
  /// \endcode
  /// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 5, 4>, it will
  /// look through %s1 and %s2 and select vectors %0 and %1 with mask
  /// <0, 1, 2, 3> for the shuffle.
  /// So, it tries to transform permutations to simple vector merge, if
  /// possible.
  /// \param V The input vector which must be shuffled using the given \p Mask.
  /// If the better candidate is found, \p V is set to this best candidate
  /// vector.
  /// \param Mask The input mask for the shuffle. If the best candidate is found
  /// during looking-through-shuffles attempt, it is updated accordingly.
  /// \param SinglePermute true if the shuffle operation is originally a
  /// single-value-permutation. In this case the look-through-shuffles procedure
  /// may look for resizing shuffles as the best candidates.
  /// \return true if the shuffle results in the non-resizing identity shuffle
  /// (and thus can be ignored), false - otherwise.
  static bool peekThroughShuffles(Value *&V, SmallVectorImpl<int> &Mask,
                                  bool SinglePermute) {
    Value *Op = V;
    ShuffleVectorInst *IdentityOp = nullptr;
    SmallVector<int> IdentityMask;
    while (auto *SV = dyn_cast<ShuffleVectorInst>(Op)) {
      // Exit if not a fixed vector type or changing size shuffle.
      auto *SVTy = dyn_cast<FixedVectorType>(SV->getType());
      if (!SVTy)
        break;
      // Remember the identity or broadcast mask, if it is not a resizing
      // shuffle. If no better candidates are found, this Op and Mask will be
      // used in the final shuffle.
      if (isIdentityMask(Mask, SVTy, /*IsStrict=*/false)) {
        if (!IdentityOp || !SinglePermute ||
            (isIdentityMask(Mask, SVTy, /*IsStrict=*/true) &&
             !ShuffleVectorInst::isZeroEltSplatMask(IdentityMask,
                                                    IdentityMask.size()))) {
          IdentityOp = SV;
          // Store current mask in the IdentityMask so later we did not lost
          // this info if IdentityOp is selected as the best candidate for the
          // permutation.
          IdentityMask.assign(Mask);
        }
      }
      // Remember the broadcast mask. If no better candidates are found, this Op
      // and Mask will be used in the final shuffle.
      // Zero splat can be used as identity too, since it might be used with
      // mask <0, 1, 2, ...>, i.e. identity mask without extra reshuffling.
      // E.g. if need to shuffle the vector with the mask <3, 1, 2, 0>, which is
      // expensive, the analysis founds out, that the source vector is just a
      // broadcast, this original mask can be transformed to identity mask <0,
      // 1, 2, 3>.
      // \code
      // %0 = shuffle %v, poison, zeroinitalizer
      // %res = shuffle %0, poison, <3, 1, 2, 0>
      // \endcode
      // may be transformed to
      // \code
      // %0 = shuffle %v, poison, zeroinitalizer
      // %res = shuffle %0, poison, <0, 1, 2, 3>
      // \endcode
      if (SV->isZeroEltSplat()) {
        IdentityOp = SV;
        IdentityMask.assign(Mask);
      }
      int LocalVF = Mask.size();
      if (auto *SVOpTy =
              dyn_cast<FixedVectorType>(SV->getOperand(0)->getType()))
        LocalVF = SVOpTy->getNumElements();
      SmallVector<int> ExtMask(Mask.size(), PoisonMaskElem);
      for (auto [Idx, I] : enumerate(Mask)) {
        if (I == PoisonMaskElem ||
            static_cast<unsigned>(I) >= SV->getShuffleMask().size())
          continue;
        ExtMask[Idx] = SV->getMaskValue(I);
      }
      bool IsOp1Undef = isUndefVector</*isPoisonOnly=*/true>(
                            SV->getOperand(0),
                            buildUseMask(LocalVF, ExtMask, UseMask::FirstArg))
                            .all();
      bool IsOp2Undef = isUndefVector</*isPoisonOnly=*/true>(
                            SV->getOperand(1),
                            buildUseMask(LocalVF, ExtMask, UseMask::SecondArg))
                            .all();
      if (!IsOp1Undef && !IsOp2Undef) {
        // Update mask and mark undef elems.
        for (int &I : Mask) {
          if (I == PoisonMaskElem)
            continue;
          if (SV->getMaskValue(I % SV->getShuffleMask().size()) ==
              PoisonMaskElem)
            I = PoisonMaskElem;
        }
        break;
      }
      SmallVector<int> ShuffleMask(SV->getShuffleMask());
      combineMasks(LocalVF, ShuffleMask, Mask);
      Mask.swap(ShuffleMask);
      if (IsOp2Undef)
        Op = SV->getOperand(0);
      else
        Op = SV->getOperand(1);
    }
    if (auto *OpTy = dyn_cast<FixedVectorType>(Op->getType());
        !OpTy || !isIdentityMask(Mask, OpTy, SinglePermute) ||
        ShuffleVectorInst::isZeroEltSplatMask(Mask, Mask.size())) {
      if (IdentityOp) {
        V = IdentityOp;
        assert(Mask.size() == IdentityMask.size() &&
               "Expected masks of same sizes.");
        // Clear known poison elements.
        for (auto [I, Idx] : enumerate(Mask))
          if (Idx == PoisonMaskElem)
            IdentityMask[I] = PoisonMaskElem;
        Mask.swap(IdentityMask);
        auto *Shuffle = dyn_cast<ShuffleVectorInst>(V);
        return SinglePermute &&
               (isIdentityMask(Mask, cast<FixedVectorType>(V->getType()),
                               /*IsStrict=*/true) ||
                (Shuffle && Mask.size() == Shuffle->getShuffleMask().size() &&
                 Shuffle->isZeroEltSplat() &&
                 ShuffleVectorInst::isZeroEltSplatMask(Mask, Mask.size()) &&
                 all_of(enumerate(Mask), [&](const auto &P) {
                   return P.value() == PoisonMaskElem ||
                          Shuffle->getShuffleMask()[P.index()] == 0;
                 })));
      }
      V = Op;
      return false;
    }
    V = Op;
    return true;
  }
 
  /// Smart shuffle instruction emission, walks through shuffles trees and
  /// tries to find the best matching vector for the actual shuffle
  /// instruction.
  template <typename T, typename ShuffleBuilderTy, typename... Args>
  static T createShuffle(Value *V1, Value *V2, ArrayRef<int> Mask,
                         ShuffleBuilderTy &Builder, Type *ScalarTy,
                         Args... Arguments) {
    assert(V1 && "Expected at least one vector value.");
    unsigned ScalarTyNumElements = getNumElements(ScalarTy);
    SmallVector<int> NewMask(Mask);
    if (ScalarTyNumElements != 1) {
      assert(SLPReVec && "FixedVectorType is not expected.");
      transformScalarShuffleIndiciesToVector(ScalarTyNumElements, NewMask);
      Mask = NewMask;
    }
    if (V2)
      Builder.resizeToMatch(V1, V2);
    int VF = Mask.size();
    if (auto *FTy = dyn_cast<FixedVectorType>(V1->getType()))
      VF = FTy->getNumElements();
    if (V2 && !isUndefVector</*IsPoisonOnly=*/true>(
                   V2, buildUseMask(VF, Mask, UseMask::SecondArg))
                   .all()) {
      // Peek through shuffles.
      Value *Op1 = V1;
      Value *Op2 = V2;
      int VF =
          cast<VectorType>(V1->getType())->getElementCount().getKnownMinValue();
      SmallVector<int> CombinedMask1(Mask.size(), PoisonMaskElem);
      SmallVector<int> CombinedMask2(Mask.size(), PoisonMaskElem);
      for (int I = 0, E = Mask.size(); I < E; ++I) {
        if (Mask[I] < VF)
          CombinedMask1[I] = Mask[I];
        else
          CombinedMask2[I] = Mask[I] - VF;
      }
      Value *PrevOp1;
      Value *PrevOp2;
      do {
        PrevOp1 = Op1;
        PrevOp2 = Op2;
        (void)peekThroughShuffles(Op1, CombinedMask1, /*SinglePermute=*/false);
        (void)peekThroughShuffles(Op2, CombinedMask2, /*SinglePermute=*/false);
        // Check if we have 2 resizing shuffles - need to peek through operands
        // again.
        if (auto *SV1 = dyn_cast<ShuffleVectorInst>(Op1))
          if (auto *SV2 = dyn_cast<ShuffleVectorInst>(Op2)) {
            SmallVector<int> ExtMask1(Mask.size(), PoisonMaskElem);
            for (auto [Idx, I] : enumerate(CombinedMask1)) {
                if (I == PoisonMaskElem)
                continue;
                ExtMask1[Idx] = SV1->getMaskValue(I);
            }
            SmallBitVector UseMask1 = buildUseMask(
                cast<FixedVectorType>(SV1->getOperand(1)->getType())
                    ->getNumElements(),
                ExtMask1, UseMask::SecondArg);
            SmallVector<int> ExtMask2(CombinedMask2.size(), PoisonMaskElem);
            for (auto [Idx, I] : enumerate(CombinedMask2)) {
                if (I == PoisonMaskElem)
                continue;
                ExtMask2[Idx] = SV2->getMaskValue(I);
            }
            SmallBitVector UseMask2 = buildUseMask(
                cast<FixedVectorType>(SV2->getOperand(1)->getType())
                    ->getNumElements(),
                ExtMask2, UseMask::SecondArg);
            if (SV1->getOperand(0)->getType() ==
                    SV2->getOperand(0)->getType() &&
                SV1->getOperand(0)->getType() != SV1->getType() &&
                isUndefVector(SV1->getOperand(1), UseMask1).all() &&
                isUndefVector(SV2->getOperand(1), UseMask2).all()) {
              Op1 = SV1->getOperand(0);
              Op2 = SV2->getOperand(0);
              SmallVector<int> ShuffleMask1(SV1->getShuffleMask());
              int LocalVF = ShuffleMask1.size();
              if (auto *FTy = dyn_cast<FixedVectorType>(Op1->getType()))
                LocalVF = FTy->getNumElements();
              combineMasks(LocalVF, ShuffleMask1, CombinedMask1);
              CombinedMask1.swap(ShuffleMask1);
              SmallVector<int> ShuffleMask2(SV2->getShuffleMask());
              LocalVF = ShuffleMask2.size();
              if (auto *FTy = dyn_cast<FixedVectorType>(Op2->getType()))
                LocalVF = FTy->getNumElements();
              combineMasks(LocalVF, ShuffleMask2, CombinedMask2);
              CombinedMask2.swap(ShuffleMask2);
            }
          }
      } while (PrevOp1 != Op1 || PrevOp2 != Op2);
      Builder.resizeToMatch(Op1, Op2);
      VF = std::max(cast<VectorType>(Op1->getType())
                        ->getElementCount()
                        .getKnownMinValue(),
                    cast<VectorType>(Op2->getType())
                        ->getElementCount()
                        .getKnownMinValue());
      for (int I = 0, E = Mask.size(); I < E; ++I) {
        if (CombinedMask2[I] != PoisonMaskElem) {
          assert(CombinedMask1[I] == PoisonMaskElem &&
                 "Expected undefined mask element");
          CombinedMask1[I] = CombinedMask2[I] + (Op1 == Op2 ? 0 : VF);
        }
      }
      if (Op1 == Op2 &&
          (ShuffleVectorInst::isIdentityMask(CombinedMask1, VF) ||
           (ShuffleVectorInst::isZeroEltSplatMask(CombinedMask1, VF) &&
            isa<ShuffleVectorInst>(Op1) &&
            cast<ShuffleVectorInst>(Op1)->getShuffleMask() ==
                ArrayRef(CombinedMask1))))
        return Builder.createIdentity(Op1);
      return Builder.createShuffleVector(
          Op1, Op1 == Op2 ? PoisonValue::get(Op1->getType()) : Op2,
          CombinedMask1);
    }
    if (isa<PoisonValue>(V1))
      return Builder.createPoison(
          cast<VectorType>(V1->getType())->getElementType(), Mask.size());
    bool IsIdentity = peekThroughShuffles(V1, NewMask, /*SinglePermute=*/true);
    assert(V1 && "Expected non-null value after looking through shuffles.");
 
    if (!IsIdentity)
      return Builder.createShuffleVector(V1, NewMask, Arguments...);
    return Builder.createIdentity(V1);
  }
 
  /// Transforms mask \p CommonMask per given \p Mask to make proper set after
  /// shuffle emission.
  static void transformMaskAfterShuffle(MutableArrayRef<int> CommonMask,
                                        ArrayRef<int> Mask) {
    for (unsigned I : seq<unsigned>(CommonMask.size()))
      if (Mask[I] != PoisonMaskElem)
        CommonMask[I] = I;
  }
};
} // namespace
 
void BoUpSLP::reorderGatherNode(TreeEntry &TE) {
  assert(TE.isGather() && TE.ReorderIndices.empty() &&
         "Expected gather node without reordering.");
  DenseMap<std::pair<size_t, Value *>, SmallVector<LoadInst *>> LoadsMap;
  SmallSet<size_t, 2> LoadKeyUsed;
 
  // Do not reorder nodes if it small (just 2 elements), all-constant or all
  // instructions have same opcode already.
  if (TE.Scalars.size() == 2 || (TE.hasState() && !TE.isAltShuffle()) ||
      all_of(TE.Scalars, isConstant))
    return;
 
  if (any_of(seq<unsigned>(TE.Idx), [&](unsigned Idx) {
        return VectorizableTree[Idx]->isSame(TE.Scalars);
      }))
    return;
 
  auto GenerateLoadsSubkey = [&](size_t Key, LoadInst *LI) {
    Key = hash_combine(hash_value(LI->getParent()->getNumber()), Key);
    Value *Ptr =
        getUnderlyingObject(LI->getPointerOperand(), RecursionMaxDepth);
    if (LoadKeyUsed.contains(Key)) {
      auto LIt = LoadsMap.find(std::make_pair(Key, Ptr));
      if (LIt != LoadsMap.end()) {
        for (LoadInst *RLI : LIt->second) {
          if (getPointersDiff(RLI->getType(), RLI->getPointerOperand(),
                              LI->getType(), LI->getPointerOperand(), *DL, *SE,
                              /*StrictCheck=*/true))
            return hash_value(RLI->getPointerOperand());
        }
        for (LoadInst *RLI : LIt->second) {
          if (arePointersCompatible(RLI->getPointerOperand(),
                                    LI->getPointerOperand(), *TLI)) {
            hash_code SubKey = hash_value(RLI->getPointerOperand());
            return SubKey;
          }
        }
        if (LIt->second.size() > 2) {
          hash_code SubKey =
              hash_value(LIt->second.back()->getPointerOperand());
          return SubKey;
        }
      }
    }
    LoadKeyUsed.insert(Key);
    LoadsMap.try_emplace(std::make_pair(Key, Ptr)).first->second.push_back(LI);
    return hash_value(LI->getPointerOperand());
  };
  MapVector<size_t, MapVector<size_t, SmallVector<Value *>>> SortedValues;
  SmallDenseMap<Value *, SmallVector<unsigned>, 8> KeyToIndex;
  bool IsOrdered = true;
  unsigned NumInstructions = 0;
  // Try to "cluster" scalar instructions, to be able to build extra vectorized
  // nodes.
  for (auto [I, V] : enumerate(TE.Scalars)) {
    size_t Key = 1, Idx = 1;
    if (auto *Inst = dyn_cast<Instruction>(V);
        Inst && !isa<ExtractElementInst, LoadInst, CastInst>(V) &&
        !isDeleted(Inst) && !isVectorized(V)) {
      std::tie(Key, Idx) = generateKeySubkey(V, TLI, GenerateLoadsSubkey,
                                             /*AllowAlternate=*/false);
      ++NumInstructions;
    }
    auto &Container = SortedValues[Key];
    if (IsOrdered && !KeyToIndex.contains(V) &&
        !(isa<Constant, ExtractElementInst>(V) ||
          isVectorLikeInstWithConstOps(V)) &&
        ((Container.contains(Idx) &&
          KeyToIndex.at(Container[Idx].back()).back() != I - 1) ||
         (!Container.empty() && !Container.contains(Idx) &&
          KeyToIndex.at(Container.back().second.back()).back() != I - 1)))
      IsOrdered = false;
    auto &KTI = KeyToIndex[V];
    if (KTI.empty())
      Container[Idx].push_back(V);
    KTI.push_back(I);
  }
  SmallVector<std::pair<unsigned, unsigned>> SubVectors;
  APInt DemandedElts = APInt::getAllOnes(TE.Scalars.size());
  if (!IsOrdered && NumInstructions > 1) {
    unsigned Cnt = 0;
    TE.ReorderIndices.resize(TE.Scalars.size(), TE.Scalars.size());
    for (const auto &D : SortedValues) {
      for (const auto &P : D.second) {
        unsigned Sz = 0;
        for (Value *V : P.second) {
          ArrayRef<unsigned> Indices = KeyToIndex.at(V);
          for (auto [K, Idx] : enumerate(Indices)) {
            TE.ReorderIndices[Cnt + K] = Idx;
            TE.Scalars[Cnt + K] = V;
          }
          Sz += Indices.size();
          Cnt += Indices.size();
        }
        if (Sz > 1 && isa<Instruction>(P.second.front())) {
          const unsigned SubVF = getFloorFullVectorNumberOfElements(
              *TTI, TE.Scalars.front()->getType(), Sz, SLPReVec);
          SubVectors.emplace_back(Cnt - Sz, SubVF);
          for (unsigned I : seq<unsigned>(Cnt - Sz, Cnt - Sz + SubVF))
            DemandedElts.clearBit(I);
        } else if (!P.second.empty() && isConstant(P.second.front())) {
          for (unsigned I : seq<unsigned>(Cnt - Sz, Cnt))
            DemandedElts.clearBit(I);
        }
      }
    }
  }
  // Reuses always require shuffles, so consider it as profitable.
  if (!TE.ReuseShuffleIndices.empty() || TE.ReorderIndices.empty())
    return;
  // Do simple cost estimation.
  InstructionCost Cost = 0;
  auto *ScalarTy = TE.Scalars.front()->getType();
  auto *VecTy = cast<VectorType>(getWidenedType(ScalarTy, TE.Scalars.size()));
  for (auto [Idx, Sz] : SubVectors) {
    Cost += getShuffleCost(*TTI, TTI::SK_InsertSubvector, VecTy, CostKind, {},
                           Idx, cast<VectorType>(getWidenedType(ScalarTy, Sz)));
  }
  Cost +=
      getScalarizationOverhead(*TTI, SLPReVec, ScalarTy, VecTy, DemandedElts,
                               /*Insert=*/true,
                               /*Extract=*/false, CostKind);
  int Sz = TE.Scalars.size();
  SmallVector<int> ReorderMask(TE.ReorderIndices.begin(),
                               TE.ReorderIndices.end());
  for (unsigned I : seq<unsigned>(Sz)) {
    Value *V = TE.getOrdered(I);
    if (isa<PoisonValue>(V)) {
      ReorderMask[I] = PoisonMaskElem;
    } else if (isConstant(V) || DemandedElts[I]) {
      ReorderMask[I] = I + TE.ReorderIndices.size();
    }
  }
  Cost += getShuffleCost(*TTI,
                         any_of(ReorderMask, [&](int I) { return I >= Sz; })
                             ? TTI::SK_PermuteTwoSrc
                             : TTI::SK_PermuteSingleSrc,
                         VecTy, CostKind, ReorderMask);
  DemandedElts = APInt::getAllOnes(TE.Scalars.size());
  ReorderMask.assign(Sz, PoisonMaskElem);
  for (unsigned I : seq<unsigned>(Sz)) {
    Value *V = TE.getOrdered(I);
    if (isConstant(V)) {
      DemandedElts.clearBit(I);
      if (!isa<PoisonValue>(V))
        ReorderMask[I] = I;
    } else {
      ReorderMask[I] = I + Sz;
    }
  }
  InstructionCost BVCost =
      getScalarizationOverhead(*TTI, SLPReVec, ScalarTy, VecTy, DemandedElts,
                               /*Insert=*/true, /*Extract=*/false, CostKind);
  if (!DemandedElts.isAllOnes())
    BVCost += getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc, VecTy, CostKind,
                             ReorderMask);
  if (Cost >= BVCost) {
    SmallVector<int> Mask(TE.ReorderIndices.begin(), TE.ReorderIndices.end());
    reorderScalars(TE.Scalars, Mask);
    TE.ReorderIndices.clear();
  }
}
 
/// Check if we can convert fadd/fsub sequence to FMAD.
/// \returns Cost of the FMAD, if conversion is possible, invalid cost otherwise.
static InstructionCost canConvertToFMA(ArrayRef<Value *> VL,
                                       const InstructionsState &S,
                                       DominatorTree &DT, const DataLayout &DL,
                                       TargetTransformInfo &TTI,
                                       const TargetLibraryInfo &TLI,
                                       const TTI::TargetCostKind CostKind) {
  assert(all_of(VL,
                [](Value *V) {
                  return V->getType()->getScalarType()->isFloatingPointTy();
                }) &&
         "Can only convert to FMA for floating point types");
  assert(S.isAddSubLikeOp() && "Can only convert to FMA for add/sub");
 
  auto CheckForContractable = [](ArrayRef<Value *> VL,
                                 const InstructionsState &S) {
    FastMathFlags FMF;
    FMF.set();
    for (Value *V : VL) {
      auto *I = dyn_cast<Instruction>(V);
      if (!I)
        continue;
      if (S.isCopyableElement(I))
        continue;
      Instruction *MatchingI = S.getMatchingMainOpOrAltOp(I);
      if (S.getMainOp() != MatchingI && S.getAltOp() != MatchingI)
        continue;
      if (auto *FPCI = dyn_cast<FPMathOperator>(I))
        FMF &= FPCI->getFastMathFlags();
    }
    return FMF.allowContract();
  };
  if (!CheckForContractable(VL, S))
    return InstructionCost::getInvalid();
  // fmul also should be contractable
  InstructionsCompatibilityAnalysis Analysis(DT, DL, TTI, TLI);
  SmallVector<BoUpSLP::ValueList> Operands = Analysis.buildOperands(S, VL);
 
  InstructionsState OpS = getSameOpcode(Operands.front(), TLI);
  if (!OpS.valid())
    return InstructionCost::getInvalid();
 
  if (OpS.isAltShuffle() || OpS.getOpcode() != Instruction::FMul)
    return InstructionCost::getInvalid();
  if (!CheckForContractable(Operands.front(), OpS))
    return InstructionCost::getInvalid();
  // Compare the costs.
  InstructionCost FMulPlusFAddCost = 0;
  InstructionCost FMACost = 0;
  // Price both sides of the fmul+fadd pair as not fused. Passing a context
  // instruction would let targets that model the fusion discount the unfused
  // side of the comparison as well.
  auto GetUnfusedFMulCost = [&](Instruction *I) {
    assert(I->getOpcode() == Instruction::FMul && "Expected an fmul");
    TTI::OperandValueInfo Op1Info = TTI::getOperandInfo(I->getOperand(0));
    TTI::OperandValueInfo Op2Info = TTI::getOperandInfo(I->getOperand(1));
    return TTI.getArithmeticInstrCost(I->getOpcode(), I->getType(), CostKind,
                                      Op1Info, Op2Info,
                                      {I->getOperand(0), I->getOperand(1)});
  };
  FastMathFlags FMF;
  FMF.set();
  const bool IsArithmeticState = S.isAddSubLikeOp() || S.isMulDivLikeOp() ||
                                 S.isShiftOp() || S.isBitwiseLogicOp();
  for (Value *V : VL) {
    auto *I = dyn_cast<Instruction>(V);
    if (!I)
      continue;
    const bool IsCopyable = S.isCopyableElement(I);
    if (!IsCopyable)
      if (auto *FPCI = dyn_cast<FPMathOperator>(I))
        FMF &= FPCI->getFastMathFlags();
    if (IsCopyable || !IsArithmeticState ||
        (I->getOpcode() != S.getOpcode() &&
         I->getOpcode() != S.getAltOpcode())) {
      FMulPlusFAddCost += TTI.getInstructionCost(I, CostKind);
      continue;
    }
    TTI::OperandValueInfo Op1Info = TTI::getOperandInfo(I->getOperand(0));
    TTI::OperandValueInfo Op2Info = TTI::getOperandInfo(I->getOperand(1));
    FMulPlusFAddCost += TTI.getArithmeticInstrCost(
        I->getOpcode(), I->getType(), CostKind, Op1Info, Op2Info,
        {I->getOperand(0), I->getOperand(1)});
  }
  unsigned NumOps = 0;
  for (auto [V, Op] : zip(VL, Operands.front())) {
    if (S.isCopyableElement(V))
      continue;
    auto *I = dyn_cast<Instruction>(Op);
    if (!I || !I->hasOneUse() || OpS.isCopyableElement(I)) {
      if (auto *OpI = dyn_cast<Instruction>(V))
        FMACost += TTI.getInstructionCost(OpI, CostKind);
      if (I)
        FMACost += TTI.getInstructionCost(I, CostKind);
      continue;
    }
    ++NumOps;
    if (auto *FPCI = dyn_cast<FPMathOperator>(I))
      FMF &= FPCI->getFastMathFlags();
    FMulPlusFAddCost += GetUnfusedFMulCost(I);
  }
  Type *Ty = VL.front()->getType();
  IntrinsicCostAttributes ICA(Intrinsic::fmuladd, Ty, {Ty, Ty, Ty}, FMF);
  FMACost += NumOps * TTI.getIntrinsicInstrCost(ICA, CostKind);
  return FMACost < FMulPlusFAddCost ? FMACost : InstructionCost::getInvalid();
}
 
bool BoUpSLP::matchesShlZExt(const TreeEntry &TE, OrdersType &Order,
                             bool &IsBSwap, bool &ForLoads) const {
  assert(TE.hasState() && TE.getOpcode() == Instruction::Shl &&
         "Expected Shl node.");
  IsBSwap = false;
  ForLoads = false;
  if (TE.State != TreeEntry::Vectorize || !TE.ReorderIndices.empty() ||
      !TE.ReuseShuffleIndices.empty() || MinBWs.contains(&TE) ||
      any_of(TE.Scalars, [](Value *V) { return !V->hasOneUse(); }))
    return false;
  Type *ScalarTy = TE.getMainOp()->getType();
  // TODO: Check if same can be done for the vector types.
  if (!ScalarTy->isIntegerTy())
    return false;
  if (ScalarTy->isVectorTy())
    return false;
  const unsigned Sz = DL->getTypeSizeInBits(ScalarTy);
  const TreeEntry *LhsTE = getOperandEntry(&TE, /*Idx=*/0);
  const TreeEntry *RhsTE = getOperandEntry(&TE, /*Idx=*/1);
  // Lhs should be zext i<stride> to I<sz>.
  if (!(LhsTE->State == TreeEntry::Vectorize &&
        LhsTE->getOpcode() == Instruction::ZExt &&
        LhsTE->ReorderIndices.empty() && LhsTE->ReuseShuffleIndices.empty() &&
        !MinBWs.contains(LhsTE) &&
        all_of(LhsTE->Scalars, [](Value *V) { return V->hasOneUse(); })))
    return false;
  Type *SrcScalarTy = cast<ZExtInst>(LhsTE->getMainOp())->getSrcTy();
  unsigned Stride = DL->getTypeSizeInBits(SrcScalarTy);
  if (!isPowerOf2_64(Stride) || Stride >= Sz || Sz % Stride != 0 ||
      !isPowerOf2_64(LhsTE->getVectorFactor()))
    return false;
  if (!(RhsTE->isGather() && RhsTE->ReorderIndices.empty() &&
        RhsTE->ReuseShuffleIndices.empty() && !MinBWs.contains(RhsTE)))
    return false;
  Order.clear();
  unsigned CurrentValue = 0;
  // Rhs should be (0, Stride, 2 * Stride, ..., N-Stride), where N <= Sz.
  if (all_of(RhsTE->Scalars,
             [&](Value *V) {
               CurrentValue += Stride;
               if (isa<UndefValue>(V))
                 return true;
               auto *C = dyn_cast<Constant>(V);
               if (!C)
                 return false;
               return C->getUniqueInteger() == CurrentValue - Stride;
             }) &&
      CurrentValue <= Sz) {
    Order.clear();
  } else {
    const unsigned VF = RhsTE->getVectorFactor();
    Order.assign(VF, VF);
    // Track which logical positions we've seen; reject duplicate shift amounts.
    SmallBitVector SeenPositions(VF);
    // Check if need to reorder Rhs to make it in form (0, Stride, 2 * Stride,
    // ..., N-Stride), where N <= Sz.
    if (VF * Stride > Sz)
      return false;
    for (const auto [Idx, V] : enumerate(RhsTE->Scalars)) {
      if (isa<UndefValue>(V))
        continue;
      auto *C = dyn_cast<Constant>(V);
      if (!C)
        return false;
      const APInt &Val = C->getUniqueInteger();
      if (Val.isNegative() || Val.uge(Sz) || Val.getZExtValue() % Stride != 0)
        return false;
      unsigned Pos = Val.getZExtValue() / Stride;
      // TODO: Support Pos >= VF, in this case need to shift the final value.
      if (Order[Idx] != VF || Pos >= VF)
        return false;
      if (SeenPositions.test(Pos))
        return false;
      SeenPositions.set(Pos);
      Order[Idx] = Pos;
    }
    // One of the indices not set - exit.
    if (is_contained(Order, VF))
      return false;
  }
  auto *SrcType = IntegerType::getIntNTy(ScalarTy->getContext(),
                                         Stride * LhsTE->getVectorFactor());
  FastMathFlags FMF;
  SmallPtrSet<Value *, 4> CheckedExtracts;
  auto *VecTy =
      cast<VectorType>(getWidenedType(ScalarTy, TE.getVectorFactor()));
  auto *SrcVecTy =
      cast<VectorType>(getWidenedType(SrcScalarTy, LhsTE->getVectorFactor()));
  TTI::CastContextHint CastCtx =
      getCastContextHint(*getOperandEntry(LhsTE, /*Idx=*/0));
  InstructionCost VecCost =
      TTI->getArithmeticReductionCost(Instruction::Or, VecTy, FMF, CostKind) +
      TTI->getArithmeticInstrCost(Instruction::Shl, VecTy, CostKind,
                                  getOperandInfo(LhsTE->Scalars)) +
      TTI->getCastInstrCost(
          Instruction::ZExt, VecTy,
          getWidenedType(SrcScalarTy, LhsTE->getVectorFactor()), CastCtx,
          CostKind);
  InstructionCost BitcastCost = TTI->getCastInstrCost(
      Instruction::BitCast, SrcType, SrcVecTy, CastCtx, CostKind);
  if (!Order.empty()) {
    fixupOrderingIndices(Order);
    SmallVector<int> Mask;
    inversePermutation(Order, Mask);
    BitcastCost += getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc, SrcVecTy,
                                  CostKind, Mask);
  }
  // Check if the combination can be modeled as a bitcast+byteswap operation.
  constexpr unsigned ByteSize = 8;
  if (!Order.empty() && isReverseOrder(Order) &&
      DL->getTypeSizeInBits(SrcScalarTy) == ByteSize) {
    IntrinsicCostAttributes CostAttrs(Intrinsic::bswap, SrcType, {SrcType});
    InstructionCost BSwapCost =
        TTI->getCastInstrCost(Instruction::BitCast, SrcType, SrcVecTy, CastCtx,
                              CostKind) +
        TTI->getIntrinsicInstrCost(CostAttrs, CostKind);
    if (BSwapCost <= BitcastCost) {
      BitcastCost = BSwapCost;
      IsBSwap = true;
      Order.clear();
      // Check for loads in the ZExt node.
      const TreeEntry *SrcTE = getOperandEntry(LhsTE, /*Idx=*/0);
      if (SrcTE->State == TreeEntry::Vectorize &&
          SrcTE->ReorderIndices.empty() && SrcTE->ReuseShuffleIndices.empty() &&
          SrcTE->getOpcode() == Instruction::Load && !SrcTE->isAltShuffle() &&
          all_of(SrcTE->Scalars, [](Value *V) { return V->hasOneUse(); })) {
        auto *LI = cast<LoadInst>(SrcTE->getMainOp());
        IntrinsicCostAttributes CostAttrs(Intrinsic::bswap, SrcType, {SrcType});
        InstructionCost BSwapCost =
            TTI->getMemoryOpCost(Instruction::Load, SrcType, LI->getAlign(),
                                 LI->getPointerAddressSpace(), CostKind) +
            TTI->getIntrinsicInstrCost(CostAttrs, CostKind);
        if (BSwapCost <= BitcastCost) {
          VecCost +=
              TTI->getMemoryOpCost(Instruction::Load, SrcVecTy, LI->getAlign(),
                                   LI->getPointerAddressSpace(), CostKind);
          BitcastCost = BSwapCost;
          ForLoads = true;
        }
      }
    }
  } else if (Order.empty() && DL->getTypeSizeInBits(SrcScalarTy) == ByteSize) {
    // Check for loads in the ZExt node.
    const TreeEntry *SrcTE = getOperandEntry(LhsTE, /*Idx=*/0);
    if (SrcTE->State == TreeEntry::Vectorize && SrcTE->ReorderIndices.empty() &&
        SrcTE->ReuseShuffleIndices.empty() &&
        SrcTE->getOpcode() == Instruction::Load && !SrcTE->isAltShuffle() &&
        all_of(SrcTE->Scalars, [](Value *V) { return V->hasOneUse(); })) {
      auto *LI = cast<LoadInst>(SrcTE->getMainOp());
      BitcastCost =
          TTI->getMemoryOpCost(Instruction::Load, SrcType, LI->getAlign(),
                               LI->getPointerAddressSpace(), CostKind);
      VecCost +=
          TTI->getMemoryOpCost(Instruction::Load, SrcVecTy, LI->getAlign(),
                               LI->getPointerAddressSpace(), CostKind);
      ForLoads = true;
    }
  }
  if (SrcType != ScalarTy) {
    BitcastCost += TTI->getCastInstrCost(Instruction::ZExt, ScalarTy, SrcType,
                                         TTI::CastContextHint::None, CostKind);
  }
  return BitcastCost < VecCost;
}
 
bool BoUpSLP::matchesInversedZExtSelect(
    const TreeEntry &SelectTE,
    SmallVectorImpl<unsigned> &InversedCmpsIndices) const {
  assert(SelectTE.hasState() && SelectTE.getOpcode() == Instruction::Select &&
         "Expected select node.");
  SmallVector<std::pair<Instruction *, unsigned>> ZExts;
  for (auto [Idx, V] : enumerate(SelectTE.Scalars)) {
    auto *Inst = dyn_cast<Instruction>(V);
    if (!Inst || Inst->getOpcode() != Instruction::ZExt)
      continue;
    ZExts.emplace_back(Inst, Idx);
  }
  if (ZExts.empty())
    return false;
  const auto *CmpTE = getOperandEntry(&SelectTE, 0);
  const auto *Op1TE = getOperandEntry(&SelectTE, 1);
  const auto *Op2TE = getOperandEntry(&SelectTE, 2);
  // Compares must be alternate vectorized, and other operands must be gathers
  // or copyables.
  // TODO: investigate opportunity for reordered/reused nodes.
  if (CmpTE->State != TreeEntry::Vectorize || !CmpTE->isAltShuffle() ||
      (CmpTE->getOpcode() != Instruction::ICmp &&
       CmpTE->getOpcode() != Instruction::FCmp) ||
      !CmpTE->ReorderIndices.empty() || !CmpTE->ReuseShuffleIndices.empty() ||
      !Op1TE->ReorderIndices.empty() || !Op1TE->ReuseShuffleIndices.empty() ||
      !Op2TE->ReorderIndices.empty() || !Op2TE->ReuseShuffleIndices.empty())
    return false;
  // The operands must be buildvectors/copyables.
  if (!Op1TE->isGather() || !Op2TE->isGather())
    return false;
  // TODO: investigate opportunity for the vector nodes with copyables.
  auto *Cmp = CmpTE->getMainOp();
  CmpPredicate Pred;
  auto MatchCmp = m_Cmp(Pred, m_Value(), m_Value());
  if (!match(Cmp, MatchCmp))
    return false;
  CmpPredicate MainPred = Pred;
  CmpPredicate InversedPred(CmpInst::getInversePredicate(Pred),
                            Pred.hasSameSign());
  for (const auto [Idx, V] : enumerate(CmpTE->Scalars)) {
    if (!match(V, MatchCmp))
      continue;
    if (CmpPredicate::getMatching(MainPred, Pred))
      continue;
    if (!CmpPredicate::getMatching(InversedPred, Pred))
      return false;
    if (!V->hasOneUse())
      return false;
    InversedCmpsIndices.push_back(Idx);
  }
 
  if (InversedCmpsIndices.empty())
    return false;
  Type *VecTy =
      getWidenedType(Cmp->getOperand(0)->getType(), CmpTE->getVectorFactor());
  Type *CmpTy = CmpInst::makeCmpResultType(VecTy);
 
  InstructionCost VecCost =
      TTI->getCmpSelInstrCost(CmpTE->getOpcode(), VecTy, CmpTy, MainPred,
                              CostKind, getOperandInfo(CmpTE->getOperand(0)),
                              getOperandInfo(CmpTE->getOperand(1)));
  InstructionCost BVCost = getScalarizationOverhead(
      *TTI, SLPReVec, Cmp->getType(), cast<VectorType>(CmpTy),
      APInt::getAllOnes(CmpTE->getVectorFactor()),
      /*Insert=*/true, /*Extract=*/false, CostKind);
  for (Value *V : CmpTE->Scalars) {
    auto *I = dyn_cast<Instruction>(V);
    if (!I)
      continue;
    BVCost += TTI->getInstructionCost(I, CostKind);
  }
  return VecCost < BVCost;
}
 
bool BoUpSLP::matchesSelectOfBits(const TreeEntry &SelectTE) const {
  assert(SelectTE.hasState() && SelectTE.getOpcode() == Instruction::Select &&
         "Expected select node.");
  if (DL->isBigEndian())
    return false;
  if (!SelectTE.ReorderIndices.empty() || !SelectTE.ReuseShuffleIndices.empty())
    return false;
  if (!UserIgnoreList || SelectTE.Idx != 0)
    return false;
  if (any_of(SelectTE.Scalars, [](Value *V) { return !V->hasOneUse(); }))
    return false;
  // Check that all reduction operands are or instructions.
  if (any_of(*UserIgnoreList,
             [](Value *V) { return !match(V, m_Or(m_Value(), m_Value())); }))
    return false;
  const TreeEntry *Op1TE = getOperandEntry(&SelectTE, 1);
  const TreeEntry *Op2TE = getOperandEntry(&SelectTE, 2);
  if (!Op1TE->isGather() || !Op2TE->isGather())
    return false;
  // No need to check for zeroes reordering.
  if (!Op1TE->ReorderIndices.empty() || !Op1TE->ReuseShuffleIndices.empty() ||
      !Op2TE->ReuseShuffleIndices.empty())
    return false;
  Type *ScalarTy = Op1TE->Scalars.front()->getType();
  if (!ScalarTy->isIntegerTy())
    return false;
  // Check that second operand is all zeroes.
  if (any_of(Op2TE->Scalars, [](Value *V) { return !match(V, m_ZeroInt()); }))
    return false;
  // Check that first operand is 1,2,4,...
  if (any_of(enumerate(Op1TE->Scalars), [](const auto &P) {
        uint64_t V;
        return !(match(P.value(), m_ConstantInt(V)) && isPowerOf2_64(V) &&
                 Log2_64(V) == P.index());
      }))
    return false;
  // Check if bitcast is cheaper than select.
  auto *DstTy = IntegerType::getIntNTy(ScalarTy->getContext(),
                                       SelectTE.getVectorFactor());
  Type *OpTy = getWidenedType(DstTy, SelectTE.getVectorFactor());
  Type *CmpTy = CmpInst::makeCmpResultType(OpTy);
  auto *VecTy =
      cast<VectorType>(getWidenedType(ScalarTy, SelectTE.getVectorFactor()));
  auto It = MinBWs.find(&SelectTE);
  if (It != MinBWs.end()) {
    auto *EffectiveScalarTy =
        IntegerType::get(F->getContext(), It->second.first);
    VecTy = cast<VectorType>(
        getWidenedType(EffectiveScalarTy, SelectTE.getVectorFactor()));
  }
  InstructionCost BitcastCost = TTI->getCastInstrCost(
      Instruction::BitCast, DstTy, CmpTy, TTI::CastContextHint::None, CostKind);
  if (DstTy != ScalarTy) {
    BitcastCost += TTI->getCastInstrCost(Instruction::ZExt, ScalarTy, DstTy,
                                         TTI::CastContextHint::None, CostKind);
  }
  FastMathFlags FMF;
  InstructionCost SelectCost =
      TTI->getCmpSelInstrCost(Instruction::Select, VecTy, CmpTy,
                              CmpInst::BAD_ICMP_PREDICATE, CostKind,
                              getOperandInfo(Op1TE->Scalars),
                              getOperandInfo(Op2TE->Scalars)) +
      TTI->getArithmeticReductionCost(Instruction::Or, VecTy, FMF, CostKind);
  return BitcastCost <= SelectCost;
}
 
void BoUpSLP::transformNodes() {
  BaseGraphSize = VectorizableTree.size();
  // Turn graph transforming mode on and off, when done.
  class GraphTransformModeRAAI {
    bool &SavedIsGraphTransformMode;
 
  public:
    GraphTransformModeRAAI(bool &IsGraphTransformMode)
        : SavedIsGraphTransformMode(IsGraphTransformMode) {
      IsGraphTransformMode = true;
    }
    ~GraphTransformModeRAAI() { SavedIsGraphTransformMode = false; }
  } TransformContext(IsGraphTransformMode);
  // Operands are profitable if they are:
  // 1. At least one constant
  // or
  // 2. Splats
  // or
  // 3. Results in good vectorization opportunity, i.e. may generate vector
  // nodes and reduce cost of the graph.
  auto CheckOperandsProfitability = [this](Instruction *I1, Instruction *I2,
                                           const InstructionsState &S) {
    SmallVector<SmallVector<std::pair<Value *, Value *>>> Candidates;
    for (unsigned Op : seq<unsigned>(S.getMainOp()->getNumOperands()))
      Candidates.emplace_back().emplace_back(I1->getOperand(Op),
                                             I2->getOperand(Op));
    return all_of(Candidates, [this](
                                  ArrayRef<std::pair<Value *, Value *>> Cand) {
      return all_of(Cand,
                    [](const std::pair<Value *, Value *> &P) {
                      return isa<Constant>(P.first) ||
                             isa<Constant>(P.second) || P.first == P.second;
                    }) ||
             findBestRootPair(Cand, LookAheadHeuristics::ScoreSplatLoads).first;
    });
  };
 
  // Try to reorder gather nodes for better vectorization opportunities.
  for (unsigned Idx : seq<unsigned>(BaseGraphSize)) {
    TreeEntry &E = *VectorizableTree[Idx];
    if (E.isGather())
      reorderGatherNode(E);
  }
 
  // Better to use full gathered loads analysis, if there are only 2 loads
  // gathered nodes each having less than 16 elements.
  constexpr unsigned VFLimit = 16;
  bool ForceLoadGather =
      count_if(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
        return TE->isGather() && TE->hasState() &&
               TE->getOpcode() == Instruction::Load &&
               TE->getVectorFactor() < VFLimit;
      }) == 2;
 
  // Checks if the scalars are used in other node.
  auto AreReusedScalars = [&](const TreeEntry *TE, ArrayRef<Value *> VL,
                              function_ref<bool(Value *)> CheckContainer) {
    return TE->isSame(VL) || all_of(VL, [&](Value *V) {
             if (isa<PoisonValue>(V))
               return true;
             auto *I = dyn_cast<Instruction>(V);
             if (!I)
               return false;
             return is_contained(TE->Scalars, I) || CheckContainer(I);
           });
  };
  auto CheckForSameVectorNodes = [&](const TreeEntry &E) {
    if (E.hasState()) {
      if (ArrayRef<TreeEntry *> TEs = getTreeEntries(E.getMainOp());
          !TEs.empty() && any_of(TEs, [&](const TreeEntry *TE) {
            return AreReusedScalars(TE, E.Scalars, [&](Value *V) {
              ArrayRef<TreeEntry *> VTEs = getTreeEntries(V);
              return !VTEs.empty() && any_of(VTEs, [&](const TreeEntry *TE) {
                return is_contained(TEs, TE);
              });
            });
          }))
        return true;
      ;
      if (ArrayRef<TreeEntry *> TEs = getSplitTreeEntries(E.getMainOp());
          !TEs.empty() && any_of(TEs, [&](const TreeEntry *TE) {
            return AreReusedScalars(TE, E.Scalars, [&](Value *V) {
              ArrayRef<TreeEntry *> VTEs = getSplitTreeEntries(V);
              return !VTEs.empty() && any_of(VTEs, [&](const TreeEntry *TE) {
                return is_contained(TEs, TE);
              });
            });
          }))
        return true;
    } else {
      // Check if the gather node full copy of split node.
      auto *It = find_if(E.Scalars, IsaPred<Instruction>);
      if (It != E.Scalars.end()) {
        if (ArrayRef<TreeEntry *> TEs = getSplitTreeEntries(*It);
            !TEs.empty() && any_of(TEs, [&](const TreeEntry *TE) {
              return AreReusedScalars(TE, E.Scalars, [&](Value *V) {
                ArrayRef<TreeEntry *> VTEs = getSplitTreeEntries(V);
                return !VTEs.empty() && any_of(VTEs, [&](const TreeEntry *TE) {
                  return is_contained(TEs, TE);
                });
              });
            }))
          return true;
      }
    }
    return false;
  };
  // The tree may grow here, so iterate over nodes, built before.
  for (unsigned Idx : seq<unsigned>(BaseGraphSize)) {
    TreeEntry &E = *VectorizableTree[Idx];
    if (E.isGather()) {
      ArrayRef<Value *> VL = E.Scalars;
      const unsigned Sz = getVectorElementSize(VL.front());
      unsigned MinVF = getMinVF(2 * Sz);
      // Do not try partial vectorization for small nodes (<= 2), nodes with the
      // same opcode and same parent block or all constants.
      if (VL.size() <= 2 || LoadEntriesToVectorize.contains(Idx) ||
          !(!E.hasState() || E.getOpcode() == Instruction::Load ||
            // We use allSameOpcode instead of isAltShuffle because we don't
            // want to use interchangeable instruction here.
            !allSameOpcode(VL) || !allSameBlock(VL)) ||
          allConstant(VL) || isSplat(VL))
        continue;
      if (ForceLoadGather && E.hasState() && E.getOpcode() == Instruction::Load)
        continue;
      // Check if the node is a copy of other vector nodes.
      if (CheckForSameVectorNodes(E))
        continue;
      // Try to find vectorizable sequences and transform them into a series of
      // insertvector instructions.
      unsigned StartIdx = 0;
      unsigned End = VL.size();
      SmallBitVector Processed(End);
      for (unsigned VF = getFloorFullVectorNumberOfElements(
               *TTI, VL.front()->getType(), VL.size() - 1, SLPReVec);
           VF >= MinVF; VF = getFloorFullVectorNumberOfElements(
                            *TTI, VL.front()->getType(), VF - 1, SLPReVec)) {
        if (StartIdx + VF > End)
          continue;
        SmallVector<std::pair<unsigned, unsigned>> Slices;
        bool AllStrided = true;
        // Walk the range in steps of VF, but allow the trailing slice to be
        // shorter (SliceVF < VF) so non-power-of-2 tails can be vectorized.
        // Processed only records slice starts (Cnt), and downstream consumers
        // (test(Cnt), AddCombinedNode's range set/StartIdx update) operate on
        // start positions, so partial coverage is consistent.
        for (unsigned Cnt = StartIdx; Cnt < End; Cnt += VF) {
          const unsigned SliceVF = std::min(VF, End - Cnt);
          if (SliceVF <= 1)
            continue;
          ArrayRef<Value *> Slice = VL.slice(Cnt, SliceVF);
          // If any instruction is vectorized already - do not try again.
          // Reuse the existing node, if it fully matches the slice.
          if ((Processed.test(Cnt) || isVectorized(Slice.front())) &&
              !getSameValuesTreeEntry(Slice.front(), Slice, /*SameVF=*/true))
            continue;
          // Constant already handled effectively - skip.
          if (allConstant(Slice))
            continue;
          // Do not try to vectorize small splats (less than vector register and
          // only with the single non-undef element).
          bool IsSplat = isSplat(Slice);
          bool IsTwoRegisterSplat = true;
          if (IsSplat && VF == 2) {
            unsigned NumRegs2VF = ::getNumberOfParts(
                *TTI,
                getWidenedType(getValueType(Slice.front(), SLPReVec), 2 * VF),
                getValueType(Slice.front(), SLPReVec));
            IsTwoRegisterSplat = NumRegs2VF == 2;
          }
          if (Slices.empty() || !IsSplat || !IsTwoRegisterSplat ||
              count(Slice, Slice.front()) ==
                  static_cast<long>(isa<UndefValue>(Slice.front()) ? SliceVF - 1
                                                                   : 1)) {
            if (IsSplat)
              continue;
            InstructionsState S = getSameOpcode(Slice, *TLI);
            if (!S || !allSameOpcode(Slice) || !allSameBlock(Slice) ||
                (S.getOpcode() == Instruction::Load &&
                 areKnownNonVectorizableLoads(Slice)))
              continue;
            if (VF == 2) {
              // Cache the cost check lazily - both branches below may need it.
              std::optional<bool> MainOpIsCheap;
              auto IsMainOpCheap = [&] {
                if (!MainOpIsCheap)
                  MainOpIsCheap =
                      TTI->getInstructionCost(S.getMainOp(), CostKind) <
                      TTI::TCC_Expensive;
                return *MainOpIsCheap;
              };
              // Try to vectorize reduced values or if all users are vectorized.
              // For expensive instructions extra extracts might be profitable.
              if ((!UserIgnoreList || E.Idx != 0) && IsMainOpCheap() &&
                  !all_of(Slice, [&](Value *V) {
                    if (isa<PoisonValue>(V))
                      return true;
                    return areAllUsersVectorized(cast<Instruction>(V),
                                                 UserIgnoreList);
                  }))
                continue;
              if (S.getOpcode() == Instruction::Load) {
                OrdersType Order;
                SmallVector<Value *> PointerOps;
                StridedPtrInfo SPtrInfo;
                LoadsState Res = canVectorizeLoads(Slice, Slice.front(), Order,
                                                   PointerOps, SPtrInfo);
                AllStrided &= Res == LoadsState::StridedVectorize ||
                              Res == LoadsState::ScatterVectorize ||
                              Res == LoadsState::Gather;
                // Do not vectorize gathers.
                if (Res == LoadsState::ScatterVectorize ||
                    Res == LoadsState::Gather) {
                  if (Res == LoadsState::Gather) {
                    registerNonVectorizableLoads(Slice);
                    // If reductions and the scalars from the root node are
                    // analyzed - mark as non-vectorizable reduction.
                    if (UserIgnoreList && E.Idx == 0)
                      analyzedReductionVals(Slice);
                  }
                  continue;
                }
              } else if (S.getOpcode() == Instruction::ExtractElement ||
                         (IsMainOpCheap() &&
                          !CheckOperandsProfitability(
                              S.getMainOp(),
                              cast<Instruction>(*find_if(reverse(Slice),
                                                         IsaPred<Instruction>)),
                              S))) {
                // Do not vectorize extractelements (handled effectively
                // alread). Do not vectorize non-profitable instructions (with
                // low cost and non-vectorizable operands.)
                continue;
              }
            }
          }
          Slices.emplace_back(Cnt, Slice.size());
        }
        // Do not try to vectorize if all slides are strided or gathered with
        // vector factor 2 and there are more than 2 slices. Better to handle
        // them in gathered loads analysis, may result in better vectorization.
        if (VF == 2 && AllStrided && Slices.size() > 2)
          continue;
        auto AddCombinedNode = [&](unsigned Idx, unsigned Cnt, unsigned Sz) {
          E.CombinedEntriesWithIndices.emplace_back(Idx, Cnt);
          Processed.set(Cnt, Cnt + Sz);
          if (StartIdx == Cnt)
            StartIdx = Cnt + Sz;
          if (End == Cnt + Sz)
            End = Cnt;
        };
        for (auto [Cnt, Sz] : Slices) {
          ArrayRef<Value *> Slice = VL.slice(Cnt, Sz);
          const TreeEntry *SameTE = nullptr;
          if (const auto *It = find_if(Slice, IsaPred<Instruction>);
              It != Slice.end()) {
            // If any instruction is vectorized already - do not try again.
            SameTE = getSameValuesTreeEntry(*It, Slice);
          }
          unsigned PrevSize = VectorizableTree.size();
          [[maybe_unused]] unsigned PrevEntriesSize =
              LoadEntriesToVectorize.size();
          buildTreeRec(Slice, 0, EdgeInfo(&E, UINT_MAX));
          if (PrevSize + 1 == VectorizableTree.size() && !SameTE &&
              VectorizableTree[PrevSize]->isGather() &&
              VectorizableTree[PrevSize]->hasState() &&
              VectorizableTree[PrevSize]->getOpcode() !=
                  Instruction::ExtractElement &&
              !isSplat(Slice)) {
            if (UserIgnoreList && E.Idx == 0 && VF == 2)
              analyzedReductionVals(Slice);
            VectorizableTree.pop_back();
            assert(PrevEntriesSize == LoadEntriesToVectorize.size() &&
                   "LoadEntriesToVectorize expected to remain the same");
            continue;
          }
          AddCombinedNode(PrevSize, Cnt, Sz);
        }
      }
      // Restore ordering, if no extra vectorization happened.
      if (E.CombinedEntriesWithIndices.empty() && !E.ReorderIndices.empty()) {
        SmallVector<int> Mask(E.ReorderIndices.begin(), E.ReorderIndices.end());
        reorderScalars(E.Scalars, Mask);
        E.ReorderIndices.clear();
      }
    }
    if (!E.hasState())
      continue;
    switch (E.getOpcode()) {
    case Instruction::Load: {
      // No need to reorder masked gather loads, just reorder the scalar
      // operands.
      if (E.State != TreeEntry::Vectorize)
        break;
      Type *ScalarTy = E.getMainOp()->getType();
      auto *VecTy =
          cast<FixedVectorType>(getWidenedType(ScalarTy, E.Scalars.size()));
      Align CommonAlignment = computeCommonAlignment<LoadInst>(E.Scalars);
      // Check if profitable to represent consecutive load + reverse as strided
      // load with stride -1.
      if (!E.ReorderIndices.empty() && isReverseOrder(E.ReorderIndices) &&
          TTI->isLegalStridedLoadStore(VecTy, CommonAlignment)) {
        SmallVector<int> Mask;
        inversePermutation(E.ReorderIndices, Mask);
        auto *BaseLI = cast<LoadInst>(E.Scalars.back());
        InstructionCost OriginalVecCost =
            TTI->getMemoryOpCost(Instruction::Load, VecTy, BaseLI->getAlign(),
                                 BaseLI->getPointerAddressSpace(), CostKind,
                                 TTI::OperandValueInfo()) +
            getShuffleCost(*TTI, TTI::SK_Reverse, VecTy, CostKind, Mask);
        InstructionCost StridedCost = TTI->getMemIntrinsicInstrCost(
            MemIntrinsicCostAttributes(Intrinsic::experimental_vp_strided_load,
                                       VecTy, BaseLI->getPointerOperand(),
                                       /*VariableMask=*/false, CommonAlignment,
                                       BaseLI),
            CostKind);
        if (StridedCost < OriginalVecCost || ForceStridedLoads) {
          // Strided load is more profitable than consecutive load + reverse -
          // transform the node to strided load.
          Type *StrideTy = DL->getIndexType(cast<LoadInst>(E.Scalars.front())
                                                ->getPointerOperand()
                                                ->getType());
          StridedPtrInfo SPtrInfo;
          SPtrInfo.StrideVal = ConstantInt::get(StrideTy, 1);
          SPtrInfo.Ty = VecTy;
          TreeEntryToStridedPtrInfoMap[&E] = SPtrInfo;
          E.State = TreeEntry::StridedVectorize;
        }
      }
      break;
    }
    case Instruction::Store: {
      if (E.State == TreeEntry::ExpandVectorize)
        break;
      Type *ScalarTy =
          cast<StoreInst>(E.getMainOp())->getValueOperand()->getType();
      auto *VecTy =
          cast<FixedVectorType>(getWidenedType(ScalarTy, E.Scalars.size()));
      Align CommonAlignment = computeCommonAlignment<StoreInst>(E.Scalars);
      // Check if profitable to represent consecutive load + reverse as strided
      // load with stride -1.
      if (!E.ReorderIndices.empty() && isReverseOrder(E.ReorderIndices) &&
          TTI->isLegalStridedLoadStore(VecTy, CommonAlignment)) {
        SmallVector<int> Mask;
        inversePermutation(E.ReorderIndices, Mask);
        auto *BaseSI = cast<StoreInst>(E.Scalars.back());
        InstructionCost OriginalVecCost =
            TTI->getMemoryOpCost(Instruction::Store, VecTy, BaseSI->getAlign(),
                                 BaseSI->getPointerAddressSpace(), CostKind,
                                 TTI::OperandValueInfo()) +
            getShuffleCost(*TTI, TTI::SK_Reverse, VecTy, CostKind, Mask);
        InstructionCost StridedCost = TTI->getMemIntrinsicInstrCost(
            MemIntrinsicCostAttributes(Intrinsic::experimental_vp_strided_store,
                                       VecTy, BaseSI->getPointerOperand(),
                                       /*VariableMask=*/false, CommonAlignment,
                                       BaseSI),
            CostKind);
        if (StridedCost < OriginalVecCost) {
          // Strided store is more profitable than reverse + consecutive store -
          // transform the node to strided store.
          E.State = TreeEntry::StridedVectorize;
          Type *StrideTy = DL->getIndexType(cast<StoreInst>(E.Scalars.front())
                                                ->getPointerOperand()
                                                ->getType());
          StridedPtrInfo SPtrInfo;
          SPtrInfo.StrideVal = ConstantInt::getSigned(StrideTy, -1);
          SPtrInfo.Ty = VecTy;
          TreeEntryToStridedPtrInfoMap[&E] = SPtrInfo;
        }
      } else if (!E.ReorderIndices.empty()) {
        // Check for interleaved stores.
        auto IsInterleaveMask = [&, &TTI = *TTI](ArrayRef<int> Mask) {
          auto *BaseSI = cast<StoreInst>(E.Scalars.front());
          assert(Mask.size() > 1 && "Expected mask greater than 1 element.");
          if (Mask.size() < 4)
            return 0u;
          for (unsigned Factor : seq<unsigned>(2, Mask.size() / 2 + 1)) {
            if (ShuffleVectorInst::isInterleaveMask(
                    Mask, Factor, VecTy->getElementCount().getFixedValue()) &&
                TTI.isLegalInterleavedAccessType(
                    VecTy, Factor, BaseSI->getAlign(),
                    BaseSI->getPointerAddressSpace()))
              return Factor;
          }
 
          return 0u;
        };
        SmallVector<int> Mask(E.ReorderIndices.begin(), E.ReorderIndices.end());
        unsigned InterleaveFactor = IsInterleaveMask(Mask);
        if (InterleaveFactor != 0)
          E.setInterleave(InterleaveFactor);
      }
      break;
    }
    case Instruction::Select: {
      if (E.State != TreeEntry::Vectorize)
        break;
      auto [MinMaxID, SelectOnly] = canConvertToMinOrMaxIntrinsic(E.Scalars);
      if (MinMaxID != Intrinsic::not_intrinsic) {
        // This node is a minmax node.
        E.CombinedOp = TreeEntry::MinMax;
        TreeEntry *CondEntry = getOperandEntry(&E, 0);
        if (SelectOnly && CondEntry->UserTreeIndex &&
            CondEntry->State == TreeEntry::Vectorize) {
          // The condition node is part of the combined minmax node.
          CondEntry->State = TreeEntry::CombinedVectorize;
        }
        break;
      }
      // Check for zext + selects, which can be reordered.
      SmallVector<unsigned> InversedCmpsIndices;
      if (matchesInversedZExtSelect(E, InversedCmpsIndices)) {
        auto *CmpTE = getOperandEntry(&E, 0);
        auto *Op1TE = getOperandEntry(&E, 1);
        auto *Op2TE = getOperandEntry(&E, 2);
        // State now is uniform, not alternate opcode.
        CmpTE->setOperations(
            InstructionsState(CmpTE->getMainOp(), CmpTE->getMainOp()));
        // Update mapping between the swapped values and their internal matching
        // nodes.
        auto UpdateGatherEntry = [&](TreeEntry *OldTE, TreeEntry *NewTE,
                                     Value *V) {
          if (isConstant(V))
            return;
          auto It = ValueToGatherNodes.find(V);
          assert(It != ValueToGatherNodes.end() &&
                 "Expected to find the value in the map.");
          auto &C = It->getSecond();
          if (!is_contained(OldTE->Scalars, V))
            C.remove(OldTE);
          C.insert(NewTE);
        };
        ValueList &Op1 = E.getOperand(1);
        ValueList &Op2 = E.getOperand(2);
        for (const unsigned Idx : InversedCmpsIndices) {
          Value *V1 = Op1TE->Scalars[Idx];
          Value *V2 = Op2TE->Scalars[Idx];
          std::swap(Op1TE->Scalars[Idx], Op2TE->Scalars[Idx]);
          std::swap(Op1[Idx], Op2[Idx]);
          UpdateGatherEntry(Op1TE, Op2TE, V1);
          UpdateGatherEntry(Op2TE, Op1TE, V2);
        }
        OperandsToTreeEntry.emplace_or_assign(std::make_pair(&E, 1), Op1TE);
        OperandsToTreeEntry.emplace_or_assign(std::make_pair(&E, 2), Op2TE);
        // NB: Fallback to check if select can be converted to cmp bitcast.
      }
      if (matchesSelectOfBits(E)) {
        // This node is a (reduced or) cmp bitcast node.
        const TreeEntry::CombinedOpcode Code = TreeEntry::ReducedCmpBitcast;
        E.CombinedOp = Code;
        auto *Op1TE = getOperandEntry(&E, 1);
        auto *Op2TE = getOperandEntry(&E, 2);
        Op1TE->State = TreeEntry::CombinedVectorize;
        Op1TE->CombinedOp = Code;
        Op2TE->State = TreeEntry::CombinedVectorize;
        Op2TE->CombinedOp = Code;
        break;
      }
      break;
    }
    case Instruction::FSub:
    case Instruction::FAdd: {
      // Check if possible to convert (a*b)+c to fma; fma takes exactly two
      // operands, so bail out on reassociated (>2-operand) nodes.
      if (E.State != TreeEntry::Vectorize ||
          !E.getOperations().isAddSubLikeOp() ||
          E.getOperations().isAltShuffle() || E.getNumOperands() != 2)
        break;
      const TreeEntry *LHS = getOperandEntry(&E, 0);
      const TreeEntry *RHS = getOperandEntry(&E, 1);
      auto IsOneUseVectorFMulOperand = [](const TreeEntry *TE) {
        return TE->State == TreeEntry::Vectorize &&
               TE->ReorderIndices.empty() && TE->ReuseShuffleIndices.empty() &&
               TE->getOpcode() == Instruction::FMul && !TE->isAltShuffle() &&
               all_of(TE->Scalars, [&](Value *V) {
                 return (TE->hasCopyableElements() &&
                         TE->isCopyableElement(V)) ||
                        V->hasOneUse();
               });
      };
      if (!IsOneUseVectorFMulOperand(LHS) &&
          (E.getOpcode() == Instruction::FSub ||
           !IsOneUseVectorFMulOperand(RHS)))
        break;
      if (!canConvertToFMA(E.Scalars, E.getOperations(), *DT, *DL, *TTI, *TLI,
                           CostKind)
               .isValid())
        break;
      // This node is a fmuladd node.
      E.CombinedOp = TreeEntry::FMulAdd;
      TreeEntry *FMulEntry = getOperandEntry(&E, 0);
      if (FMulEntry->UserTreeIndex &&
          FMulEntry->State == TreeEntry::Vectorize) {
        // The FMul node is part of the combined fmuladd node.
        FMulEntry->State = TreeEntry::CombinedVectorize;
      }
      break;
    }
    case Instruction::Shl: {
      // Shl is not reassociated; guard since this case indexes operands 0/1.
      if (E.getNumOperands() != 2)
        break;
      if (E.Idx != 0 || DL->isBigEndian())
        break;
      if (!UserIgnoreList)
        break;
      // Check that all reduction operands are disjoint or instructions.
      if (any_of(*UserIgnoreList, [](Value *V) {
            return !match(V, m_DisjointOr(m_Value(), m_Value()));
          }))
        break;
      OrdersType Order;
      bool IsBSwap;
      bool ForLoads;
      if (!matchesShlZExt(E, Order, IsBSwap, ForLoads))
        break;
      // This node is a (reduced disjoint or) bitcast node.
      TreeEntry::CombinedOpcode Code =
          IsBSwap ? (ForLoads ? TreeEntry::ReducedBitcastBSwapLoads
                              : TreeEntry::ReducedBitcastBSwap)
                  : (ForLoads ? TreeEntry::ReducedBitcastLoads
                              : TreeEntry::ReducedBitcast);
      E.CombinedOp = Code;
      E.ReorderIndices = std::move(Order);
      TreeEntry *ZExtEntry = getOperandEntry(&E, 0);
      assert(ZExtEntry->UserTreeIndex &&
             ZExtEntry->State == TreeEntry::Vectorize &&
             ZExtEntry->getOpcode() == Instruction::ZExt &&
             "Expected ZExt node.");
      // The ZExt node is part of the combined node.
      ZExtEntry->State = TreeEntry::CombinedVectorize;
      ZExtEntry->CombinedOp = Code;
      if (ForLoads) {
        TreeEntry *LoadsEntry = getOperandEntry(ZExtEntry, 0);
        assert(LoadsEntry->UserTreeIndex &&
               LoadsEntry->State == TreeEntry::Vectorize &&
               LoadsEntry->getOpcode() == Instruction::Load &&
               "Expected Load node.");
        // The Load node is part of the combined node.
        LoadsEntry->State = TreeEntry::CombinedVectorize;
        LoadsEntry->CombinedOp = Code;
      }
      TreeEntry *ConstEntry = getOperandEntry(&E, 1);
      assert(ConstEntry->UserTreeIndex && ConstEntry->isGather() &&
             "Expected ZExt node.");
      // The ConstNode node is part of the combined node.
      ConstEntry->State = TreeEntry::CombinedVectorize;
      ConstEntry->CombinedOp = Code;
      break;
    }
    default:
      break;
    }
  }
 
  if (LoadEntriesToVectorize.empty()) {
    // Single load node - exit.
    if (VectorizableTree.size() <= 1 && getRootNode().hasState() &&
        getRootNode().getOpcode() == Instruction::Load)
      return;
    // Small graph with small VF - exit.
    constexpr unsigned SmallTree = 3;
    constexpr unsigned SmallVF = 2;
    if ((VectorizableTree.size() <= SmallTree &&
         getRootNodeScalars().size() == SmallVF) ||
        (VectorizableTree.size() <= 2 && UserIgnoreList))
      return;
 
    if (getRootNode().isNonPowOf2Vec() &&
        getCanonicalGraphSize() != getTreeSize() && UserIgnoreList &&
        getCanonicalGraphSize() <= SmallTree &&
        count_if(ArrayRef(VectorizableTree).drop_front(getCanonicalGraphSize()),
                 [](const std::unique_ptr<TreeEntry> &TE) {
                   return TE->isGather() && TE->hasState() &&
                          TE->getOpcode() == Instruction::Load &&
                          !allSameBlock(TE->Scalars);
                 }) == 1)
      return;
  }
 
  // A list of loads to be gathered during the vectorization process. We can
  // try to vectorize them at the end, if profitable.
  SmallMapVector<std::tuple<BasicBlock *, Value *, Type *>,
                 SmallVector<SmallVector<std::pair<LoadInst *, int64_t>>>, 8>
      GatheredLoads;
 
  for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
    TreeEntry &E = *TE;
    if (E.isGather() &&
        ((E.hasState() && E.getOpcode() == Instruction::Load) ||
         (!E.hasState() && any_of(E.Scalars,
                                  [&](Value *V) {
                                    return isa<LoadInst>(V) &&
                                           !isVectorized(V) &&
                                           !isDeleted(cast<Instruction>(V));
                                  }))) &&
        !isSplat(E.Scalars)) {
      for (Value *V : E.Scalars) {
        auto *LI = dyn_cast<LoadInst>(V);
        if (!LI)
          continue;
        if (isDeleted(LI) || isVectorized(LI) || !LI->isSimple())
          continue;
        gatherPossiblyVectorizableLoads(
            *this, V, *DL, *SE, *TTI,
            GatheredLoads[std::make_tuple(
                LI->getParent(),
                getUnderlyingObject(LI->getPointerOperand(), RecursionMaxDepth),
                LI->getType())]);
      }
    }
  }
  // Try to vectorize gathered loads if this is not just a gather of loads.
  if (!GatheredLoads.empty())
    tryToVectorizeGatheredLoads(GatheredLoads);
}
 
/// Merges shuffle masks and emits final shuffle instruction, if required. It
/// supports shuffling of 2 input vectors. It implements lazy shuffles emission,
/// when the actual shuffle instruction is generated only if this is actually
/// required. Otherwise, the shuffle instruction emission is delayed till the
/// end of the process, to reduce the number of emitted instructions and further
/// analysis/transformations.
class BoUpSLP::ShuffleCostEstimator : public BaseShuffleAnalysis {
  bool IsFinalized = false;
  SmallVector<int> CommonMask;
  SmallVector<PointerUnion<Value *, const TreeEntry *>, 2> InVectors;
  /// Captures the original scalar VL of a single, "clean" gather() call so
  /// the values can be forwarded as the Args operand to getShuffleCost() for
  /// the final permutation in finalize(). This lets the target cost model
  /// recognize patterns such as broadcast-of-load (e.g. on X86,
  /// vbroadcast{ss,sd} folds the broadcast and the load into one instruction
  /// under AVX/AVX2 and is reported as TCC_Free by getShuffleCost). The
  /// state machine is:
  ///   * engaged + empty: tracking active, no qualifying gather seen yet.
  ///   * engaged + non-empty: exactly one qualifying gather observed and its
  ///     VL still corresponds to InVectors.front().
  ///   * disengaged: the cached VL is no longer trustworthy (multiple
  ///     gather() calls, or a state-mutating add() happened).
  std::optional<SmallVector<Value *>> BVValues = SmallVector<Value *>();
  const TargetTransformInfo &TTI;
  InstructionCost Cost = 0;
  SmallDenseSet<Value *> VectorizedVals;
  BoUpSLP &R;
  SmallPtrSetImpl<Value *> &CheckedExtracts;
  const TTI::TargetCostKind CostKind;
  /// While set, still trying to estimate the cost for the same nodes and we
  /// can delay actual cost estimation (virtual shuffle instruction emission).
  /// May help better estimate the cost if same nodes must be permuted + allows
  /// to move most of the long shuffles cost estimation to TTI.
  bool SameNodesEstimated = true;
 
  static Constant *getAllOnesValue(const DataLayout &DL, Type *Ty) {
    if (Ty->getScalarType()->isPointerTy()) {
      Constant *Res = ConstantExpr::getIntToPtr(
          ConstantInt::getAllOnesValue(
              IntegerType::get(Ty->getContext(),
                               DL.getTypeStoreSizeInBits(Ty->getScalarType()))),
          Ty->getScalarType());
      if (auto *VTy = dyn_cast<VectorType>(Ty))
        Res = ConstantVector::getSplat(VTy->getElementCount(), Res);
      return Res;
    }
    return Constant::getAllOnesValue(Ty);
  }
 
  InstructionCost getBuildVectorCost(ArrayRef<Value *> VL, Value *Root) {
    if ((!Root && allConstant(VL)) || all_of(VL, IsaPred<UndefValue>))
      return TTI::TCC_Free;
    auto *VecTy = cast<VectorType>(getWidenedType(ScalarTy, VL.size()));
    InstructionCost GatherCost = 0;
    SmallVector<Value *> Gathers(VL);
    if (!Root && isSplat(VL)) {
      // Found the broadcasting of the single scalar, calculate the cost as
      // the broadcast.
      const auto *It = find_if_not(VL, IsaPred<UndefValue>);
      assert(It != VL.end() && "Expected at least one non-undef value.");
      // Add broadcast for non-identity shuffle only.
      bool NeedShuffle =
          count(VL, *It) > 1 &&
          (VL.front() != *It || !all_of(VL.drop_front(), IsaPred<UndefValue>));
      if (!NeedShuffle) {
        if (isa<FixedVectorType>(ScalarTy)) {
          assert(SLPReVec && "FixedVectorType is not expected.");
          return TTI.getShuffleCost(
              TTI::SK_InsertSubvector, VecTy, VecTy, CostKind, {},
              std::distance(VL.begin(), It) * getNumElements(ScalarTy),
              cast<FixedVectorType>(ScalarTy));
        }
        return TTI.getVectorInstrCost(Instruction::InsertElement, VecTy,
                                      CostKind, std::distance(VL.begin(), It),
                                      PoisonValue::get(VecTy), *It);
      }
 
      SmallVector<int> ShuffleMask(VL.size(), PoisonMaskElem);
      transform(VL, ShuffleMask.begin(), [](Value *V) {
        return isa<PoisonValue>(V) ? PoisonMaskElem : 0;
      });
      InstructionCost InsertCost =
          TTI.getVectorInstrCost(Instruction::InsertElement, VecTy, CostKind, 0,
                                 PoisonValue::get(VecTy), *It);
      return InsertCost + getShuffleCost(TTI, TargetTransformInfo::SK_Broadcast,
                                         VecTy, CostKind, ShuffleMask,
                                         /*Index=*/0, /*SubTp=*/nullptr,
                                         /*Args=*/*It);
    }
    return GatherCost +
           (all_of(Gathers, IsaPred<UndefValue>)
                ? TTI::TCC_Free
                : R.getGatherCost(Gathers, !Root && VL.equals(Gathers),
                                  ScalarTy));
  };
 
  /// Compute the cost of creating a vector containing the extracted values from
  /// \p VL.
  InstructionCost
  computeExtractCost(ArrayRef<Value *> VL, ArrayRef<int> Mask,
                     ArrayRef<std::optional<TTI::ShuffleKind>> ShuffleKinds,
                     unsigned NumParts) {
    assert(VL.size() > NumParts && "Unexpected scalarized shuffle.");
    unsigned NumElts = accumulate(VL, 0, [](unsigned Sz, Value *V) {
      auto *EE = dyn_cast<ExtractElementInst>(V);
      if (!EE)
        return Sz;
      auto *VecTy = dyn_cast<FixedVectorType>(EE->getVectorOperandType());
      if (!VecTy)
        return Sz;
      return std::max(Sz, VecTy->getNumElements());
    });
    // FIXME: this must be moved to TTI for better estimation.
    unsigned EltsPerVector = getPartNumElems(VL.size(), NumParts);
    auto CheckPerRegistersShuffle = [&](MutableArrayRef<int> Mask,
                                        SmallVectorImpl<unsigned> &Indices,
                                        SmallVectorImpl<unsigned> &SubVecSizes)
        -> std::optional<TTI::ShuffleKind> {
      if (NumElts <= EltsPerVector)
        return std::nullopt;
      int OffsetReg0 = alignDown(accumulate(Mask, INT_MAX,
                                            [](int S, int I) {
                                              if (I == PoisonMaskElem)
                                                return S;
                                              return std::min(S, I);
                                            }),
                                 EltsPerVector);
      int OffsetReg1 = OffsetReg0;
      DenseSet<int> RegIndices;
      // Check that if trying to permute same single/2 input vectors.
      TTI::ShuffleKind ShuffleKind = TTI::SK_PermuteSingleSrc;
      int FirstRegId = -1;
      Indices.assign(1, OffsetReg0);
      for (auto [Pos, I] : enumerate(Mask)) {
        if (I == PoisonMaskElem)
          continue;
        int Idx = I - OffsetReg0;
        int RegId =
            (Idx / NumElts) * NumParts + (Idx % NumElts) / EltsPerVector;
        if (FirstRegId < 0)
          FirstRegId = RegId;
        RegIndices.insert(RegId);
        if (RegIndices.size() > 2)
          return std::nullopt;
        if (RegIndices.size() == 2) {
          ShuffleKind = TTI::SK_PermuteTwoSrc;
          if (Indices.size() == 1) {
            OffsetReg1 = alignDown(
                std::accumulate(
                    std::next(Mask.begin(), Pos), Mask.end(), INT_MAX,
                    [&](int S, int I) {
                      if (I == PoisonMaskElem)
                        return S;
                      int RegId = ((I - OffsetReg0) / NumElts) * NumParts +
                                  ((I - OffsetReg0) % NumElts) / EltsPerVector;
                      if (RegId == FirstRegId)
                        return S;
                      return std::min(S, I);
                    }),
                EltsPerVector);
            unsigned Index = OffsetReg1 % NumElts;
            Indices.push_back(Index);
            SubVecSizes.push_back(std::min(NumElts - Index, EltsPerVector));
          }
          Idx = I - OffsetReg1;
        }
        I = (Idx % NumElts) % EltsPerVector +
            (RegId == FirstRegId ? 0 : EltsPerVector);
      }
      return ShuffleKind;
    };
    InstructionCost Cost = 0;
 
    // Process extracts in blocks of EltsPerVector to check if the source vector
    // operand can be re-used directly. If not, add the cost of creating a
    // shuffle to extract the values into a vector register.
    for (unsigned Part : seq<unsigned>(NumParts)) {
      if (!ShuffleKinds[Part])
        continue;
      ArrayRef<int> MaskSlice = Mask.slice(
          Part * EltsPerVector, getNumElems(Mask.size(), EltsPerVector, Part));
      SmallVector<int> SubMask(EltsPerVector, PoisonMaskElem);
      copy(MaskSlice, SubMask.begin());
      SmallVector<unsigned, 2> Indices;
      SmallVector<unsigned, 2> SubVecSizes;
      std::optional<TTI::ShuffleKind> RegShuffleKind =
          CheckPerRegistersShuffle(SubMask, Indices, SubVecSizes);
      if (!RegShuffleKind) {
        if (*ShuffleKinds[Part] != TTI::SK_PermuteSingleSrc ||
            !ShuffleVectorInst::isIdentityMask(
                MaskSlice, std::max<unsigned>(NumElts, MaskSlice.size())))
          Cost += getShuffleCost(
              TTI, *ShuffleKinds[Part],
              cast<VectorType>(getWidenedType(ScalarTy, NumElts)), CostKind,
              MaskSlice);
        continue;
      }
      if (*RegShuffleKind != TTI::SK_PermuteSingleSrc ||
          !ShuffleVectorInst::isIdentityMask(SubMask, EltsPerVector)) {
        Cost += getShuffleCost(
            TTI, *RegShuffleKind,
            cast<VectorType>(getWidenedType(ScalarTy, EltsPerVector)), CostKind,
            SubMask);
      }
      const unsigned BaseVF = getFullVectorNumberOfElements(
          *R.TTI, VL.front()->getType(), alignTo(NumElts, EltsPerVector),
          SLPReVec);
      for (const auto [Idx, SubVecSize] : zip(Indices, SubVecSizes)) {
        assert((Idx + SubVecSize) <= BaseVF &&
               "SK_ExtractSubvector index out of range");
        Cost += getShuffleCost(
            TTI, TTI::SK_ExtractSubvector,
            cast<VectorType>(getWidenedType(ScalarTy, BaseVF)), CostKind, {},
            Idx, cast<VectorType>(getWidenedType(ScalarTy, SubVecSize)));
      }
      // Second attempt to check, if just a permute is better estimated than
      // subvector extract.
      SubMask.assign(NumElts, PoisonMaskElem);
      copy(MaskSlice, SubMask.begin());
      InstructionCost OriginalCost =
          getShuffleCost(TTI, *ShuffleKinds[Part],
                         cast<VectorType>(getWidenedType(ScalarTy, NumElts)),
                         CostKind, SubMask);
      if (OriginalCost < Cost)
        Cost = OriginalCost;
    }
    return Cost;
  }
  /// Adds the cost of reshuffling \p E1 and \p E2 (if present), using given
  /// mask \p Mask, register number \p Part, that includes \p SliceSize
  /// elements.
  void estimateNodesPermuteCost(const TreeEntry &E1, const TreeEntry *E2,
                                ArrayRef<int> Mask, unsigned Part,
                                unsigned SliceSize) {
    if (SameNodesEstimated) {
      // Delay the cost estimation if the same nodes are reshuffling.
      // If we already requested the cost of reshuffling of E1 and E2 before, no
      // need to estimate another cost with the sub-Mask, instead include this
      // sub-Mask into the CommonMask to estimate it later and avoid double cost
      // estimation.
      if ((InVectors.size() == 2 &&
           cast<const TreeEntry *>(InVectors.front()) == &E1 &&
           cast<const TreeEntry *>(InVectors.back()) == E2) ||
          (!E2 && cast<const TreeEntry *>(InVectors.front()) == &E1)) {
        unsigned Limit = getNumElems(Mask.size(), SliceSize, Part);
        assert(all_of(ArrayRef(CommonMask).slice(Part * SliceSize, Limit),
                      [](int Idx) { return Idx == PoisonMaskElem; }) &&
               "Expected all poisoned elements.");
        ArrayRef<int> SubMask = ArrayRef(Mask).slice(Part * SliceSize, Limit);
        copy(SubMask, std::next(CommonMask.begin(), SliceSize * Part));
        return;
      }
      // Found non-matching nodes - need to estimate the cost for the matched
      // and transform mask.
      Cost += createShuffle(InVectors.front(),
                            InVectors.size() == 1 ? nullptr : InVectors.back(),
                            CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
    } else if (InVectors.size() == 2) {
      Cost += createShuffle(InVectors.front(), InVectors.back(), CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
    }
    SameNodesEstimated = false;
    if (!E2 && InVectors.size() == 1) {
      unsigned VF = E1.getVectorFactor();
      if (Value *V1 = dyn_cast<Value *>(InVectors.front())) {
        VF = std::max(VF, getVF(V1));
      } else {
        const auto *E = cast<const TreeEntry *>(InVectors.front());
        VF = std::max(VF, E->getVectorFactor());
      }
      for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
        if (Mask[Idx] != PoisonMaskElem && CommonMask[Idx] == PoisonMaskElem)
          CommonMask[Idx] = Mask[Idx] + VF;
      Cost += createShuffle(InVectors.front(), &E1, CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
    } else {
      auto P = InVectors.front();
      Cost += createShuffle(&E1, E2, Mask);
      unsigned VF = Mask.size();
      if (Value *V1 = dyn_cast<Value *>(P)) {
        VF = std::max(VF,
                      getNumElements(V1->getType()));
      } else {
        const auto *E = cast<const TreeEntry *>(P);
        VF = std::max(VF, E->getVectorFactor());
      }
      for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
        if (Mask[Idx] != PoisonMaskElem)
          CommonMask[Idx] = Idx + (InVectors.empty() ? 0 : VF);
      Cost += createShuffle(P, InVectors.front(), CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
    }
  }
 
  class ShuffleCostBuilder {
    const TargetTransformInfo &TTI;
    const TTI::TargetCostKind CostKind;
 
    static bool isEmptyOrIdentity(ArrayRef<int> Mask, unsigned VF) {
      int Index = -1;
      return Mask.empty() ||
             (VF == Mask.size() &&
              ShuffleVectorInst::isIdentityMask(Mask, VF)) ||
             (ShuffleVectorInst::isExtractSubvectorMask(Mask, VF, Index) &&
              Index == 0);
    }
 
  public:
    ShuffleCostBuilder(const TargetTransformInfo &TTI,
                       const TTI::TargetCostKind CostKind)
        : TTI(TTI), CostKind(CostKind) {}
    ~ShuffleCostBuilder() = default;
    InstructionCost createShuffleVector(Value *V1, Value *,
                                        ArrayRef<int> Mask) const {
      // Empty mask or identity mask are free.
      unsigned VF =
          cast<VectorType>(V1->getType())->getElementCount().getKnownMinValue();
      if (isEmptyOrIdentity(Mask, VF))
        return TTI::TCC_Free;
      return getShuffleCost(TTI, TTI::SK_PermuteTwoSrc,
                            cast<VectorType>(V1->getType()), CostKind, Mask);
    }
    InstructionCost createShuffleVector(Value *V1, ArrayRef<int> Mask,
                                        ArrayRef<Value *> VL) const {
      // Empty mask or identity mask are free.
      unsigned VF =
          cast<VectorType>(V1->getType())->getElementCount().getKnownMinValue();
      if (isEmptyOrIdentity(Mask, VF))
        return TTI::TCC_Free;
      return getShuffleCost(TTI, TTI::SK_PermuteSingleSrc,
                            cast<VectorType>(V1->getType()), CostKind, Mask,
                            /*Index=*/0, /*SubTp=*/nullptr, VL);
    }
    InstructionCost createIdentity(Value *) const { return TTI::TCC_Free; }
    InstructionCost createPoison(Type *Ty, unsigned VF) const {
      return TTI::TCC_Free;
    }
    void resizeToMatch(Value *&, Value *&) const {}
  };
 
  /// Smart shuffle instruction emission, walks through shuffles trees and
  /// tries to find the best matching vector for the actual shuffle
  /// instruction.
  InstructionCost
  createShuffle(const PointerUnion<Value *, const TreeEntry *> &P1,
                const PointerUnion<Value *, const TreeEntry *> &P2,
                ArrayRef<int> Mask, ArrayRef<Value *> VL = {}) {
    ShuffleCostBuilder Builder(TTI, CostKind);
    SmallVector<int> CommonMask(Mask);
    Value *V1 = P1.dyn_cast<Value *>(), *V2 = P2.dyn_cast<Value *>();
    unsigned CommonVF = Mask.size();
    InstructionCost ExtraCost = 0;
    auto GetNodeMinBWAffectedCost = [&](const TreeEntry &E,
                                        unsigned VF) -> InstructionCost {
      if (E.isGather() && allConstant(E.Scalars))
        return TTI::TCC_Free;
      Type *EScalarTy = E.Scalars.front()->getType();
      bool IsSigned = true;
      if (auto It = R.MinBWs.find(&E); It != R.MinBWs.end()) {
        EScalarTy = IntegerType::get(EScalarTy->getContext(), It->second.first);
        IsSigned = It->second.second;
      }
      if (EScalarTy != ScalarTy) {
        unsigned CastOpcode = Instruction::Trunc;
        unsigned DstSz = R.DL->getTypeSizeInBits(ScalarTy);
        unsigned SrcSz = R.DL->getTypeSizeInBits(EScalarTy);
        if (DstSz > SrcSz)
          CastOpcode = IsSigned ? Instruction::SExt : Instruction::ZExt;
        return TTI.getCastInstrCost(CastOpcode, getWidenedType(ScalarTy, VF),
                                    getWidenedType(EScalarTy, VF),
                                    TTI::CastContextHint::None, CostKind);
      }
      return TTI::TCC_Free;
    };
    auto GetValueMinBWAffectedCost = [&](const Value *V) -> InstructionCost {
      if (isa<Constant>(V))
        return TTI::TCC_Free;
      auto *VecTy = cast<VectorType>(V->getType());
      Type *EScalarTy = VecTy->getElementType();
      if (EScalarTy != ScalarTy) {
        bool IsSigned = !isKnownNonNegative(V, SimplifyQuery(*R.DL));
        unsigned CastOpcode = Instruction::Trunc;
        unsigned DstSz = R.DL->getTypeSizeInBits(ScalarTy);
        unsigned SrcSz = R.DL->getTypeSizeInBits(EScalarTy);
        if (DstSz > SrcSz)
          CastOpcode = IsSigned ? Instruction::SExt : Instruction::ZExt;
        return TTI.getCastInstrCost(
            CastOpcode, VectorType::get(ScalarTy, VecTy->getElementCount()),
            VecTy, TTI::CastContextHint::None, CostKind);
      }
      return TTI::TCC_Free;
    };
    if (!V1 && !V2 && !P2.isNull()) {
      // Shuffle 2 entry nodes.
      const TreeEntry *E = cast<const TreeEntry *>(P1);
      unsigned VF = E->getVectorFactor();
      const TreeEntry *E2 = cast<const TreeEntry *>(P2);
      CommonVF = std::max(VF, E2->getVectorFactor());
      assert(all_of(Mask,
                    [=](int Idx) {
                      return Idx < 2 * static_cast<int>(CommonVF);
                    }) &&
             "All elements in mask must be less than 2 * CommonVF.");
      if (E->Scalars.size() == E2->Scalars.size()) {
        SmallVector<int> EMask = E->getCommonMask();
        SmallVector<int> E2Mask = E2->getCommonMask();
        if (!EMask.empty() || !E2Mask.empty()) {
          for (int &Idx : CommonMask) {
            if (Idx == PoisonMaskElem)
              continue;
            if (Idx < static_cast<int>(CommonVF) && !EMask.empty())
              Idx = EMask[Idx];
            else if (Idx >= static_cast<int>(CommonVF))
              Idx = (E2Mask.empty() ? Idx - CommonVF : E2Mask[Idx - CommonVF]) +
                    E->Scalars.size();
          }
        }
        CommonVF = E->Scalars.size();
        ExtraCost += GetNodeMinBWAffectedCost(*E, CommonVF) +
                     GetNodeMinBWAffectedCost(*E2, CommonVF);
      } else {
        ExtraCost += GetNodeMinBWAffectedCost(*E, E->getVectorFactor()) +
                     GetNodeMinBWAffectedCost(*E2, E2->getVectorFactor());
      }
      V1 = Constant::getNullValue(getWidenedType(ScalarTy, CommonVF));
      V2 = getAllOnesValue(*R.DL, getWidenedType(ScalarTy, CommonVF));
    } else if (!V1 && P2.isNull()) {
      // Shuffle single entry node.
      const TreeEntry *E = cast<const TreeEntry *>(P1);
      unsigned VF = E->getVectorFactor();
      CommonVF = VF;
      assert(
          all_of(Mask,
                 [=](int Idx) { return Idx < static_cast<int>(CommonVF); }) &&
          "All elements in mask must be less than CommonVF.");
      if (E->Scalars.size() == Mask.size() && VF != Mask.size()) {
        SmallVector<int> EMask = E->getCommonMask();
        assert(!EMask.empty() && "Expected non-empty common mask.");
        for (int &Idx : CommonMask) {
          if (Idx != PoisonMaskElem)
            Idx = EMask[Idx];
        }
        CommonVF = E->Scalars.size();
      } else if (unsigned Factor = E->getInterleaveFactor();
                 Factor > 0 && E->Scalars.size() != Mask.size() &&
                 ShuffleVectorInst::isDeInterleaveMaskOfFactor(CommonMask,
                                                               Factor)) {
        // Deinterleaved nodes are free.
        std::iota(CommonMask.begin(), CommonMask.end(), 0);
      }
      ExtraCost += GetNodeMinBWAffectedCost(*E, CommonVF);
      V1 = Constant::getNullValue(getWidenedType(ScalarTy, CommonVF));
      // Not identity/broadcast? Try to see if the original vector is better.
      if (!E->ReorderIndices.empty() && CommonVF == E->ReorderIndices.size() &&
          CommonVF == CommonMask.size() &&
          any_of(enumerate(CommonMask),
                 [](const auto &&P) {
                   return P.value() != PoisonMaskElem &&
                          static_cast<unsigned>(P.value()) != P.index();
                 }) &&
          any_of(CommonMask,
                 [](int Idx) { return Idx != PoisonMaskElem && Idx != 0; })) {
        SmallVector<int> ReorderMask;
        inversePermutation(E->ReorderIndices, ReorderMask);
        addMask(CommonMask, ReorderMask);
      }
    } else if (V1 && P2.isNull()) {
      // Shuffle single vector.
      ExtraCost += GetValueMinBWAffectedCost(V1);
      CommonVF = getVF(V1);
      assert(
          all_of(Mask,
                 [=](int Idx) { return Idx < static_cast<int>(CommonVF); }) &&
          "All elements in mask must be less than CommonVF.");
    } else if (V1 && !V2) {
      // Shuffle vector and tree node.
      unsigned VF = getVF(V1);
      const TreeEntry *E2 = cast<const TreeEntry *>(P2);
      CommonVF = std::max(VF, E2->getVectorFactor());
      assert(all_of(Mask,
                    [=](int Idx) {
                      return Idx < 2 * static_cast<int>(CommonVF);
                    }) &&
             "All elements in mask must be less than 2 * CommonVF.");
      if (E2->Scalars.size() == VF && VF != CommonVF) {
        SmallVector<int> E2Mask = E2->getCommonMask();
        assert(!E2Mask.empty() && "Expected non-empty common mask.");
        for (int &Idx : CommonMask) {
          if (Idx == PoisonMaskElem)
            continue;
          if (Idx >= static_cast<int>(CommonVF))
            Idx = E2Mask[Idx - CommonVF] + VF;
        }
        CommonVF = VF;
      }
      ExtraCost += GetValueMinBWAffectedCost(V1);
      V1 = Constant::getNullValue(getWidenedType(ScalarTy, CommonVF));
      ExtraCost += GetNodeMinBWAffectedCost(
          *E2, std::min(CommonVF, E2->getVectorFactor()));
      V2 = getAllOnesValue(*R.DL, getWidenedType(ScalarTy, CommonVF));
    } else if (!V1 && V2) {
      // Shuffle vector and tree node.
      unsigned VF = getVF(V2);
      const TreeEntry *E1 = cast<const TreeEntry *>(P1);
      CommonVF = std::max(VF, E1->getVectorFactor());
      assert(all_of(Mask,
                    [=](int Idx) {
                      return Idx < 2 * static_cast<int>(CommonVF);
                    }) &&
             "All elements in mask must be less than 2 * CommonVF.");
      if (E1->Scalars.size() == VF && VF != CommonVF) {
        SmallVector<int> E1Mask = E1->getCommonMask();
        assert(!E1Mask.empty() && "Expected non-empty common mask.");
        for (int &Idx : CommonMask) {
          if (Idx == PoisonMaskElem)
            continue;
          if (Idx >= static_cast<int>(CommonVF))
            Idx = E1Mask[Idx - CommonVF] + VF;
          else
            Idx = E1Mask[Idx];
        }
        CommonVF = VF;
      }
      ExtraCost += GetNodeMinBWAffectedCost(
          *E1, std::min(CommonVF, E1->getVectorFactor()));
      V1 = Constant::getNullValue(getWidenedType(ScalarTy, CommonVF));
      ExtraCost += GetValueMinBWAffectedCost(V2);
      V2 = getAllOnesValue(*R.DL, getWidenedType(ScalarTy, CommonVF));
    } else {
      assert(V1 && V2 && "Expected both vectors.");
      unsigned VF = getVF(V1);
      CommonVF = std::max(VF, getVF(V2));
      assert(all_of(Mask,
                    [=](int Idx) {
                      return Idx < 2 * static_cast<int>(CommonVF);
                    }) &&
             "All elements in mask must be less than 2 * CommonVF.");
      ExtraCost +=
          GetValueMinBWAffectedCost(V1) + GetValueMinBWAffectedCost(V2);
      if (V1->getType() != V2->getType()) {
        V1 = Constant::getNullValue(getWidenedType(ScalarTy, CommonVF));
        V2 = getAllOnesValue(*R.DL, getWidenedType(ScalarTy, CommonVF));
      } else {
        if (cast<VectorType>(V1->getType())->getElementType() != ScalarTy)
          V1 = Constant::getNullValue(getWidenedType(ScalarTy, CommonVF));
        if (cast<VectorType>(V2->getType())->getElementType() != ScalarTy)
          V2 = getAllOnesValue(*R.DL, getWidenedType(ScalarTy, CommonVF));
      }
    }
    InVectors.front() =
        Constant::getNullValue(getWidenedType(ScalarTy, CommonMask.size()));
    if (InVectors.size() == 2)
      InVectors.pop_back();
    return ExtraCost + BaseShuffleAnalysis::createShuffle<InstructionCost>(
                           V1, V2, CommonMask, Builder, ScalarTy, VL);
  }
 
public:
  ShuffleCostEstimator(Type *ScalarTy, TargetTransformInfo &TTI,
                       ArrayRef<Value *> VectorizedVals, BoUpSLP &R,
                       SmallPtrSetImpl<Value *> &CheckedExtracts)
      : BaseShuffleAnalysis(ScalarTy), TTI(TTI),
        VectorizedVals(VectorizedVals.begin(), VectorizedVals.end()), R(R),
        CheckedExtracts(CheckedExtracts), CostKind(R.getCostKind()) {}
  Value *adjustExtracts(const TreeEntry *E, MutableArrayRef<int> Mask,
                        ArrayRef<std::optional<TTI::ShuffleKind>> ShuffleKinds,
                        unsigned NumParts, bool &UseVecBaseAsInput) {
    UseVecBaseAsInput = false;
    if (Mask.empty())
      return nullptr;
    Value *VecBase = nullptr;
    SmallVector<Value *> VL(E->Scalars.begin(), E->Scalars.end());
    if (!E->ReorderIndices.empty()) {
      SmallVector<int> ReorderMask(E->ReorderIndices.begin(),
                                   E->ReorderIndices.end());
      reorderScalars(VL, ReorderMask);
    }
    // Check if it can be considered reused if same extractelements were
    // vectorized already.
    bool PrevNodeFound = any_of(
        ArrayRef(R.VectorizableTree).take_front(E->Idx),
        [&](const std::unique_ptr<TreeEntry> &TE) {
          return ((TE->hasState() && !TE->isAltShuffle() &&
                   TE->getOpcode() == Instruction::ExtractElement) ||
                  TE->isGather()) &&
                 all_of(enumerate(TE->Scalars), [&](auto &&Data) {
                   return VL.size() > Data.index() &&
                          (Mask[Data.index()] == PoisonMaskElem ||
                           isa<UndefValue>(VL[Data.index()]) ||
                           Data.value() == VL[Data.index()]);
                 });
        });
    SmallPtrSet<Value *, 4> UniqueBases;
    unsigned SliceSize = getPartNumElems(VL.size(), NumParts);
    SmallDenseMap<Value *, APInt, 4> VectorOpsToExtracts;
    for (unsigned Part : seq<unsigned>(NumParts)) {
      unsigned Limit = getNumElems(VL.size(), SliceSize, Part);
      ArrayRef<int> SubMask = Mask.slice(Part * SliceSize, Limit);
      for (auto [I, V] :
           enumerate(ArrayRef(VL).slice(Part * SliceSize, Limit))) {
        // Ignore non-extractelement scalars.
        if (isa<UndefValue>(V) ||
            (!SubMask.empty() && SubMask[I] == PoisonMaskElem))
          continue;
        // If all users of instruction are going to be vectorized and this
        // instruction itself is not going to be vectorized, consider this
        // instruction as dead and remove its cost from the final cost of the
        // vectorized tree.
        // Also, avoid adjusting the cost for extractelements with multiple uses
        // in different graph entries.
        auto *EE = cast<ExtractElementInst>(V);
        VecBase = EE->getVectorOperand();
        UniqueBases.insert(VecBase);
        ArrayRef<TreeEntry *> VEs = R.getTreeEntries(V);
        if (!CheckedExtracts.insert(V).second ||
            !R.areAllUsersVectorized(cast<Instruction>(V), &VectorizedVals) ||
            any_of(VEs,
                   [&](const TreeEntry *TE) {
                     return R.DeletedNodes.contains(TE) ||
                            R.TransformedToGatherNodes.contains(TE);
                   }) ||
            (E->UserTreeIndex && E->UserTreeIndex.EdgeIdx == UINT_MAX &&
             !R.isVectorized(EE) &&
             count_if(E->Scalars, [&](Value *V) { return V == EE; }) !=
                 count_if(E->UserTreeIndex.UserTE->Scalars,
                          [&](Value *V) { return V == EE; })) ||
            any_of(EE->users(),
                   [&](User *U) {
                     return isa<GetElementPtrInst>(U) &&
                            !R.areAllUsersVectorized(cast<Instruction>(U),
                                                     &VectorizedVals);
                   }) ||
            (!VEs.empty() && !is_contained(VEs, E)))
          continue;
        std::optional<unsigned> EEIdx = getExtractIndex(EE);
        if (!EEIdx)
          continue;
        unsigned Idx = *EEIdx;
        // Take credit for instruction that will become dead.
        if (EE->hasOneUse() || !PrevNodeFound) {
          Instruction *Ext = EE->user_back();
          if (isa<SExtInst, ZExtInst>(Ext) &&
              all_of(Ext->users(), IsaPred<GetElementPtrInst>)) {
            // Use getExtractWithExtendCost() to calculate the cost of
            // extractelement/ext pair.
            Cost -= TTI.getExtractWithExtendCost(
                Ext->getOpcode(), Ext->getType(), EE->getVectorOperandType(),
                Idx, CostKind);
            // Add back the cost of s|zext which is subtracted separately.
            Cost += TTI.getCastInstrCost(
                Ext->getOpcode(), Ext->getType(), EE->getType(),
                TTI::getCastContextHint(Ext), CostKind, Ext);
            continue;
          }
        }
        APInt &DemandedElts =
            VectorOpsToExtracts
                .try_emplace(VecBase,
                             APInt::getZero(getNumElements(VecBase->getType())))
                .first->getSecond();
        DemandedElts.setBit(Idx);
      }
    }
    for (const auto &[Vec, DemandedElts] : VectorOpsToExtracts)
      Cost -= TTI.getScalarizationOverhead(cast<VectorType>(Vec->getType()),
                                           DemandedElts, /*Insert=*/false,
                                           /*Extract=*/true, CostKind);
    // Check that gather of extractelements can be represented as just a
    // shuffle of a single/two vectors the scalars are extracted from.
    // Found the bunch of extractelement instructions that must be gathered
    // into a vector and can be represented as a permutation elements in a
    // single input vector or of 2 input vectors.
    // Done for reused if same extractelements were vectorized already.
    if (!PrevNodeFound)
      Cost += computeExtractCost(VL, Mask, ShuffleKinds, NumParts);
    InVectors.assign(1, E);
    CommonMask.assign(Mask.begin(), Mask.end());
    transformMaskAfterShuffle(CommonMask, CommonMask);
    SameNodesEstimated = false;
    if (NumParts != 1 && UniqueBases.size() != 1) {
      UseVecBaseAsInput = true;
      VecBase =
          Constant::getNullValue(getWidenedType(ScalarTy, CommonMask.size()));
    }
    return VecBase;
  }
  /// Checks if the specified entry \p E needs to be delayed because of its
  /// dependency nodes.
  std::optional<InstructionCost>
  needToDelay(const TreeEntry *,
              ArrayRef<SmallVector<const TreeEntry *>>) const {
    // No need to delay the cost estimation during analysis.
    return std::nullopt;
  }
  /// Reset the builder to handle perfect diamond match.
  void resetForSameNode() {
    IsFinalized = false;
    CommonMask.clear();
    InVectors.clear();
    Cost = 0;
    VectorizedVals.clear();
    SameNodesEstimated = true;
  }
  void add(const TreeEntry &E1, const TreeEntry &E2, ArrayRef<int> Mask) {
    BVValues.reset();
    if (&E1 == &E2) {
      assert(all_of(Mask,
                    [&](int Idx) {
                      return Idx < static_cast<int>(E1.getVectorFactor());
                    }) &&
             "Expected single vector shuffle mask.");
      add(E1, Mask);
      return;
    }
    if (InVectors.empty()) {
      CommonMask.assign(Mask.begin(), Mask.end());
      InVectors.assign({&E1, &E2});
      return;
    }
    assert(!CommonMask.empty() && "Expected non-empty common mask.");
    auto *MaskVecTy = getWidenedType(ScalarTy, Mask.size());
    unsigned NumParts =
        ::getNumberOfParts(TTI, MaskVecTy, ScalarTy, Mask.size());
    unsigned SliceSize = getPartNumElems(Mask.size(), NumParts);
    const auto *It = find_if(Mask, not_equal_to(PoisonMaskElem));
    unsigned Part = std::distance(Mask.begin(), It) / SliceSize;
    estimateNodesPermuteCost(E1, &E2, Mask, Part, SliceSize);
  }
  void add(const TreeEntry &E1, ArrayRef<int> Mask) {
    BVValues.reset();
    if (InVectors.empty()) {
      CommonMask.assign(Mask.begin(), Mask.end());
      InVectors.assign(1, &E1);
      return;
    }
    assert(!CommonMask.empty() && "Expected non-empty common mask.");
    auto *MaskVecTy = getWidenedType(ScalarTy, Mask.size());
    unsigned NumParts =
        ::getNumberOfParts(TTI, MaskVecTy, ScalarTy, Mask.size());
    unsigned SliceSize = getPartNumElems(Mask.size(), NumParts);
    const auto *It = find_if(Mask, not_equal_to(PoisonMaskElem));
    unsigned Part = std::distance(Mask.begin(), It) / SliceSize;
    estimateNodesPermuteCost(E1, nullptr, Mask, Part, SliceSize);
    if (!SameNodesEstimated && InVectors.size() == 1)
      InVectors.emplace_back(&E1);
  }
  /// Adds 2 input vectors and the mask for their shuffling.
  void add(Value *V1, Value *V2, ArrayRef<int> Mask) {
    // May come only for shuffling of 2 vectors with extractelements, already
    // handled in adjustExtracts.
    assert(InVectors.size() == 1 &&
           all_of(enumerate(CommonMask),
                  [&](auto P) {
                    if (P.value() == PoisonMaskElem)
                      return Mask[P.index()] == PoisonMaskElem;
                    auto *EI = cast<ExtractElementInst>(
                        cast<const TreeEntry *>(InVectors.front())
                            ->getOrdered(P.index()));
                    return EI->getVectorOperand() == V1 ||
                           EI->getVectorOperand() == V2;
                  }) &&
           "Expected extractelement vectors.");
  }
  /// Adds another one input vector and the mask for the shuffling.
  void add(Value *V1, ArrayRef<int> Mask, bool ForExtracts = false) {
    if (BVValues && !isa<Constant>(V1))
      BVValues.reset();
    if (InVectors.empty()) {
      assert(CommonMask.empty() && !ForExtracts &&
             "Expected empty input mask/vectors.");
      CommonMask.assign(Mask.begin(), Mask.end());
      InVectors.assign(1, V1);
      return;
    }
    if (ForExtracts) {
      // No need to add vectors here, already handled them in adjustExtracts.
      assert(InVectors.size() == 1 && isa<const TreeEntry *>(InVectors[0]) &&
             !CommonMask.empty() &&
             all_of(enumerate(CommonMask),
                    [&](auto P) {
                      Value *Scalar = cast<const TreeEntry *>(InVectors[0])
                                          ->getOrdered(P.index());
                      if (P.value() == PoisonMaskElem)
                        return P.value() == Mask[P.index()] ||
                               isa<UndefValue>(Scalar);
                      if (isa<Constant>(V1))
                        return true;
                      auto *EI = cast<ExtractElementInst>(Scalar);
                      return EI->getVectorOperand() == V1;
                    }) &&
             "Expected only tree entry for extractelement vectors.");
      return;
    }
    assert(!InVectors.empty() && !CommonMask.empty() &&
           "Expected only tree entries from extracts/reused buildvectors.");
    unsigned VF = getVF(V1);
    if (InVectors.size() == 2) {
      Cost += createShuffle(InVectors.front(), InVectors.back(), CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
      VF = std::max<unsigned>(VF, CommonMask.size());
    } else if (const auto *InTE =
                   InVectors.front().dyn_cast<const TreeEntry *>()) {
      VF = std::max(VF, InTE->getVectorFactor());
    } else {
      VF = std::max(
          VF, cast<FixedVectorType>(cast<Value *>(InVectors.front())->getType())
                  ->getNumElements());
    }
    InVectors.push_back(V1);
    for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
      if (Mask[Idx] != PoisonMaskElem && CommonMask[Idx] == PoisonMaskElem)
        CommonMask[Idx] = Mask[Idx] + VF;
  }
  Value *gather(ArrayRef<Value *> VL, unsigned MaskVF = 0,
                Value *Root = nullptr) {
    Cost += getBuildVectorCost(VL, Root);
    if (BVValues) {
      if (BVValues->empty() && InVectors.empty())
        BVValues->assign(VL.begin(), VL.end());
      else
        BVValues.reset();
    }
    if (!Root) {
      // FIXME: Need to find a way to avoid use of getNullValue here.
      SmallVector<Constant *> Vals;
      unsigned VF = VL.size();
      if (MaskVF != 0)
        VF = std::min(VF, MaskVF);
      Type *VLScalarTy = VL.front()->getType();
      for (Value *V : VL.take_front(VF)) {
        Type *ScalarTy = VLScalarTy->getScalarType();
        if (isa<PoisonValue>(V)) {
          Vals.push_back(PoisonValue::get(ScalarTy));
          continue;
        }
        if (isa<UndefValue>(V)) {
          Vals.push_back(UndefValue::get(ScalarTy));
          continue;
        }
        Vals.push_back(Constant::getNullValue(ScalarTy));
      }
      if (auto *VecTy = dyn_cast<FixedVectorType>(VLScalarTy)) {
        assert(SLPReVec && "FixedVectorType is not expected.");
        // When REVEC is enabled, we need to expand vector types into scalar
        // types.
        Vals = replicateMask(Vals, VecTy->getNumElements());
      }
      return ConstantVector::get(Vals);
    }
    return ConstantVector::getSplat(
        ElementCount::getFixed(
            cast<FixedVectorType>(Root->getType())->getNumElements()),
        getAllOnesValue(*R.DL, ScalarTy->getScalarType()));
  }
  InstructionCost createFreeze(InstructionCost Cost) { return Cost; }
  /// Finalize emission of the shuffles.
  InstructionCost finalize(
      ArrayRef<int> ExtMask,
      ArrayRef<std::pair<const TreeEntry *, unsigned>> SubVectors,
      ArrayRef<int> SubVectorsMask, unsigned VF = 0,
      function_ref<void(Value *&, SmallVectorImpl<int> &,
                        function_ref<Value *(Value *, Value *, ArrayRef<int>)>)>
          Action = {}) {
    IsFinalized = true;
    if (Action) {
      const PointerUnion<Value *, const TreeEntry *> &Vec = InVectors.front();
      if (InVectors.size() == 2)
        Cost += createShuffle(Vec, InVectors.back(), CommonMask);
      else
        Cost += createShuffle(Vec, nullptr, CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
      assert(VF > 0 &&
             "Expected vector length for the final value before action.");
      Value *V = cast<Value *>(Vec);
      Action(V, CommonMask, [this](Value *V1, Value *V2, ArrayRef<int> Mask) {
        Cost += createShuffle(V1, V2, Mask);
        return V1;
      });
      InVectors.front() = V;
    }
    if (!SubVectors.empty()) {
      const PointerUnion<Value *, const TreeEntry *> &Vec = InVectors.front();
      if (InVectors.size() == 2)
        Cost += createShuffle(Vec, InVectors.back(), CommonMask);
      else
        Cost += createShuffle(Vec, nullptr, CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
      // Add subvectors permutation cost.
      if (!SubVectorsMask.empty()) {
        assert(SubVectorsMask.size() <= CommonMask.size() &&
               "Expected same size of masks for subvectors and common mask.");
        SmallVector<int> SVMask(CommonMask.size(), PoisonMaskElem);
        copy(SubVectorsMask, SVMask.begin());
        for (auto [I1, I2] : zip(SVMask, CommonMask)) {
          if (I2 != PoisonMaskElem) {
            assert(I1 == PoisonMaskElem && "Expected unused subvectors mask");
            I1 = I2 + CommonMask.size();
          }
        }
        Cost += getShuffleCost(
            TTI, TTI::SK_PermuteTwoSrc,
            cast<VectorType>(getWidenedType(ScalarTy, CommonMask.size())),
            CostKind, SVMask);
      }
      for (auto [E, Idx] : SubVectors) {
        Type *EScalarTy = E->Scalars.front()->getType();
        bool IsSigned = true;
        if (auto It = R.MinBWs.find(E); It != R.MinBWs.end()) {
          EScalarTy =
              IntegerType::get(EScalarTy->getContext(), It->second.first);
          IsSigned = It->second.second;
        }
        if (ScalarTy != EScalarTy) {
          unsigned CastOpcode = Instruction::Trunc;
          unsigned DstSz = R.DL->getTypeSizeInBits(ScalarTy);
          unsigned SrcSz = R.DL->getTypeSizeInBits(EScalarTy);
          if (DstSz > SrcSz)
            CastOpcode = IsSigned ? Instruction::SExt : Instruction::ZExt;
          Cost += TTI.getCastInstrCost(
              CastOpcode, getWidenedType(ScalarTy, E->getVectorFactor()),
              getWidenedType(EScalarTy, E->getVectorFactor()),
              TTI::CastContextHint::Normal, CostKind);
        }
        Cost += getShuffleCost(
            TTI, TTI::SK_InsertSubvector,
            cast<VectorType>(getWidenedType(ScalarTy, CommonMask.size())),
            CostKind, {}, Idx,
            cast<VectorType>(getWidenedType(ScalarTy, E->getVectorFactor())));
        if (!CommonMask.empty()) {
          std::iota(std::next(CommonMask.begin(), Idx),
                    std::next(CommonMask.begin(), Idx + E->getVectorFactor()),
                    Idx);
        }
      }
    }
 
    if (!ExtMask.empty()) {
      if (CommonMask.empty()) {
        CommonMask.assign(ExtMask.begin(), ExtMask.end());
      } else {
        SmallVector<int> NewMask(ExtMask.size(), PoisonMaskElem);
        for (int I = 0, Sz = ExtMask.size(); I < Sz; ++I) {
          if (ExtMask[I] == PoisonMaskElem)
            continue;
          NewMask[I] = CommonMask[ExtMask[I]];
        }
        CommonMask.swap(NewMask);
      }
    }
    if (CommonMask.empty()) {
      assert(InVectors.size() == 1 && "Expected only one vector with no mask");
      return Cost;
    }
    ArrayRef<Value *> VL;
    if (BVValues)
      VL = *BVValues;
    return Cost +
           createShuffle(InVectors.front(),
                         InVectors.size() == 2 ? InVectors.back() : nullptr,
                         CommonMask, VL);
  }
 
  ~ShuffleCostEstimator() {
    assert((IsFinalized || CommonMask.empty()) &&
           "Shuffle construction must be finalized.");
  }
};
 
const BoUpSLP::TreeEntry *BoUpSLP::getOperandEntry(const TreeEntry *E,
                                                   unsigned Idx) const {
  TreeEntry *Op = OperandsToTreeEntry.at({E, Idx});
  assert(Op->isSame(E->getOperand(Idx)) && "Operands mismatch!");
  return Op;
}
 
TTI::CastContextHint BoUpSLP::getCastContextHint(const TreeEntry &TE) const {
  if (TE.State == TreeEntry::ScatterVectorize ||
      TE.State == TreeEntry::StridedVectorize)
    return TTI::CastContextHint::GatherScatter;
  if (TE.State == TreeEntry::CompressVectorize ||
      TE.State == TreeEntry::BlendedLoadVectorize)
    return TTI::CastContextHint::Masked;
  if (TE.State == TreeEntry::Vectorize && TE.getOpcode() == Instruction::Load &&
      !TE.isAltShuffle()) {
    if (TE.ReorderIndices.empty())
      return TTI::CastContextHint::Normal;
    SmallVector<int> Mask;
    inversePermutation(TE.ReorderIndices, Mask);
    if (ShuffleVectorInst::isReverseMask(Mask, Mask.size()))
      return TTI::CastContextHint::Reversed;
  }
  return TTI::CastContextHint::None;
}
 
/// Get the assumed loop trip count for the loop \p L.
static unsigned getLoopTripCount(const Loop *L, ScalarEvolution &SE) {
  if (LoopAwareTripCount == 0)
    return 1;
  unsigned Scale = SE.getSmallConstantTripCount(L);
  if (Scale == 0)
    Scale = getLoopEstimatedTripCount(const_cast<Loop *>(L)).value_or(0);
  if (Scale != 0) {
    // Multiple exiting blocks - choose the minimum between trip count (scale)
    // and LoopAwareTripCount, since the multiple exit loops can be terminated
    // early.
    if (!L->getExitingBlock())
      return std::min<unsigned>(LoopAwareTripCount, Scale);
    return Scale;
  }
  return LoopAwareTripCount;
}
 
uint64_t BoUpSLP::getScaleToLoopIterations(const TreeEntry &TE, Value *Scalar,
                                           Instruction *U) {
  BasicBlock *Parent = nullptr;
  if (U) {
    // The extractelement for a PHI-node user is created in the incoming
    // block that feeds the matching operand, not in the PHI block itself
    // When the PHI is inside a loop that incoming block can belong to a deeper
    // loop than the PHI block. Scaling by the PHI block would use
    // the outer trip count instead of inner*outer, and because
    // ExtractCostCalculated deduplicates by scalar (only the first external
    // user fixes the scale) it would also make the cost depend on external-user
    // ordering. A PHI outside all loops is a plain loop-exit phi: its live-out
    // lanes are normally rebuilt as a vector LCSSA phi in the exit block, which
    // hoists the extract out of the loop, so scale = 1 (via U->getParent()
    // below) is kept and the adjustment is restricted to in-loop PHIs.
    if (auto *PHI = dyn_cast<PHINode>(U); PHI && Scalar) {
      if (LI->getLoopFor(PHI->getParent())) {
        // Use the deepest incoming block among all slots where Scalar
        // appears, to be conservative when the same value appears in
        // multiple predecessors.
        for (unsigned I : seq<unsigned>(PHI->getNumIncomingValues())) {
          if (PHI->getIncomingValue(I) != Scalar)
            continue;
          BasicBlock *InBB = PHI->getIncomingBlock(I);
          if (!Parent || LI->getLoopDepth(InBB) > LI->getLoopDepth(Parent))
            Parent = InBB;
        }
      }
    }
    if (!Parent)
      Parent = U->getParent();
  } else if (TE.isGather() || TE.State == TreeEntry::SplitVectorize) {
    EdgeInfo EI = TE.UserTreeIndex;
    while (EI.UserTE) {
      if (EI.UserTE->isGather() ||
          EI.UserTE->State == TreeEntry::SplitVectorize) {
        EI = EI.UserTE->UserTreeIndex;
        continue;
      }
      if (EI.UserTE->State == TreeEntry::Vectorize &&
          EI.UserTE->getOpcode() == Instruction::PHI) {
        auto *PH = cast<PHINode>(EI.UserTE->getMainOp());
        Parent = PH->getIncomingBlock(EI.EdgeIdx);
      } else {
        Parent = EI.UserTE->getMainOp()->getParent();
      }
      break;
    }
    if (!Parent)
      return 1;
  } else {
    Parent = TE.getMainOp()->getParent();
  }
  const Loop *L = LI->getLoopFor(Parent);
  if (!L)
    return 1;
  // The entry's cost is paid once per execution of the innermost loop in
  // which some of its operands are variant. Operands that are invariant in
  // all enclosing loops are executed once (LICM will hoist them out).
  return getLoopNestScale(findInnermostNonInvariantLoop(
      L, Scalar ? ArrayRef(Scalar) : ArrayRef(TE.Scalars)));
}
 
uint64_t BoUpSLP::getLoopNestScale(const Loop *L) {
  if (!L || LoopAwareTripCount == 0)
    return 1;
  if (auto It = LoopNestScaleCache.find(L); It != LoopNestScaleCache.end())
    return It->second;
  // Collect loops from L outward up to (but not including) the first cached
  // ancestor or the function top, then walk back inward multiplying trip
  // counts. Use uint64_t to avoid silent overflow on deep/large nests.
  SmallVector<const Loop *> Chain;
  for (const Loop *Cur = L; Cur; Cur = Cur->getParentLoop()) {
    if (LoopNestScaleCache.contains(Cur))
      break;
    Chain.push_back(Cur);
  }
  assert(!Chain.empty() && "Early-return above should have handled cache hit.");
  uint64_t Scale = 1;
  if (const Loop *Parent = Chain.back()->getParentLoop())
    Scale = LoopNestScaleCache.lookup(Parent);
  // Walk from the outermost uncached loop inward, accumulating trip counts.
  // Use SaturatingMultiply to clamp at uint64_t max on deep/large nests
  // rather than wrapping around.
  for (const Loop *Cur : reverse(Chain)) {
    uint64_t TC = std::max<uint64_t>(1, getLoopTripCount(Cur, *SE));
    Scale = SaturatingMultiply(Scale, TC);
    LoopNestScaleCache.try_emplace(Cur, std::max<uint64_t>(1, Scale));
  }
  return std::max<uint64_t>(1, Scale);
}
 
uint64_t BoUpSLP::getGatherNodeEffectiveScale(const TreeEntry &TE,
                                              Instruction *U) {
  // Only meaningful for gather/buildvector-like entries; the per-lane
  // insertelements that make up such an entry are LICM-hoistable by
  // optimizeGatherSequence() when their operand is loop-invariant.
  assert((TE.isGather() || TE.State == TreeEntry::SplitVectorize) &&
         "Expected gather/split tree entry.");
 
  uint64_t BaseScale = getScaleToLoopIterations(TE, nullptr, U);
  if (!PerLaneGatherScale || LoopAwareTripCount == 0 || BaseScale <= 1)
    return BaseScale;
 
  // Average the per-lane execution scales: for each lane, reuse the same
  // scale helper the rest of the cost model uses, but ask it about that
  // one lane's value. Lanes that are loop-invariant in the current nest
  // collapse to their outer-loop scale (or 1 for fully invariant/constant
  // lanes), which matches the LICM hoisting performed by
  // optimizeGatherSequence(). Cap per-lane contributions by BaseScale so a
  // refinement can never raise the cost above the whole-entry scale.
  // Each lane contributes at most BaseScale, so Sum is bounded above by
  // N * BaseScale. If BaseScale is near uint64_t max (saturated by
  // getLoopNestScale on a deep nest) Sum can still overflow uint64_t,
  // which would silently wrap and produce a wrong average. Use
  // SaturatingAdd and bail out to BaseScale on overflow: the true average
  // is bounded above by BaseScale anyway, so this preserves the
  // refinement's invariant that it can never raise cost.
  uint64_t Sum = 0;
  unsigned N = 0;
  bool Overflow = false;
  for (Value *V : TE.Scalars) {
    if (isConstant(V))
      continue;
    ++N;
    uint64_t LaneScale =
        std::min(getScaleToLoopIterations(TE, V, U), BaseScale);
    Sum = SaturatingAdd(Sum, LaneScale, &Overflow);
    if (Overflow)
      return BaseScale;
  }
  if (N == 0)
    return BaseScale;
  // Ceil-divide so we never round the effective scale down below 1.
  uint64_t Numerator = SaturatingAdd(Sum, uint64_t(N - 1), &Overflow);
  if (Overflow)
    return BaseScale;
  uint64_t Avg = Numerator / N;
  return std::clamp<uint64_t>(Avg, 1, BaseScale);
}
 
uint64_t BoUpSLP::getEntryEffectiveScale(const TreeEntry &TE, Instruction *U) {
  if (TE.isGather() || TE.State == TreeEntry::SplitVectorize)
    return getGatherNodeEffectiveScale(TE, U);
  return getScaleToLoopIterations(TE);
}
 
InstructionCost
BoUpSLP::getVectorSpillReloadCost(const TreeEntry *E, Type *ScalarTy,
                                  Type *VecTy, Type *FinalVecTy,
                                  const TTI::TargetCostKind CostKind) const {
  InstructionCost SpillsReloads = 0;
 
  // Estimate vector register pressure per target register class: operand
  // vectors plus the result. The same vector operand is counted once via
  // CountedOpEntries deduplication. PHIs take the max operand pressure across
  // incoming slots (only one predecessor is live at a time) plus the result.
  // All-constant operand bundles are skipped.
  if (!E->hasState() || E->getOpcode() == Instruction::Store ||
      E->getOpcode() == Instruction::ExtractElement ||
      E->getOpcode() == Instruction::ExtractValue ||
      E->getOpcode() == Instruction::Freeze ||
      (E->getOpcode() == Instruction::Load &&
       E->State != TreeEntry::ScatterVectorize))
    return SpillsReloads;
 
  const bool IsPHI =
      E->State == TreeEntry::Vectorize && E->getOpcode() == Instruction::PHI;
  SmallPtrSet<const TreeEntry *, 8> CountedOpEntries;
  SmallDenseMap<unsigned, unsigned> PressureByClass;
  auto AddPartsToClass = [&](unsigned RegClass, unsigned Parts) {
    assert(Parts != 0 && "Expected non-zero number of parts (registers).");
    PressureByClass[RegClass] += Parts;
  };
 
  auto GetEntryVecTy = [&](const TreeEntry *TE) -> std::pair<Type *, Type *> {
    Type *ScalarTy = getValueType(TE->Scalars.front(), SLPReVec);
    auto BWIt = MinBWs.find(TE);
    if (BWIt != MinBWs.end()) {
      auto *VTy = dyn_cast<FixedVectorType>(ScalarTy);
      ScalarTy = IntegerType::get(F->getContext(), BWIt->second.first);
      if (VTy)
        ScalarTy = getWidenedType(ScalarTy, VTy->getNumElements());
    }
    return std::make_pair(ScalarTy,
                          getWidenedType(ScalarTy, TE->getVectorFactor()));
  };
 
  if (E->State == TreeEntry::SplitVectorize) {
    for (const auto &[Idx, _] : E->CombinedEntriesWithIndices) {
      const TreeEntry *OpTE = VectorizableTree[Idx].get();
 
      if (!CountedOpEntries.insert(OpTE).second)
        continue;
      auto [ScalarTy, OpVecTy] = GetEntryVecTy(OpTE);
      const unsigned Parts = getNumberOfParts(OpVecTy, ScalarTy);
      if (Parts == 0)
        continue;
      const unsigned RC =
          TTI->getRegisterClassForType(/*Vector=*/true, OpVecTy);
      AddPartsToClass(RC, Parts);
    }
  } else if (IsPHI) {
    // Only one predecessor is live at a time - take the max operand pressure
    // across incoming slots.
    SmallDenseMap<unsigned, unsigned> MaxOpPressureByClass;
    for (unsigned Idx : seq<unsigned>(E->getNumOperands())) {
      const TreeEntry *OpTE = getOperandEntry(E, Idx);
      auto [ScalarTy, OpVecTy] = GetEntryVecTy(OpTE);
      const unsigned Parts = getNumberOfParts(OpVecTy, ScalarTy);
      if (Parts == 0)
        continue;
      const unsigned RC =
          TTI->getRegisterClassForType(/*Vector=*/true, OpVecTy);
      MaxOpPressureByClass[RC] = std::max(MaxOpPressureByClass[RC], Parts);
    }
    for (auto [RC, Parts] : MaxOpPressureByClass)
      AddPartsToClass(RC, Parts);
  } else {
    for (unsigned Idx : seq<unsigned>(E->getNumOperands())) {
      // InsertElement operand 0 is the vector being inserted into, which is
      // built incrementally and does not occupy an extra register.
      if ((E->getOpcode() == Instruction::InsertElement ||
           E->getOpcode() == Instruction::InsertValue) &&
          Idx == 0)
        continue;
      ArrayRef<Value *> Ops = E->getOperand(Idx);
      if (Ops.empty() || allConstant(Ops) || isSplat(Ops))
        continue;
      Value *Op = Ops.front();
      if (!Op)
        continue;
      const TreeEntry *OpTE = getOperandEntry(E, Idx);
 
      if (!CountedOpEntries.insert(OpTE).second)
        continue;
      auto *OpVecTy = getWidenedType(Op->getType(), Ops.size());
      const unsigned Parts = getNumberOfParts(OpVecTy, Op->getType());
      if (Parts == 0)
        continue;
      const unsigned RC =
          TTI->getRegisterClassForType(/*Vector=*/true, OpVecTy);
      AddPartsToClass(RC, Parts);
    }
  }
 
  if (E->getOpcode() != Instruction::Load) {
    const unsigned ResParts = getNumberOfParts(VecTy, ScalarTy);
    if (ResParts != 0) {
      const unsigned RC = TTI->getRegisterClassForType(/*Vector=*/true, VecTy);
      AddPartsToClass(RC, ResParts);
    }
    if (VecTy != FinalVecTy) {
      const unsigned FinalResParts = getNumberOfParts(FinalVecTy, ScalarTy);
      if (FinalResParts != 0) {
        const unsigned RC =
            TTI->getRegisterClassForType(/*Vector=*/true, FinalVecTy);
        AddPartsToClass(RC, FinalResParts);
      }
    }
  }
 
  for (auto [RegClass, UsedRegs] : PressureByClass) {
    const unsigned NumAvailRegs = TTI->getNumberOfRegisters(RegClass);
    if (NumAvailRegs == 0 || UsedRegs <= NumAvailRegs)
      continue;
    const unsigned SpillCount = UsedRegs - NumAvailRegs;
    InstructionCost SingleRegSpillReload =
        TTI->getRegisterClassReloadCost(RegClass, CostKind);
    // No need to spill cost only for the root entry (Idx == 0), for reduction
    // and non-returning instructions, like void calls.
    if (E->Idx > 0 || !UserIgnoreList || !E->Scalars[0]->getType()->isVoidTy())
      SingleRegSpillReload +=
          TTI->getRegisterClassSpillCost(RegClass, CostKind);
    SpillsReloads += SingleRegSpillReload * SpillCount;
  }
  return SpillsReloads;
}
 
/// Calculates a VectorInstrContext from the values in \p VL at indices in
/// \p DemandedElts.
static TTI::VectorInstrContext
getVectorInstrContextHint(ArrayRef<Value *> VL, const APInt &DemandedElts) {
  TTI::VectorInstrContext VIC = TTI::VectorInstrContext::None;
  for (unsigned I : seq(VL.size())) {
    if (!DemandedElts[I])
      continue;
    Value *V = VL[I];
    if (isa<UndefValue>(V))
      continue;
    auto *Inst = dyn_cast<Instruction>(V);
    if (!Inst)
      return TTI::VectorInstrContext::None;
    TTI::VectorInstrContext IVIC = TTI::getVectorInstrContextHint(Inst);
    if (VIC == TTI::VectorInstrContext::None) {
      VIC = IVIC;
      continue;
    }
    if (VIC != IVIC)
      return TTI::VectorInstrContext::None;
  }
  return VIC;
}
 
InstructionCost
BoUpSLP::getEntryCost(const TreeEntry *E, ArrayRef<Value *> VectorizedVals,
                      SmallPtrSetImpl<Value *> &CheckedExtracts) {
  ArrayRef<Value *> VL = E->Scalars;
 
  Type *ScalarTy = getValueType(VL[0], SLPReVec);
  if (SLPReVec && E->State == TreeEntry::Vectorize &&
      E->getOpcode() == Instruction::InsertElement &&
      !E->getOperand(1).back()->getType()->isVectorTy())
    ScalarTy = ScalarTy->getScalarType();
  if (!isValidElementType(ScalarTy, SLPReVec))
    return InstructionCost::getInvalid();
 
  // If we have computed a smaller type for the expression, update VecTy so
  // that the costs will be accurate.
  auto It = MinBWs.find(E);
  Type *OrigScalarTy = ScalarTy;
  if (It != MinBWs.end()) {
    auto *VecTy = dyn_cast<FixedVectorType>(ScalarTy);
    ScalarTy = IntegerType::get(F->getContext(), It->second.first);
    if (VecTy)
      ScalarTy = getWidenedType(ScalarTy, VecTy->getNumElements());
  } else if (E->Idx == 0 && isReducedBitcastRoot()) {
    const TreeEntry *ZExt = getOperandEntry(E, /*Idx=*/0);
    ScalarTy = cast<CastInst>(ZExt->getMainOp())->getSrcTy();
  }
  auto *VecTy = getWidenedType(ScalarTy, VL.size());
  unsigned EntryVF = E->getVectorFactor();
  auto *FinalVecTy = getWidenedType(ScalarTy, EntryVF);
 
  const InstructionCost SpillsReloads =
      getVectorSpillReloadCost(E, ScalarTy, VecTy, FinalVecTy, CostKind);
  if (E->isGather() || TransformedToGatherNodes.contains(E)) {
    if (allConstant(VL))
      return 0;
    if (isa<InsertElementInst, InsertValueInst>(VL[0]))
      return InstructionCost::getInvalid();
    return SpillsReloads +
           processBuildVector<ShuffleCostEstimator, InstructionCost>(
               E, ScalarTy, *TTI, VectorizedVals, *this, CheckedExtracts);
  }
  if (E->State == TreeEntry::SplitVectorize) {
    assert(E->CombinedEntriesWithIndices.size() == 2 &&
           "Expected exactly 2 combined entries.");
    assert(E->ReuseShuffleIndices.empty() && "Expected empty reuses mask.");
    InstructionCost VectorCost = 0;
    if (E->ReorderIndices.empty()) {
      VectorCost = getShuffleCost(
          *TTI, TTI::SK_InsertSubvector, cast<VectorType>(FinalVecTy), CostKind,
          {}, E->CombinedEntriesWithIndices.back().second,
          cast<VectorType>(getWidenedType(
              ScalarTy,
              VectorizableTree[E->CombinedEntriesWithIndices.back().first]
                  ->getVectorFactor())));
    } else {
      unsigned CommonVF =
          std::max(VectorizableTree[E->CombinedEntriesWithIndices.front().first]
                       ->getVectorFactor(),
                   VectorizableTree[E->CombinedEntriesWithIndices.back().first]
                       ->getVectorFactor());
      VectorCost =
          getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc,
                         cast<VectorType>(getWidenedType(ScalarTy, CommonVF)),
                         CostKind, E->getSplitMask());
    }
    VectorCost += SpillsReloads;
    LLVM_DEBUG(dumpTreeCosts(E, 0, VectorCost, 0, "Calculated costs for Tree"));
    return VectorCost;
  }
  InstructionCost CommonCost = 0;
  SmallVector<int> Mask;
  if (!E->ReorderIndices.empty() && E->State != TreeEntry::CompressVectorize &&
      (E->State != TreeEntry::StridedVectorize ||
       !isReverseOrder(E->ReorderIndices))) {
    SmallVector<int> NewMask;
    if (E->getOpcode() == Instruction::Store) {
      // For stores the order is actually a mask.
      NewMask.resize(E->ReorderIndices.size());
      copy(E->ReorderIndices, NewMask.begin());
    } else {
      inversePermutation(E->ReorderIndices, NewMask);
    }
    addMask(Mask, NewMask);
  }
  if (!E->ReuseShuffleIndices.empty())
    addMask(Mask, E->ReuseShuffleIndices);
  if (!Mask.empty() && !ShuffleVectorInst::isIdentityMask(Mask, Mask.size())) {
    assert(!isa<StructType>(FinalVecTy) &&
           "Expected non-struct vector type for shuffle cost calculation.");
    CommonCost = getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc,
                                cast<VectorType>(FinalVecTy), CostKind, Mask,
                                /*Index=*/0, cast<VectorType>(VecTy));
  }
  assert((E->State == TreeEntry::Vectorize ||
          E->State == TreeEntry::ScatterVectorize ||
          E->State == TreeEntry::StridedVectorize ||
          E->State == TreeEntry::ExpandVectorize ||
          E->State == TreeEntry::CompressVectorize ||
          E->State == TreeEntry::BlendedLoadVectorize) &&
         "Unhandled state");
  assert(E->getOpcode() &&
         ((allSameType(VL) && allSameBlock(VL)) ||
          (E->getOpcode() == Instruction::GetElementPtr &&
           E->getMainOp()->getType()->isPointerTy()) ||
          E->hasCopyableElements()) &&
         "Invalid VL");
  Instruction *VL0 = E->getMainOp();
  unsigned ShuffleOrOp =
      E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode();
  if (E->CombinedOp != TreeEntry::NotCombinedOp)
    ShuffleOrOp = E->CombinedOp;
  SmallSetVector<Value *, 16> UniqueValues;
  SmallVector<unsigned, 16> UniqueIndexes;
  for (auto [Idx, V] : enumerate(VL))
    if (UniqueValues.insert(V))
      UniqueIndexes.push_back(Idx);
  const unsigned Sz = UniqueValues.size();
  SmallBitVector UsedScalars(Sz, false);
  for (unsigned I = 0; I < Sz; ++I) {
    if (isa<Instruction>(UniqueValues[I]) &&
        !E->isCopyableElement(UniqueValues[I]) &&
        getTreeEntries(UniqueValues[I]).front() == E)
      continue;
    UsedScalars.set(I);
  }
  auto GetCastContextHint = [&](Value *V) {
    if (ArrayRef<TreeEntry *> OpTEs = getTreeEntries(V); OpTEs.size() == 1)
      return getCastContextHint(*OpTEs.front());
    InstructionsState SrcState = getSameOpcode(E->getOperand(0), *TLI);
    if (SrcState && SrcState.getOpcode() == Instruction::Load &&
        !SrcState.isAltShuffle())
      return TTI::CastContextHint::GatherScatter;
    return TTI::CastContextHint::None;
  };
  auto GetCostDiff =
      [=](function_ref<InstructionCost(unsigned)> ScalarEltCost,
          function_ref<InstructionCost(InstructionCost)> VectorCost) {
        // Calculate the cost of this instruction.
        InstructionCost ScalarCost = 0;
        if (isa<CastInst, CallInst>(VL0)) {
          // For some of the instructions no need to calculate cost for each
          // particular instruction, we can use the cost of the single
          // instruction x total number of scalar instructions.
          ScalarCost = (Sz - UsedScalars.count()) * ScalarEltCost(0);
        } else {
          for (unsigned I = 0; I < Sz; ++I) {
            if (UsedScalars.test(I))
              continue;
            ScalarCost += ScalarEltCost(I);
          }
        }
 
        InstructionCost VecCost = VectorCost(CommonCost);
        // Check if the current node must be resized, if the parent node is not
        // resized.
        if (It != MinBWs.end() && !UnaryInstruction::isCast(E->getOpcode()) &&
            E->Idx != 0 &&
            (E->getOpcode() != Instruction::Load || E->UserTreeIndex)) {
          const EdgeInfo &EI = E->UserTreeIndex;
          if (!EI.UserTE->hasState() ||
              EI.UserTE->getOpcode() != Instruction::Select ||
              EI.EdgeIdx != 0) {
            auto UserBWIt = MinBWs.find(EI.UserTE);
            Type *UserScalarTy =
                (EI.UserTE->isGather() ||
                 EI.UserTE->State == TreeEntry::SplitVectorize)
                    ? EI.UserTE->Scalars.front()->getType()
                    : EI.UserTE->getOperand(EI.EdgeIdx).front()->getType();
            if (UserBWIt != MinBWs.end())
              UserScalarTy = IntegerType::get(ScalarTy->getContext(),
                                              UserBWIt->second.first);
            if (ScalarTy != UserScalarTy) {
              unsigned BWSz = DL->getTypeSizeInBits(ScalarTy);
              unsigned SrcBWSz = DL->getTypeSizeInBits(UserScalarTy);
              unsigned VecOpcode;
              auto *UserVecTy = getWidenedType(UserScalarTy, E->Scalars.size());
              if (BWSz > SrcBWSz)
                VecOpcode = Instruction::Trunc;
              else
                VecOpcode =
                    It->second.second ? Instruction::SExt : Instruction::ZExt;
              TTI::CastContextHint CCH = GetCastContextHint(VL0);
              VecCost += TTI->getCastInstrCost(VecOpcode, UserVecTy, VecTy, CCH,
                                               CostKind);
            }
          }
        }
        VecCost += SpillsReloads;
        LLVM_DEBUG(dumpTreeCosts(E, CommonCost, VecCost - CommonCost,
                                 ScalarCost, "Calculated costs for Tree"));
        return VecCost - ScalarCost;
      };
  // Price peeled intermediate instructions on the scalar side: they are
  // erased when the node vectorizes. The peeled cost is folded into the
  // first scalar-cost query so the cost dump reports the full scalar cost.
  // Peeled chain links are always 2-operand associative binops, priced per
  // instruction so the operand properties (constants, uniformity) apply.
  auto GetCostDiffWithPeeled =
      [&](function_ref<InstructionCost(unsigned)> ScalarEltCost,
          function_ref<InstructionCost(InstructionCost)> VectorCost) {
        InstructionCost PeeledScalarCost = 0;
        for (Value *V : E->getReassocScalars()) {
          auto *I = cast<Instruction>(V);
          TTI::OperandValueInfo Op1Info = TTI::getOperandInfo(I->getOperand(0));
          TTI::OperandValueInfo Op2Info = TTI::getOperandInfo(I->getOperand(1));
          PeeledScalarCost += TTI->getArithmeticInstrCost(
              I->getOpcode(), OrigScalarTy, CostKind, Op1Info, Op2Info);
        }
        bool PeeledCostAdded = false;
        InstructionCost CostDiff = GetCostDiff(
            [&](unsigned Idx) {
              InstructionCost Cost = ScalarEltCost(Idx);
              if (!PeeledCostAdded) {
                PeeledCostAdded = true;
                Cost += PeeledScalarCost;
              }
              return Cost;
            },
            VectorCost);
        // Every scalar may be marked as used elsewhere, leaving the
        // scalar-cost query uncalled and the peeled cost unapplied.
        if (!PeeledCostAdded)
          CostDiff -= PeeledScalarCost;
        return CostDiff;
      };
  // Calculate cost difference from vectorizing set of GEPs.
  // Negative value means vectorizing is profitable.
  auto GetGEPCostDiff = [=](ArrayRef<Value *> Ptrs, Value *BasePtr) {
    assert((E->State == TreeEntry::Vectorize ||
            E->State == TreeEntry::StridedVectorize ||
            E->State == TreeEntry::ExpandVectorize ||
            E->State == TreeEntry::CompressVectorize) &&
           "Entry state expected to be Vectorize, StridedVectorize, "
           "ExpandVectorize or CompressVectorize here.");
    InstructionCost ScalarCost = 0;
    InstructionCost VecCost = 0;
    std::tie(ScalarCost, VecCost) =
        getGEPCosts(*TTI, Ptrs, BasePtr, E->getOpcode(), CostKind, OrigScalarTy,
                    cast<VectorType>(VecTy));
    LLVM_DEBUG(dumpTreeCosts(E, 0, VecCost, ScalarCost,
                             "Calculated GEPs cost for Tree"));
 
    return VecCost - ScalarCost + SpillsReloads;
  };
 
  auto GetMinMaxCost = [&](Type *Ty, Instruction *VI = nullptr) {
    auto [MinMaxID, SelectOnly] = canConvertToMinOrMaxIntrinsic(VI ? VI : VL);
    if (MinMaxID == Intrinsic::not_intrinsic)
      return InstructionCost::getInvalid();
    Type *CanonicalType = Ty;
    if (CanonicalType->isPtrOrPtrVectorTy())
      CanonicalType = CanonicalType->getWithNewType(IntegerType::get(
          CanonicalType->getContext(),
          DL->getTypeSizeInBits(CanonicalType->getScalarType())));
 
    IntrinsicCostAttributes CostAttrs(MinMaxID, CanonicalType,
                                      {CanonicalType, CanonicalType});
    InstructionCost IntrinsicCost =
        TTI->getIntrinsicInstrCost(CostAttrs, CostKind);
    // If the selects are the only uses of the compares, they will be
    // dead and we can adjust the cost by removing their cost.
    if (VI && SelectOnly) {
      assert((!Ty->isVectorTy() || SLPReVec) &&
             "Expected only for scalar type.");
      auto *CI = cast<CmpInst>(VI->getOperand(0));
      IntrinsicCost -= TTI->getCmpSelInstrCost(
          CI->getOpcode(), Ty, Builder.getInt1Ty(), CI->getPredicate(),
          CostKind, {TTI::OK_AnyValue, TTI::OP_None},
          {TTI::OK_AnyValue, TTI::OP_None}, CI);
    }
    return IntrinsicCost;
  };
  auto GetFMulAddCost = [&, &TTI = *TTI](const InstructionsState &S,
                                         Instruction *VI) {
    InstructionCost Cost =
        canConvertToFMA(VI, S, *DT, *DL, TTI, *TLI, CostKind);
    return Cost;
  };
  switch (ShuffleOrOp) {
  case Instruction::PHI: {
    // Count reused scalars.
    InstructionCost ScalarCost = 0;
    SmallPtrSet<const TreeEntry *, 4> CountedOps;
    for (Value *V : UniqueValues) {
      auto *PHI = dyn_cast<PHINode>(V);
      if (!PHI)
        continue;
 
      ValueList Operands(PHI->getNumIncomingValues(), nullptr);
      for (unsigned I = 0, N = PHI->getNumIncomingValues(); I < N; ++I) {
        Value *Op = PHI->getIncomingValue(I);
        Operands[I] = Op;
      }
      if (const TreeEntry *OpTE =
              getSameValuesTreeEntry(Operands.front(), Operands))
        if (CountedOps.insert(OpTE).second &&
            !OpTE->ReuseShuffleIndices.empty())
          ScalarCost += TTI::TCC_Basic * (OpTE->ReuseShuffleIndices.size() -
                                          OpTE->Scalars.size());
    }
 
    return CommonCost - ScalarCost + SpillsReloads;
  }
  case Instruction::ExtractValue:
  case Instruction::ExtractElement: {
    // For ExtractValue entries vectorized via the struct-call path the scalar
    // extractvalue instructions are free (they become extractvalue from a
    // struct-of-vectors, which is also free). The cost is fully captured by
    // the underlying Call entry and the external-use extraction costs.
    if (ShuffleOrOp == Instruction::ExtractValue && !E->StructEVIndices.empty())
      return CommonCost;
    APInt DemandedElts;
    APInt CopyableInsertElts;
    VectorType *SrcVecTy = nullptr;
    auto GetScalarCost = [&](unsigned Idx) {
      if (isa<PoisonValue>(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
 
      // Copyable lanes are not extracts; they are inserted into the reused
      // source vector, so charge the insert to the vector side only.
      if (E->isCopyableElement(UniqueValues[Idx])) {
        if (CopyableInsertElts.isZero())
          CopyableInsertElts = APInt::getZero(E->getVectorFactor());
        CopyableInsertElts.setBit(Idx);
        return InstructionCost(TTI::TCC_Free);
      }
 
      auto *I = cast<Instruction>(UniqueValues[Idx]);
      if (!SrcVecTy) {
        if (ShuffleOrOp == Instruction::ExtractElement) {
          auto *EE = cast<ExtractElementInst>(I);
          SrcVecTy = EE->getVectorOperandType();
        } else {
          auto *EV = cast<ExtractValueInst>(I);
          Type *AggregateTy = EV->getAggregateOperand()->getType();
          unsigned NumElts;
          if (auto *ATy = dyn_cast<ArrayType>(AggregateTy))
            NumElts = ATy->getNumElements();
          else
            NumElts = AggregateTy->getStructNumElements();
          SrcVecTy = cast<VectorType>(getWidenedType(OrigScalarTy, NumElts));
        }
      }
      if (I->hasOneUse()) {
        Instruction *Ext = I->user_back();
        if ((isa<SExtInst>(Ext) || isa<ZExtInst>(Ext)) &&
            all_of(Ext->users(), IsaPred<GetElementPtrInst>)) {
          // Use getExtractWithExtendCost() to calculate the cost of
          // extractelement/ext pair.
          InstructionCost Cost = TTI->getExtractWithExtendCost(
              Ext->getOpcode(), Ext->getType(), SrcVecTy, *getExtractIndex(I),
              CostKind);
          // Subtract the cost of s|zext which is subtracted separately.
          Cost -= TTI->getCastInstrCost(
              Ext->getOpcode(), Ext->getType(), I->getType(),
              TTI::getCastContextHint(Ext), CostKind, Ext);
          return Cost;
        }
      }
      if (DemandedElts.isZero())
        DemandedElts = APInt::getZero(getNumElements(SrcVecTy));
      DemandedElts.setBit(*getExtractIndex(I));
      return InstructionCost(TTI::TCC_Free);
    };
    auto GetVectorCost = [&, &TTI = *TTI](InstructionCost CommonCost) {
      return CommonCost +
             (CopyableInsertElts.isZero()
                  ? TTI::TCC_Free
                  : TTI.getScalarizationOverhead(
                        cast<VectorType>(
                            getWidenedType(OrigScalarTy, E->getVectorFactor())),
                        CopyableInsertElts, /*Insert=*/true,
                        /*Extract=*/false, CostKind)) -
             (DemandedElts.isZero()
                  ? TTI::TCC_Free
                  : TTI.getScalarizationOverhead(SrcVecTy, DemandedElts,
                                                 /*Insert=*/false,
                                                 /*Extract=*/true, CostKind));
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case Instruction::InsertValue:
  case Instruction::InsertElement: {
    assert(E->ReuseShuffleIndices.empty() &&
           "Unique insertelements only are expected.");
    FixedVectorType *SrcVecTy = getInsertBuildVectorSrcTy(E);
    unsigned const NumElts = getNumElements(SrcVecTy);
    unsigned const NumScalars = VL.size();
 
    unsigned NumOfParts =
        getNumberOfParts(SrcVecTy, VL0->getOperand(1)->getType());
 
    SmallVector<int> InsertMask(NumElts, PoisonMaskElem);
    unsigned OffsetBeg = *getElementIndex(VL.front());
    unsigned OffsetEnd = OffsetBeg;
    InsertMask[OffsetBeg] = 0;
    for (auto [I, V] : enumerate(VL.drop_front())) {
      unsigned Idx = *getElementIndex(V);
      if (OffsetBeg > Idx)
        OffsetBeg = Idx;
      else if (OffsetEnd < Idx)
        OffsetEnd = Idx;
      InsertMask[Idx] = I + 1;
    }
    unsigned VecScalarsSz = PowerOf2Ceil(NumElts);
    if (NumOfParts > 0 && NumOfParts < NumElts)
      VecScalarsSz = PowerOf2Ceil((NumElts + NumOfParts - 1) / NumOfParts);
    unsigned VecSz = (1 + OffsetEnd / VecScalarsSz - OffsetBeg / VecScalarsSz) *
                     VecScalarsSz;
    unsigned Offset = VecScalarsSz * (OffsetBeg / VecScalarsSz);
    unsigned InsertVecSz = std::min<unsigned>(
        PowerOf2Ceil(OffsetEnd - OffsetBeg + 1),
        ((OffsetEnd - OffsetBeg + VecScalarsSz) / VecScalarsSz) * VecScalarsSz);
    bool IsWholeSubvector =
        OffsetBeg == Offset && ((OffsetEnd + 1) % VecScalarsSz == 0);
    // Check if we can safely insert a subvector. If it is not possible, just
    // generate a whole-sized vector and shuffle the source vector and the new
    // subvector.
    if (OffsetBeg + InsertVecSz > VecSz) {
      // Align OffsetBeg to generate correct mask.
      OffsetBeg = alignDown(OffsetBeg, VecSz, Offset);
      InsertVecSz = VecSz;
    }
 
    APInt DemandedElts = APInt::getZero(NumElts);
    // TODO: Add support for Instruction::InsertValue.
    SmallVector<int> Mask;
    if (!E->ReorderIndices.empty()) {
      inversePermutation(E->ReorderIndices, Mask);
      Mask.append(InsertVecSz - Mask.size(), PoisonMaskElem);
    } else {
      Mask.assign(VecSz, PoisonMaskElem);
      std::iota(Mask.begin(), std::next(Mask.begin(), InsertVecSz), 0);
    }
    bool IsIdentity = true;
    SmallVector<int> PrevMask(InsertVecSz, PoisonMaskElem);
    SmallVector<Value *> AdjustedVL(NumElts, PoisonValue::get(ScalarTy));
    Mask.swap(PrevMask);
    for (unsigned I = 0; I < NumScalars; ++I) {
      unsigned InsertIdx = *getElementIndex(VL[PrevMask[I]]);
      DemandedElts.setBit(InsertIdx);
      AdjustedVL[InsertIdx] = VL[PrevMask[I]];
      IsIdentity &= InsertIdx - OffsetBeg == I;
      Mask[InsertIdx - OffsetBeg] = I;
    }
    assert(Offset < NumElts && "Failed to find vector index offset");
 
    InstructionCost Cost = 0;
    // First cost - resize to actual vector size if not identity shuffle or
    // need to shift the vector.
    // Do not calculate the cost if the actual size is the register size and
    // we can merge this shuffle with the following SK_Select.
    auto *InsertVecTy = cast<VectorType>(getWidenedType(ScalarTy, InsertVecSz));
    if (!IsIdentity)
      Cost += getShuffleCost(*TTI, TargetTransformInfo::SK_PermuteSingleSrc,
                             InsertVecTy, CostKind, Mask);
    auto *FirstInsert = cast<Instruction>(*find_if(E->Scalars, [E](Value *V) {
      return !is_contained(E->Scalars, cast<Instruction>(V)->getOperand(0));
    }));
    // Second cost - permutation with subvector, if some elements are from the
    // initial vector or inserting a subvector.
    // TODO: Implement the analysis of the FirstInsert->getOperand(0)
    // subvector of ActualVecTy.
    SmallBitVector InMask =
        isUndefVector(FirstInsert->getOperand(0),
                      buildUseMask(NumElts, InsertMask, UseMask::UndefsAsMask));
    Cost -= getScalarizationOverhead(
        *TTI, SLPReVec, ScalarTy, SrcVecTy, DemandedElts,
        /*Insert=*/true, /*Extract=*/false, CostKind, InMask.all(), AdjustedVL,
        getVectorInstrContextHint(AdjustedVL, DemandedElts));
    if (!InMask.all() && NumScalars != NumElts && !IsWholeSubvector) {
      if (InsertVecSz != VecSz) {
        auto *ActualVecTy = cast<VectorType>(getWidenedType(ScalarTy, VecSz));
        Cost += getShuffleCost(*TTI, TTI::SK_InsertSubvector, ActualVecTy,
                               CostKind, {}, OffsetBeg - Offset, InsertVecTy);
      } else {
        for (unsigned I = 0, End = OffsetBeg - Offset; I < End; ++I)
          Mask[I] = InMask.test(I) ? PoisonMaskElem : I;
        for (unsigned I = OffsetBeg - Offset, End = OffsetEnd - Offset;
             I <= End; ++I)
          if (Mask[I] != PoisonMaskElem)
            Mask[I] = I + VecSz;
        for (unsigned I = OffsetEnd + 1 - Offset; I < VecSz; ++I)
          Mask[I] =
              ((I >= InMask.size()) || InMask.test(I)) ? PoisonMaskElem : I;
        Cost += getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc, InsertVecTy,
                               CostKind, Mask);
      }
    }
    if (ShuffleOrOp == Instruction::InsertValue &&
        !canVectorStoreInsertValue(E)) {
      // Match the stack slot alignment used during codegen (see the store +
      // load roundtrip in vectorizeTree()): the slot holds both the source
      // vector and the original aggregate.
      Align VecAlign = std::max(DL->getPrefTypeAlign(SrcVecTy),
                                DL->getPrefTypeAlign(VL0->getType()));
      Cost += TTI->getMemoryOpCost(Instruction::Store, SrcVecTy, VecAlign,
                                   /*AddressSpace=*/0, CostKind) +
              TTI->getMemoryOpCost(Instruction::Load, VL0->getType(), VecAlign,
                                   /*AddressSpace=*/0, CostKind);
    }
    return Cost + SpillsReloads;
  }
  case Instruction::ZExt:
  case Instruction::SExt:
  case Instruction::FPToUI:
  case Instruction::FPToSI:
  case Instruction::FPExt:
  case Instruction::PtrToInt:
  case Instruction::PtrToAddr:
  case Instruction::IntToPtr:
  case Instruction::SIToFP:
  case Instruction::UIToFP:
  case Instruction::Trunc:
  case Instruction::FPTrunc:
  case Instruction::BitCast: {
    auto SrcIt = MinBWs.find(getOperandEntry(E, 0));
    Type *SrcScalarTy = VL0->getOperand(0)->getType();
    auto *SrcVecTy = getWidenedType(SrcScalarTy, VL.size());
    unsigned Opcode = ShuffleOrOp;
    unsigned VecOpcode = Opcode;
    if (!ScalarTy->isFPOrFPVectorTy() && !SrcScalarTy->isFPOrFPVectorTy() &&
        (SrcIt != MinBWs.end() || It != MinBWs.end())) {
      // Check if the values are candidates to demote.
      unsigned SrcBWSz = DL->getTypeSizeInBits(SrcScalarTy->getScalarType());
      if (SrcIt != MinBWs.end()) {
        SrcBWSz = SrcIt->second.first;
        unsigned SrcScalarTyNumElements = getNumElements(SrcScalarTy);
        SrcScalarTy = IntegerType::get(F->getContext(), SrcBWSz);
        SrcVecTy =
            getWidenedType(SrcScalarTy, VL.size() * SrcScalarTyNumElements);
      }
      unsigned BWSz = DL->getTypeSizeInBits(ScalarTy->getScalarType());
      if (BWSz == SrcBWSz) {
        VecOpcode = Instruction::BitCast;
      } else if (BWSz < SrcBWSz) {
        VecOpcode = Instruction::Trunc;
      } else if (It != MinBWs.end()) {
        assert(BWSz > SrcBWSz && "Invalid cast!");
        VecOpcode = It->second.second ? Instruction::SExt : Instruction::ZExt;
      } else if (SrcIt != MinBWs.end()) {
        assert(BWSz > SrcBWSz && "Invalid cast!");
        VecOpcode =
            SrcIt->second.second ? Instruction::SExt : Instruction::ZExt;
      }
    } else if (VecOpcode == Instruction::SIToFP && SrcIt != MinBWs.end() &&
               !SrcIt->second.second) {
      VecOpcode = Instruction::UIToFP;
    }
    auto GetScalarCost = [&](unsigned Idx) -> InstructionCost {
      assert(Idx == 0 && "Expected 0 index only");
      return TTI->getCastInstrCost(Opcode, VL0->getType(),
                                   VL0->getOperand(0)->getType(),
                                   TTI::getCastContextHint(VL0), CostKind, VL0);
    };
    auto GetVectorCost = [=](InstructionCost CommonCost) {
      // Do not count cost here if minimum bitwidth is in effect and it is just
      // a bitcast (here it is just a noop).
      if (VecOpcode != Opcode && VecOpcode == Instruction::BitCast)
        return CommonCost;
      auto *VI = VL0->getOpcode() == Opcode ? VL0 : nullptr;
      TTI::CastContextHint CCH = GetCastContextHint(VL0->getOperand(0));
 
      bool IsArithmeticExtendedReduction =
          E->Idx == 0 && UserIgnoreList &&
          all_of(*UserIgnoreList, [](Value *V) {
            auto *I = cast<Instruction>(V);
            return is_contained({Instruction::Add, Instruction::FAdd,
                                 Instruction::Mul, Instruction::FMul,
                                 Instruction::And, Instruction::Or,
                                 Instruction::Xor},
                                I->getOpcode());
          });
      if (IsArithmeticExtendedReduction &&
          (VecOpcode == Instruction::ZExt || VecOpcode == Instruction::SExt))
        return CommonCost;
      return CommonCost +
             TTI->getCastInstrCost(VecOpcode, VecTy, SrcVecTy, CCH, CostKind,
                                   VecOpcode == Opcode ? VI : nullptr);
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case Instruction::FCmp:
  case Instruction::ICmp:
    // Override ScalarTy/VecTy with the compared operand type (not i1). The
    // cost of a compare instruction is determined by the operand width, and
    // getCmpSelInstrCost expects the compared type as its first type arg.
    OrigScalarTy = ScalarTy =
        getValueType(VL0, SLPReVec, /*LookThroughCmp=*/true);
    VecTy = getWidenedType(ScalarTy, VL.size());
    [[fallthrough]];
  case Instruction::Select: {
    CmpPredicate VecPred, SwappedVecPred;
    auto MatchCmp = m_Cmp(VecPred, m_Value(), m_Value());
    if (match(VL0, m_Select(MatchCmp, m_Value(), m_Value())) ||
        match(VL0, MatchCmp))
      SwappedVecPred = CmpInst::getSwappedPredicate(VecPred);
    else
      SwappedVecPred = VecPred = ScalarTy->isFloatingPointTy()
                                     ? CmpInst::BAD_FCMP_PREDICATE
                                     : CmpInst::BAD_ICMP_PREDICATE;
    auto GetScalarCost = [&](unsigned Idx) {
      if (isa<PoisonValue>(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
 
      if (!isa<SelectInst>(UniqueValues[Idx]))
        return TTI->getInstructionCost(cast<Instruction>(UniqueValues[Idx]),
                                       CostKind);
 
      auto *VI = cast<Instruction>(UniqueValues[Idx]);
      CmpPredicate CurrentPred = ScalarTy->isFloatingPointTy()
                                     ? CmpInst::BAD_FCMP_PREDICATE
                                     : CmpInst::BAD_ICMP_PREDICATE;
      Value *LHS = nullptr, *RHS = nullptr;
      auto MatchCmp = m_Cmp(CurrentPred, m_Value(), m_Value());
      bool IsSelect =
          ShuffleOrOp == Instruction::Select &&
          (match(VI, m_Select(MatchCmp, m_Value(LHS), m_Value(RHS))) ||
           match(VI, m_Select(m_Value(), m_Value(LHS), m_Value(RHS))));
      if ((!IsSelect && !match(VI, MatchCmp)) ||
          (CurrentPred != static_cast<CmpInst::Predicate>(VecPred) &&
           CurrentPred != static_cast<CmpInst::Predicate>(SwappedVecPred)))
        VecPred = SwappedVecPred = ScalarTy->isFloatingPointTy()
                                       ? CmpInst::BAD_FCMP_PREDICATE
                                       : CmpInst::BAD_ICMP_PREDICATE;
 
      // Check if operands are of i1 types, like a condition expression.
      // TODO: consider implementing this in TTI.
      InstructionCost ScalarCost = InstructionCost::getInvalid();
      if (IsSelect && LHS->getType() == VI->getOperand(0)->getType()) {
        assert(LHS->getType() == RHS->getType() &&
               "Expected same type for LHS/RHS");
        // select i1 v, i1 true, i1 b -> or i1 v, i1 b
        if (match(LHS, m_AllOnes())) {
          ScalarCost = TTI->getArithmeticInstrCost(
              Instruction::Or, LHS->getType(), CostKind,
              getOperandInfo(VI->getOperand(0)), getOperandInfo(RHS));
        } else if (match(RHS, m_Zero())) {
          // select i1 v, i1 b, i1 false -> and i1 v, i1 b
          ScalarCost = TTI->getArithmeticInstrCost(
              Instruction::And, LHS->getType(), CostKind,
              getOperandInfo(VI->getOperand(0)), getOperandInfo(LHS));
        }
      }
      if (!ScalarCost.isValid()) {
        // For selects, the "condition type" arg is the condition operand's
        // type; for standalone compares, it is the result type (i1).
        ScalarCost = TTI->getCmpSelInstrCost(
            E->getOpcode(), OrigScalarTy,
            ShuffleOrOp == Instruction::Select ? VL0->getOperand(0)->getType()
                                               : VL0->getType(),
            CurrentPred, CostKind,
            getOperandInfo(
                VI->getOperand(ShuffleOrOp == Instruction::Select ? 1 : 0)),
            getOperandInfo(
                VI->getOperand(ShuffleOrOp == Instruction::Select ? 2 : 1)),
            VI);
      }
      InstructionCost IntrinsicCost = GetMinMaxCost(OrigScalarTy, VI);
      if (IntrinsicCost.isValid())
        ScalarCost = IntrinsicCost;
 
      return ScalarCost;
    };
    auto GetVectorCost = [&](InstructionCost CommonCost) {
      // For selects, the condition type may differ from the result type
      // (e.g. condition is <N x i1> while result is <N x i32>). For
      // compares, the result type IS the mask (i1/vNi1). Construct the
      // right type so getCmpSelInstrCost sees the actual mask/result width.
      auto *MaskTy = cast<VectorType>(getWidenedType(
          ShuffleOrOp == Instruction::Select ? VL0->getOperand(0)->getType()
                                             : VL0->getType(),
          VL.size()));
 
      InstructionCost VecCost = InstructionCost::getInvalid();
      if (ShuffleOrOp == Instruction::Select) {
        ArrayRef<Value *> Cond = E->getOperand(0);
        ArrayRef<Value *> LHS = E->getOperand(1);
        ArrayRef<Value *> RHS = E->getOperand(2);
        // select <VF x i1>, <VF x i1>, <VF x i1>?
        // TODO: consider implementing this in TTI.
        if (Cond.front()->getType() == LHS.front()->getType()) {
          // select <VF x i1> v, <VF x i1> true, <VF x i1> b -> or <VF x i1> v,
          // <VF x i1> b
          if (all_of(LHS, [&](Value *V) { return match(V, m_AllOnes()); })) {
            VecCost = TTI->getArithmeticInstrCost(
                Instruction::Or, VecTy, CostKind, getOperandInfo(Cond),
                getOperandInfo(RHS));
          } else if (all_of(RHS,
                            [&](Value *V) { return match(V, m_Zero()); })) {
            // select <VF x i1> v, <VF x i1> b, <VF x i1> false -> and <VF x i1>
            // v, <VF x i1> b
            VecCost = TTI->getArithmeticInstrCost(
                Instruction::And, VecTy, CostKind, getOperandInfo(Cond),
                getOperandInfo(LHS));
          }
        }
      }
      if (!VecCost.isValid()) {
        VecCost = TTI->getCmpSelInstrCost(
            E->getOpcode(), VecTy, MaskTy, VecPred, CostKind,
            getOperandInfo(
                E->getOperand(ShuffleOrOp == Instruction::Select ? 1 : 0)),
            getOperandInfo(
                E->getOperand(ShuffleOrOp == Instruction::Select ? 2 : 1)),
            VL0);
        if (isa<SelectInst>(VL0)) {
          unsigned CondNumElements = getNumElements(MaskTy);
          unsigned VecTyNumElements = getNumElements(VecTy);
          assert(VecTyNumElements >= CondNumElements &&
                 VecTyNumElements % CondNumElements == 0 &&
                 "Cannot vectorize Instruction::Select");
          if (CondNumElements != VecTyNumElements) {
            // When the return type is i1 but the source is fixed vector type,
            // we need to duplicate the condition value.
            VecCost += getShuffleCost(
                *TTI, TTI::SK_PermuteSingleSrc, MaskTy, CostKind,
                createReplicatedMask(VecTyNumElements / CondNumElements,
                                     CondNumElements));
          }
        }
      }
      return VecCost + CommonCost;
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case TreeEntry::MinMax: {
    auto GetScalarCost = [&](unsigned Idx) {
      return GetMinMaxCost(OrigScalarTy);
    };
    auto GetVectorCost = [&](InstructionCost CommonCost) {
      InstructionCost VecCost = GetMinMaxCost(VecTy);
      return VecCost + CommonCost;
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case TreeEntry::FMulAdd: {
    auto GetScalarCost = [&](unsigned Idx) {
      if (isa<PoisonValue>(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
      return GetFMulAddCost(E->getOperations(),
                            cast<Instruction>(UniqueValues[Idx]));
    };
    auto GetVectorCost = [&, &TTI = *TTI](InstructionCost CommonCost) {
      FastMathFlags FMF;
      FMF.set();
      for (Value *V : E->Scalars) {
        if (auto *FPCI = dyn_cast<FPMathOperator>(V)) {
          FMF &= FPCI->getFastMathFlags();
          if (auto *FPCIOp = dyn_cast<FPMathOperator>(FPCI->getOperand(0)))
            FMF &= FPCIOp->getFastMathFlags();
        }
      }
      IntrinsicCostAttributes ICA(Intrinsic::fmuladd, VecTy,
                                  {VecTy, VecTy, VecTy}, FMF);
      InstructionCost VecCost = TTI.getIntrinsicInstrCost(ICA, CostKind);
      return VecCost + CommonCost;
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case TreeEntry::ReducedBitcast:
  case TreeEntry::ReducedBitcastBSwap: {
    auto GetScalarCost = [&, &TTI = *TTI](unsigned Idx) {
      if (isa<PoisonValue>(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
      auto *Shl = dyn_cast<Instruction>(UniqueValues[Idx]);
      if (!Shl)
        return InstructionCost(TTI::TCC_Free);
      InstructionCost ScalarCost = TTI.getInstructionCost(Shl, CostKind);
      auto *ZExt = dyn_cast<Instruction>(Shl->getOperand(0));
      if (!ZExt)
        return ScalarCost;
      ScalarCost += TTI.getInstructionCost(ZExt, CostKind);
      return ScalarCost;
    };
    auto GetVectorCost = [&, &TTI = *TTI](InstructionCost CommonCost) {
      const TreeEntry *LhsTE = getOperandEntry(E, /*Idx=*/0);
      TTI::CastContextHint CastCtx =
          getCastContextHint(*getOperandEntry(LhsTE, /*Idx=*/0));
      Type *SrcScalarTy = cast<ZExtInst>(LhsTE->getMainOp())->getSrcTy();
      auto *SrcVecTy = getWidenedType(SrcScalarTy, LhsTE->getVectorFactor());
      InstructionCost BitcastCost = TTI.getCastInstrCost(
          Instruction::BitCast, ScalarTy, SrcVecTy, CastCtx, CostKind);
      if (ShuffleOrOp == TreeEntry::ReducedBitcastBSwap) {
        auto *SrcType = IntegerType::getIntNTy(
            ScalarTy->getContext(),
            DL->getTypeSizeInBits(SrcScalarTy) * EntryVF);
        IntrinsicCostAttributes CostAttrs(Intrinsic::bswap, SrcType, {SrcType});
        InstructionCost IntrinsicCost =
            TTI.getIntrinsicInstrCost(CostAttrs, CostKind);
        BitcastCost += IntrinsicCost;
        if (SrcType != ScalarTy) {
          BitcastCost +=
              TTI.getCastInstrCost(Instruction::ZExt, ScalarTy, SrcType,
                                   TTI::CastContextHint::None, CostKind);
        }
      }
      return BitcastCost + CommonCost;
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case TreeEntry::ReducedBitcastLoads:
  case TreeEntry::ReducedBitcastBSwapLoads: {
    auto GetScalarCost = [&, &TTI = *TTI](unsigned Idx) {
      if (isa<PoisonValue>(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
      auto *Shl = dyn_cast<Instruction>(UniqueValues[Idx]);
      if (!Shl)
        return InstructionCost(TTI::TCC_Free);
      InstructionCost ScalarCost = TTI.getInstructionCost(Shl, CostKind);
      auto *ZExt = dyn_cast<Instruction>(Shl->getOperand(0));
      if (!ZExt)
        return ScalarCost;
      ScalarCost += TTI.getInstructionCost(ZExt, CostKind);
      auto *Load = dyn_cast<Instruction>(ZExt->getOperand(0));
      if (!Load)
        return ScalarCost;
      ScalarCost += TTI.getInstructionCost(Load, CostKind);
      return ScalarCost;
    };
    auto GetVectorCost = [&, &TTI = *TTI](InstructionCost CommonCost) {
      const TreeEntry *LhsTE = getOperandEntry(E, /*Idx=*/0);
      const TreeEntry *LoadTE = getOperandEntry(LhsTE, /*Idx=*/0);
      auto *LI0 = cast<LoadInst>(LoadTE->getMainOp());
      auto *SrcType = IntegerType::getIntNTy(
          ScalarTy->getContext(),
          DL->getTypeSizeInBits(LI0->getType()) * EntryVF);
      InstructionCost LoadCost =
          TTI.getMemoryOpCost(Instruction::Load, SrcType, LI0->getAlign(),
                              LI0->getPointerAddressSpace(), CostKind);
      if (ShuffleOrOp == TreeEntry::ReducedBitcastBSwapLoads) {
        IntrinsicCostAttributes CostAttrs(Intrinsic::bswap, SrcType, {SrcType});
        InstructionCost IntrinsicCost =
            TTI.getIntrinsicInstrCost(CostAttrs, CostKind);
        LoadCost += IntrinsicCost;
        if (SrcType != ScalarTy) {
          LoadCost +=
              TTI.getCastInstrCost(Instruction::ZExt, ScalarTy, SrcType,
                                   TTI::CastContextHint::None, CostKind);
        }
      }
      return LoadCost + CommonCost;
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case TreeEntry::ReducedCmpBitcast: {
    auto GetScalarCost = [&, &TTI = *TTI](unsigned Idx) {
      if (isa<PoisonValue>(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
      auto *Sel = dyn_cast<Instruction>(UniqueValues[Idx]);
      if (!Sel)
        return InstructionCost(TTI::TCC_Free);
      InstructionCost ScalarCost = TTI.getInstructionCost(Sel, CostKind);
      return ScalarCost;
    };
    auto GetVectorCost = [&, &TTI = *TTI](InstructionCost CommonCost) {
      Type *CmpTy = CmpInst::makeCmpResultType(VecTy);
      auto *DstTy =
          IntegerType::getIntNTy(ScalarTy->getContext(), E->getVectorFactor());
      InstructionCost BitcastCost =
          TTI.getCastInstrCost(Instruction::BitCast, DstTy, CmpTy,
                               TTI::CastContextHint::None, CostKind);
      if (DstTy != ScalarTy) {
        BitcastCost +=
            TTI.getCastInstrCost(Instruction::ZExt, ScalarTy, DstTy,
                                 TTI::CastContextHint::None, CostKind);
      }
      return BitcastCost + CommonCost;
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case Instruction::FNeg:
  case Instruction::Add:
  case Instruction::FAdd:
  case Instruction::Sub:
  case Instruction::FSub:
  case Instruction::Mul:
  case Instruction::FMul:
  case Instruction::UDiv:
  case Instruction::SDiv:
  case Instruction::FDiv:
  case Instruction::URem:
  case Instruction::SRem:
  case Instruction::FRem:
  case Instruction::Shl:
  case Instruction::LShr:
  case Instruction::AShr:
  case Instruction::And:
  case Instruction::Or:
  case Instruction::Xor: {
    auto GetScalarCost = [&](unsigned Idx) {
      if (isa<PoisonValue>(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
 
      // We cannot retrieve the operand from UniqueValues[Idx] because an
      // interchangeable instruction may be used. The order and the actual
      // operand might differ from what is retrieved from UniqueValues[Idx].
      unsigned Lane = UniqueIndexes[Idx];
      Value *Op1 = E->getOperand(0)[Lane];
      Value *Op2;
      SmallVector<const Value *, 2> Operands(1, Op1);
      if (isa<UnaryOperator>(UniqueValues[Idx])) {
        Op2 = Op1;
      } else {
        Op2 = E->getOperand(1)[Lane];
        Operands.push_back(Op2);
      }
      TTI::OperandValueInfo Op1Info = TTI::getOperandInfo(Op1);
      TTI::OperandValueInfo Op2Info = TTI::getOperandInfo(Op2);
      InstructionCost ScalarCost = TTI->getArithmeticInstrCost(
          ShuffleOrOp, OrigScalarTy, CostKind, Op1Info, Op2Info, Operands);
      if (auto *I = dyn_cast<Instruction>(UniqueValues[Idx]);
          I && (ShuffleOrOp == Instruction::FAdd ||
                ShuffleOrOp == Instruction::FSub)) {
        InstructionCost IntrinsicCost = GetFMulAddCost(E->getOperations(), I);
        if (IntrinsicCost.isValid())
          ScalarCost = IntrinsicCost;
      }
      return ScalarCost;
    };
    auto GetVectorCost = [=](InstructionCost CommonCost) {
      // And peephole only applies to plain 2-operand nodes.
      if (ShuffleOrOp == Instruction::And && It != MinBWs.end() &&
          !E->hasReassocScalars()) {
        for (unsigned I : seq<unsigned>(0, E->getNumOperands())) {
          ArrayRef<Value *> Ops = E->getOperand(I);
          if (all_of(Ops, [&](Value *Op) {
                auto *CI = dyn_cast<ConstantInt>(Op);
                return CI && CI->getValue().countr_one() >= It->second.first;
              }))
            return CommonCost;
        }
      }
      // Masked path ignores Op1Info/Op2Info like its codegen does; keep it
      // out of the operand-aware cost comparison below.
      if (InstructionCost MaskedCost = getMaskedDivRemCost(
              *TTI, SLPReVec, ShuffleOrOp, ScalarTy, VL.size(), CostKind);
          MaskedCost.isValid())
        return MaskedCost + CommonCost;
      unsigned OpIdx = isa<UnaryOperator>(VL0) ? 0 : 1;
      TTI::OperandValueInfo Op1Info = getOperandInfo(E->getOperand(0));
      TTI::OperandValueInfo Op2Info = getOperandInfo(E->getOperand(OpIdx));
      InstructionCost Cost = TTI->getArithmeticInstrCost(
          ShuffleOrOp, VecTy, CostKind, Op1Info, Op2Info, {}, nullptr, TLI);
      // N columns need N-1 vector combines; price extra columns
      // conservatively, skipping identity-only columns (not combined by
      // codegen).
      if (E->hasReassocScalars()) {
        const unsigned CombineOpcode = getReassocCombineOpcode(E->getOpcode());
        for (unsigned Idx : seq<unsigned>(2, E->getNumOperands())) {
          if (all_of(E->getOperand(Idx), [&](Value *V) {
                return isBinOpIdentityConstant(V, CombineOpcode);
              }))
            continue;
          Cost += TTI->getArithmeticInstrCost(
              ShuffleOrOp, VecTy, CostKind, {},
              getOperandInfo(E->getOperand(Idx)), {}, nullptr, TLI);
        }
      }
      return Cost + CommonCost;
    };
    return GetCostDiffWithPeeled(GetScalarCost, GetVectorCost);
  }
  case Instruction::GetElementPtr: {
    return CommonCost + GetGEPCostDiff(VL, VL0);
  }
  case Instruction::Load: {
    auto GetScalarCost = [&](unsigned Idx) {
      auto *VI = cast<LoadInst>(UniqueValues[Idx]);
      return TTI->getMemoryOpCost(Instruction::Load, OrigScalarTy,
                                  VI->getAlign(), VI->getPointerAddressSpace(),
                                  CostKind, TTI::OperandValueInfo(), VI);
    };
    auto *LI0 = cast<LoadInst>(VL0);
    auto GetVectorCost = [&](InstructionCost CommonCost) {
      InstructionCost VecLdCost;
      switch (E->State) {
      case TreeEntry::Vectorize:
        if (unsigned Factor = E->getInterleaveFactor()) {
          VecLdCost = TTI->getInterleavedMemoryOpCost(
              Instruction::Load, VecTy, Factor, {}, LI0->getAlign(),
              LI0->getPointerAddressSpace(), CostKind);
 
        } else {
          VecLdCost = TTI->getMemoryOpCost(
              Instruction::Load, VecTy, LI0->getAlign(),
              LI0->getPointerAddressSpace(), CostKind, TTI::OperandValueInfo());
        }
        break;
      case TreeEntry::StridedVectorize: {
        const StridedPtrInfo &SPtrInfo = TreeEntryToStridedPtrInfoMap.at(E);
        FixedVectorType *StridedLoadTy = SPtrInfo.Ty;
        assert(StridedLoadTy && "Missing StridedPointerInfo for tree entry.");
        Align CommonAlignment =
            computeCommonAlignment<LoadInst>(UniqueValues.getArrayRef());
        VecLdCost = TTI->getMemIntrinsicInstrCost(
            MemIntrinsicCostAttributes(Intrinsic::experimental_vp_strided_load,
                                       StridedLoadTy, LI0->getPointerOperand(),
                                       /*VariableMask=*/false, CommonAlignment),
            CostKind);
        if (StridedLoadTy != VecTy)
          VecLdCost +=
              TTI->getCastInstrCost(Instruction::BitCast, VecTy, StridedLoadTy,
                                    getCastContextHint(*E), CostKind);
 
        break;
      }
      case TreeEntry::CompressVectorize: {
        bool IsMasked;
        unsigned InterleaveFactor;
        SmallVector<int> CompressMask;
        VectorType *LoadVecTy;
        SmallVector<Value *> Scalars(VL);
        if (!E->ReorderIndices.empty()) {
          SmallVector<int> Mask(E->ReorderIndices.begin(),
                                E->ReorderIndices.end());
          reorderScalars(Scalars, Mask);
        }
        SmallVector<Value *> PointerOps(Scalars.size());
        for (auto [I, V] : enumerate(Scalars))
          PointerOps[I] = cast<LoadInst>(V)->getPointerOperand();
        [[maybe_unused]] bool IsVectorized = isMaskedLoadCompress(
            Scalars, PointerOps, E->ReorderIndices, *TTI, *DL, *SE, *AC, *DT,
            *TLI, CostKind, [](Value *) { return true; }, IsMasked,
            InterleaveFactor, CompressMask, LoadVecTy);
        CompressEntryToData.try_emplace(E, CompressMask, LoadVecTy,
                                        InterleaveFactor, IsMasked);
        Align CommonAlignment = LI0->getAlign();
        if (InterleaveFactor) {
          VecLdCost = TTI->getInterleavedMemoryOpCost(
              Instruction::Load, LoadVecTy, InterleaveFactor, {},
              CommonAlignment, LI0->getPointerAddressSpace(), CostKind);
        } else if (IsMasked) {
          VecLdCost = TTI->getMemIntrinsicInstrCost(
              MemIntrinsicCostAttributes(Intrinsic::masked_load, LoadVecTy,
                                         CommonAlignment,
                                         LI0->getPointerAddressSpace()),
              CostKind);
          // TODO: include this cost into CommonCost.
          VecLdCost += getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc, LoadVecTy,
                                      CostKind, CompressMask);
        } else {
          VecLdCost = TTI->getMemoryOpCost(
              Instruction::Load, LoadVecTy, CommonAlignment,
              LI0->getPointerAddressSpace(), CostKind, TTI::OperandValueInfo());
          // TODO: include this cost into CommonCost.
          VecLdCost += getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc, LoadVecTy,
                                      CostKind, CompressMask);
        }
        break;
      }
      case TreeEntry::ScatterVectorize: {
        Align CommonAlignment =
            computeCommonAlignment<LoadInst>(UniqueValues.getArrayRef());
        VecLdCost = TTI->getMemIntrinsicInstrCost(
            MemIntrinsicCostAttributes(Intrinsic::masked_gather, VecTy,
                                       LI0->getPointerOperand(),
                                       /*VariableMask=*/false, CommonAlignment),
            CostKind);
        break;
      }
      case TreeEntry::BlendedLoadVectorize: {
        // Two masked loads (one per candidate base) blended by a select.
        Align CommonAlignment =
            computeCommonAlignment<LoadInst>(UniqueValues.getArrayRef());
        VecLdCost = getBlendedLoadCost(*TTI, VecTy, CommonAlignment,
                                       LI0->getPointerAddressSpace(), CostKind);
        break;
      }
      case TreeEntry::ExpandVectorize:
      case TreeEntry::CombinedVectorize:
      case TreeEntry::SplitVectorize:
      case TreeEntry::NeedToGather:
        llvm_unreachable("Unexpected vectorization state.");
      }
      return VecLdCost + CommonCost;
    };
 
    InstructionCost Cost = GetCostDiff(GetScalarCost, GetVectorCost);
    // Masked gather and blended loads are not terminal nodes: their address
    // cost is estimated separately (blended loads have no per-lane address).
    if (E->State == TreeEntry::ScatterVectorize ||
        E->State == TreeEntry::BlendedLoadVectorize)
      return Cost;
 
    // Estimate cost of GEPs since this tree node is a terminator.
    SmallVector<Value *> PointerOps(VL.size());
    for (auto [I, V] : enumerate(VL))
      PointerOps[I] = cast<LoadInst>(V)->getPointerOperand();
    return Cost + GetGEPCostDiff(PointerOps, LI0->getPointerOperand());
  }
  case Instruction::Store: {
    bool IsReorder = !E->ReorderIndices.empty();
    auto GetScalarCost = [=](unsigned Idx) {
      auto *VI = cast<StoreInst>(VL[Idx]);
      TTI::OperandValueInfo OpInfo = TTI::getOperandInfo(VI->getValueOperand());
      return TTI->getMemoryOpCost(Instruction::Store, OrigScalarTy,
                                  VI->getAlign(), VI->getPointerAddressSpace(),
                                  CostKind, OpInfo, VI);
    };
    auto *BaseSI =
        cast<StoreInst>(IsReorder ? VL[E->ReorderIndices.front()] : VL0);
    auto GetVectorCost = [=](InstructionCost CommonCost) {
      // We know that we can merge the stores. Calculate the cost.
      InstructionCost VecStCost;
      if (E->State == TreeEntry::StridedVectorize) {
        const StridedPtrInfo &SPtrInfo = TreeEntryToStridedPtrInfoMap.at(E);
        FixedVectorType *StridedStoreTy = SPtrInfo.Ty;
        assert(StridedStoreTy && "Missing StridedPointerInfo for tree entry.");
        Align CommonAlignment =
            computeCommonAlignment<StoreInst>(UniqueValues.getArrayRef());
        VecStCost = TTI->getMemIntrinsicInstrCost(
            MemIntrinsicCostAttributes(Intrinsic::experimental_vp_strided_store,
                                       StridedStoreTy,
                                       BaseSI->getPointerOperand(),
                                       /*VariableMask=*/false, CommonAlignment),
            CostKind);
        if (StridedStoreTy != VecTy)
          VecStCost +=
              TTI->getCastInstrCost(Instruction::BitCast, VecTy, StridedStoreTy,
                                    getCastContextHint(*E), CostKind);
      } else if (E->State == TreeEntry::ExpandVectorize) {
        const StridedPtrInfo &SPtrInfo = TreeEntryToStridedPtrInfoMap.at(E);
        FixedVectorType *MaskedStoreTy = SPtrInfo.Ty;
        assert(MaskedStoreTy && "Missing StridedPointerInfo for tree entry.");
        Align CommonAlignment =
            computeCommonAlignment<StoreInst>(UniqueValues.getArrayRef());
        // Masked store: the values are expanded into the widened vector and
        // stored with a constant mask.
        VecStCost = TTI->getMemIntrinsicInstrCost(
            MemIntrinsicCostAttributes(Intrinsic::masked_store, MaskedStoreTy,
                                       CommonAlignment,
                                       BaseSI->getPointerAddressSpace()),
            CostKind);
      } else {
        assert(E->State == TreeEntry::Vectorize &&
               "Expected either strided, consecutive, or expanded stores.");
        if (unsigned Factor = E->getInterleaveFactor()) {
          assert(E->ReuseShuffleIndices.empty() && !E->ReorderIndices.empty() &&
                 "No reused shuffles expected");
          CommonCost = 0;
          VecStCost = TTI->getInterleavedMemoryOpCost(
              Instruction::Store, VecTy, Factor, {}, BaseSI->getAlign(),
              BaseSI->getPointerAddressSpace(), CostKind);
        } else {
          TTI::OperandValueInfo OpInfo = getOperandInfo(E->getOperand(0));
          VecStCost = TTI->getMemoryOpCost(
              Instruction::Store, VecTy, BaseSI->getAlign(),
              BaseSI->getPointerAddressSpace(), CostKind, OpInfo);
        }
      }
      return VecStCost + CommonCost;
    };
    SmallVector<Value *> PointerOps(VL.size());
    for (auto [I, V] : enumerate(VL)) {
      unsigned Idx = IsReorder ? E->ReorderIndices[I] : I;
      PointerOps[Idx] = cast<StoreInst>(V)->getPointerOperand();
    }
 
    return GetCostDiff(GetScalarCost, GetVectorCost) +
           GetGEPCostDiff(PointerOps, BaseSI->getPointerOperand());
  }
  case Instruction::Call: {
    auto GetScalarCost = [&](unsigned Idx) {
      // A copyable lane has no scalar call of its own; its real cost is
      // accounted for wherever it is otherwise computed.
      if (isa<PoisonValue>(UniqueValues[Idx]) ||
          E->isCopyableElement(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
      auto *CI = cast<CallInst>(UniqueValues[Idx]);
      Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
      if (ID != Intrinsic::not_intrinsic) {
        IntrinsicCostAttributes CostAttrs(ID, *CI, 1);
        return TTI->getIntrinsicInstrCost(CostAttrs, CostKind);
      }
      return TTI->getCallInstrCost(CI->getCalledFunction(),
                                   CI->getFunctionType()->getReturnType(),
                                   CI->getFunctionType()->params(), CostKind);
    };
    auto GetVectorCost = [=](InstructionCost CommonCost) {
      auto *CI = cast<CallInst>(VL0);
      Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
      SmallVector<Type *> ArgTys = buildIntrinsicArgTypes(
          CI, ID, getNumElements(VecTy),
          It != MinBWs.end() ? It->second.first : 0, TTI);
      auto VecCallCosts =
          getVectorCallCosts(CI, VecTy, TTI, TLI, ArgTys, CostKind);
      return std::min(VecCallCosts.first, VecCallCosts.second) + CommonCost;
    };
    return GetCostDiff(GetScalarCost, GetVectorCost);
  }
  case Instruction::ShuffleVector: {
    if (!SLPReVec || E->isAltShuffle())
      assert(E->isAltShuffle() &&
             ((Instruction::isBinaryOp(E->getOpcode()) &&
               Instruction::isBinaryOp(E->getAltOpcode())) ||
              (Instruction::isCast(E->getOpcode()) &&
               Instruction::isCast(E->getAltOpcode())) ||
              (isa<CmpInst>(VL0) && isa<CmpInst>(E->getAltOp()))) &&
             "Invalid Shuffle Vector Operand");
    // Try to find the previous shuffle node with the same operands and same
    // main/alternate ops.
    auto TryFindNodeWithEqualOperands = [=]() {
      for (const std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
        if (TE.get() == E)
          break;
        if (TE->hasState() && TE->isAltShuffle() &&
            ((TE->getOpcode() == E->getOpcode() &&
              TE->getAltOpcode() == E->getAltOpcode()) ||
             (TE->getOpcode() == E->getAltOpcode() &&
              TE->getAltOpcode() == E->getOpcode())) &&
            TE->hasEqualOperands(*E))
          return true;
      }
      return false;
    };
    auto GetScalarCost = [&](unsigned Idx) {
      if (isa<PoisonValue>(UniqueValues[Idx]))
        return InstructionCost(TTI::TCC_Free);
 
      auto *VI = cast<Instruction>(UniqueValues[Idx]);
      assert(E->getMatchingMainOpOrAltOp(VI) &&
             "Unexpected main/alternate opcode");
      (void)E;
      return TTI->getInstructionCost(VI, CostKind);
    };
    // Need to clear CommonCost since the final shuffle cost is included into
    // vector cost.
    auto GetVectorCost = [&, &TTIRef = *TTI](InstructionCost) {
      // VecCost is equal to sum of the cost of creating 2 vectors
      // and the cost of creating shuffle.
      InstructionCost VecCost = 0;
      if (TryFindNodeWithEqualOperands()) {
        LLVM_DEBUG({
          dbgs() << "SLP: diamond match for alternate node found.\n";
          E->dump();
        });
        // No need to add new vector costs here since we're going to reuse
        // same main/alternate vector ops, just do different shuffling.
      } else if (Instruction::isBinaryOp(E->getOpcode())) {
        // Peeled alternate chains fold the operand columns into one pure
        // main-opcode chain and one pure alt-opcode chain, followed by a
        // single lane-select shuffle; a plain alternate node is a single
        // combine. Each combine is priced with the properties of the column
        // it folds in; the other operand is the running fold, which stays
        // constant while every folded column is constant (such combines
        // constant-fold away in codegen) and stays uniform while every
        // folded column is uniform.
        auto ChainCost = [&](unsigned Opcode) {
          InstructionCost Cost = 0;
          TTI::OperandValueInfo RunningInfo = getOperandInfo(E->getOperand(0));
          for (unsigned Idx : seq<unsigned>(1, E->getNumOperands())) {
            TTI::OperandValueInfo ColInfo = getOperandInfo(E->getOperand(Idx));
            if (!RunningInfo.isConstant() || !ColInfo.isConstant())
              Cost += TTIRef.getArithmeticInstrCost(Opcode, VecTy, CostKind,
                                                    RunningInfo, ColInfo, {},
                                                    nullptr, TLI);
            TTI::OperandValueKind Kind = TTI::OK_AnyValue;
            if (RunningInfo.isConstant() && ColInfo.isConstant())
              Kind = RunningInfo.Kind == TTI::OK_UniformConstantValue &&
                             ColInfo.Kind == TTI::OK_UniformConstantValue
                         ? TTI::OK_UniformConstantValue
                         : TTI::OK_NonUniformConstantValue;
            else if (RunningInfo.isUniform() && ColInfo.isUniform())
              Kind = TTI::OK_UniformValue;
            RunningInfo = {Kind, TTI::OP_None};
          }
          return Cost;
        };
        VecCost = ChainCost(E->getOpcode()) + ChainCost(E->getAltOpcode());
      } else if (auto *CI0 = dyn_cast<CmpInst>(VL0)) {
        auto *MaskTy = getWidenedType(Builder.getInt1Ty(), VL.size());
        VecCost = TTIRef.getCmpSelInstrCost(
            E->getOpcode(), VecTy, MaskTy, CI0->getPredicate(), CostKind,
            {TTI::OK_AnyValue, TTI::OP_None}, {TTI::OK_AnyValue, TTI::OP_None},
            VL0);
        VecCost += TTIRef.getCmpSelInstrCost(
            E->getOpcode(), VecTy, MaskTy,
            cast<CmpInst>(E->getAltOp())->getPredicate(), CostKind,
            {TTI::OK_AnyValue, TTI::OP_None}, {TTI::OK_AnyValue, TTI::OP_None},
            E->getAltOp());
      } else {
        Type *SrcSclTy = E->getMainOp()->getOperand(0)->getType();
        auto *SrcTy = getWidenedType(SrcSclTy, VL.size());
        if (SrcSclTy->isIntegerTy() && ScalarTy->isIntegerTy()) {
          auto SrcIt = MinBWs.find(getOperandEntry(E, 0));
          unsigned BWSz = DL->getTypeSizeInBits(ScalarTy);
          unsigned SrcBWSz =
              DL->getTypeSizeInBits(E->getMainOp()->getOperand(0)->getType());
          if (SrcIt != MinBWs.end()) {
            SrcBWSz = SrcIt->second.first;
            SrcSclTy = IntegerType::get(SrcSclTy->getContext(), SrcBWSz);
            SrcTy = getWidenedType(SrcSclTy, VL.size());
          }
          if (BWSz <= SrcBWSz) {
            if (BWSz < SrcBWSz)
              VecCost =
                  TTIRef.getCastInstrCost(Instruction::Trunc, VecTy, SrcTy,
                                          TTI::CastContextHint::None, CostKind);
            LLVM_DEBUG({
              dbgs()
                  << "SLP: alternate extension, which should be truncated.\n";
              E->dump();
            });
            return VecCost;
          }
        }
        VecCost = TTIRef.getCastInstrCost(E->getOpcode(), VecTy, SrcTy,
                                          TTI::CastContextHint::None, CostKind);
        VecCost +=
            TTIRef.getCastInstrCost(E->getAltOpcode(), VecTy, SrcTy,
                                    TTI::CastContextHint::None, CostKind);
      }
      SmallVector<int> Mask;
      E->buildAltOpShuffleMask(
          [&](Instruction *I) {
            assert(E->getMatchingMainOpOrAltOp(I) &&
                   "Unexpected main/alternate opcode");
            return isAlternateInstruction(I, E->getMainOp(), E->getAltOp(),
                                          *TLI);
          },
          Mask);
      VecCost += getShuffleCost(TTIRef, TargetTransformInfo::SK_PermuteTwoSrc,
                                cast<VectorType>(FinalVecTy), CostKind, Mask);
      // Patterns like [fadd,fsub] can be combined into a single instruction
      // in x86. Reordering them into [fsub,fadd] blocks this pattern. So we
      // need to take into account their order when looking for the most used
      // order. Linearized chains emit no alternate-ops pattern.
      if (E->hasReassocScalars())
        return VecCost;
      unsigned Opcode0 = E->getOpcode();
      unsigned Opcode1 = E->getAltOpcode();
      SmallBitVector OpcodeMask(
          getAltInstrMask(E->Scalars, ScalarTy, Opcode0, Opcode1));
      // If this pattern is supported by the target then we consider the
      // order.
      if (TTIRef.isLegalAltInstr(cast<VectorType>(VecTy), Opcode0, Opcode1,
                                 OpcodeMask)) {
        InstructionCost AltVecCost = TTIRef.getAltInstrCost(
            cast<VectorType>(VecTy), Opcode0, Opcode1, OpcodeMask, CostKind);
        return AltVecCost < VecCost ? AltVecCost : VecCost;
      }
      // TODO: Check the reverse order too.
      return VecCost;
    };
    if (SLPReVec && !E->isAltShuffle())
      return GetCostDiff(
          GetScalarCost, [&](InstructionCost) -> InstructionCost {
            // If a group uses mask in order, the shufflevector can be
            // eliminated by instcombine. Then the cost is 0.
            assert(isa<ShuffleVectorInst>(VL.front()) &&
                   "Not supported shufflevector usage.");
            auto *SV = cast<ShuffleVectorInst>(VL.front());
            unsigned SVNumElements =
                cast<FixedVectorType>(SV->getOperand(0)->getType())
                    ->getNumElements();
            unsigned GroupSize = SVNumElements / SV->getShuffleMask().size();
            for (size_t I = 0, End = VL.size(); I != End; I += GroupSize) {
              ArrayRef<Value *> Group = VL.slice(I, GroupSize);
              int NextIndex = 0;
              if (!all_of(Group, [&](Value *V) {
                    assert(isa<ShuffleVectorInst>(V) &&
                           "Not supported shufflevector usage.");
                    auto *SV = cast<ShuffleVectorInst>(V);
                    int Index;
                    [[maybe_unused]] bool IsExtractSubvectorMask =
                        SV->isExtractSubvectorMask(Index);
                    assert(IsExtractSubvectorMask &&
                           "Not supported shufflevector usage.");
                    if (NextIndex != Index)
                      return false;
                    NextIndex += SV->getShuffleMask().size();
                    return true;
                  }))
                return getShuffleCost(*TTI,
                                      TargetTransformInfo::SK_PermuteSingleSrc,
                                      cast<VectorType>(VecTy), CostKind,
                                      calculateShufflevectorMask(E->Scalars));
            }
            return TTI::TCC_Free;
          });
    return GetCostDiffWithPeeled(GetScalarCost, GetVectorCost);
  }
  case Instruction::Freeze:
    return CommonCost;
  default:
    llvm_unreachable("Unknown instruction");
  }
}
 
bool BoUpSLP::isFullyVectorizableTinyTree(bool ForReduction) const {
  LLVM_DEBUG(dbgs() << "SLP: Check whether the tree with height "
                    << VectorizableTree.size() << " is fully vectorizable .\n");
 
  auto &&AreVectorizableGathers = [this](const TreeEntry *TE, unsigned Limit) {
    SmallVector<int> Mask;
    return TE->isGather() &&
           !any_of(TE->Scalars,
                   [this](Value *V) { return EphValues.contains(V); }) &&
           (allConstant(TE->Scalars) || isSplat(TE->Scalars) ||
            TE->Scalars.size() < Limit ||
            // Nodes with copyable lanes may mix in non-extract lanes, which
            // are not representable as a shuffle of the source vector.
            (((TE->hasState() &&
               TE->getOpcode() == Instruction::ExtractElement &&
               !TE->hasCopyableElements()) ||
              all_of(TE->Scalars, IsaPred<ExtractElementInst, UndefValue>)) &&
             isFixedVectorShuffle(TE->Scalars, Mask, AC)) ||
            (TE->hasState() && TE->getOpcode() == Instruction::Load &&
             !TE->isAltShuffle()) ||
            any_of(TE->Scalars, IsaPred<LoadInst>));
  };
 
  // We only handle trees of heights 1 and 2.
  if (VectorizableTree.size() == 1 &&
      (VectorizableTree[0]->State == TreeEntry::Vectorize ||
       VectorizableTree[0]->State == TreeEntry::StridedVectorize ||
       VectorizableTree[0]->State == TreeEntry::ExpandVectorize ||
       VectorizableTree[0]->State == TreeEntry::CompressVectorize ||
       VectorizableTree[0]->State == TreeEntry::BlendedLoadVectorize ||
       (ForReduction &&
        AreVectorizableGathers(VectorizableTree[0].get(),
                               VectorizableTree[0]->Scalars.size()) &&
        VectorizableTree[0]->getVectorFactor() > 2)))
    return true;
 
  if (VectorizableTree.size() != 2)
    return false;
 
  // Handle splat and all-constants stores. Also try to vectorize tiny trees
  // with the second gather nodes if they have less scalar operands rather than
  // the initial tree element (may be profitable to shuffle the second gather)
  // or they are extractelements, which form shuffle.
  if (VectorizableTree[0]->State == TreeEntry::Vectorize &&
      AreVectorizableGathers(VectorizableTree[1].get(),
                             VectorizableTree[0]->Scalars.size()))
    return true;
 
  // Gathering cost would be too much for tiny trees.
  if (VectorizableTree[0]->isGather() ||
      (VectorizableTree[1]->isGather() &&
       VectorizableTree[0]->State != TreeEntry::ScatterVectorize &&
       VectorizableTree[0]->State != TreeEntry::StridedVectorize &&
       VectorizableTree[0]->State != TreeEntry::ExpandVectorize &&
       VectorizableTree[0]->State != TreeEntry::CompressVectorize &&
       VectorizableTree[0]->State != TreeEntry::BlendedLoadVectorize))
    return false;
 
  return true;
}
 
bool BoUpSLP::isTreeTinyAndNotFullyVectorizable(bool ForReduction) const {
  if (!DebugCounter::shouldExecute(VectorizedGraphs))
    return true;
 
  // If we are revectorizing reduction, it may result in same reduction pattern
  // with shuffles as leaves of the reduction. Prevent SLP from revectorizing
  // that shuffle pattern.
  if (SLPReVec && ForReduction && VectorizableTree.size() == 3 &&
      VectorizableTree[0]->State == TreeEntry::Vectorize &&
      VectorizableTree[0]->getOpcode() == Instruction::ShuffleVector &&
      VectorizableTree[1]->isGather() &&
      isSplat(VectorizableTree[1]->Scalars) &&
      VectorizableTree[2]->isGather() &&
      allConstant(VectorizableTree[2]->Scalars)) {
    LLVM_DEBUG(dbgs() << "SLP: Rejecting reduction tree with 3 nodes(shuffle "
                         "as root and remaining are gather nodes).\n");
    return true;
  }
 
  // Graph is empty - do nothing.
  if (VectorizableTree.empty()) {
    assert(ExternalUses.empty() && "We shouldn't have any external users");
 
    return true;
  }
 
  // FIXME: support buildvector of the gather nodes with struct types.
  if (any_of(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
        return TE->isGather() &&
               isa<StructType>(TE->Scalars.front()->getType());
      })) {
    LLVM_DEBUG(
        dbgs() << "SLP: rejecting tree with buildvector struct values of size "
               << VectorizableTree.size() << ".\n");
    return true;
  }
 
  // Cache values from the root node and the cost-threshold options to avoid
  // re-querying them inside hot predicates below.
  const unsigned TreeSize = VectorizableTree.size();
  const TreeEntry &Front = getRootNode();
  const bool FrontIsGather = Front.isGather();
  const bool FrontHasState = Front.hasState();
  const unsigned FrontOpcode = FrontHasState ? Front.getOpcode() : 0u;
  const bool ThresholdSet = SLPCostThreshold.getNumOccurrences() > 0;
  const bool ThresholdNonNegative = SLPCostThreshold >= 0;
 
  constexpr unsigned Limit = 4;
  constexpr unsigned LargeTree = 20;
  constexpr unsigned LimitTreeSize = 36;
 
  // The remaining size-1/size-<=MinTreeSize early bail-outs only apply to
  // non-reduction trees; group them under a single guard to avoid 3 separate
  // !ForReduction short-circuits when reducing.
  if (!ForReduction) {
    // Single gather node: bail out for ExtractElement or any node containing a
    // real Instruction scalar.
    if (TreeSize == 1 && FrontIsGather) {
      if (FrontHasState && FrontOpcode == Instruction::ExtractElement)
        return true;
      if (any_of(Front.Scalars, IsaPred<Instruction>))
        return true;
    }
    if (TreeSize <= MinTreeSize &&
        all_of(VectorizableTree, [](const std::unique_ptr<TreeEntry> &TE) {
          return TE->isGather() || TE->State == TreeEntry::SplitVectorize;
        }))
      return true;
    if (TreeSize == 1 && SLPCostThreshold < 0 && FrontHasState &&
        FrontOpcode == Instruction::ExtractElement &&
        (Front.getVectorFactor() == 2 ||
         all_of(
             Front.Scalars,
             [&](Value *V) {
               auto *I = dyn_cast<Instruction>(V);
               return !I || !areAllUsersVectorized(I, UserIgnoreList);
             })))
      return true;
  }
  // No need to vectorize inserts of gathered values.
  if (TreeSize == 2 && isa<InsertElementInst>(Front.Scalars[0]) &&
      VectorizableTree[1]->isGather() &&
      (VectorizableTree[1]->getVectorFactor() <= 2 ||
       !(isSplat(VectorizableTree[1]->Scalars) ||
         allConstant(VectorizableTree[1]->Scalars))))
    return true;
 
  // The tree with only 3 nodes, where 2 last are gathers/buildvectors, not
  // profitable for vectorization.
  if (TreeSize == 3 && SLPCostThreshold == 0 &&
      (!ForReduction || Front.getVectorFactor() <= 2) &&
      all_of(ArrayRef(VectorizableTree).drop_front(),
             [&](const std::unique_ptr<TreeEntry> &TE) {
               return TE->isGather() && TE->getVectorFactor() <= Limit &&
                      !all_of(
                          TE->Scalars,
                          IsaPred<ExtractElementInst, UndefValue, Constant>);
             }))
    return true;
 
  // All remaining bail-out heuristics require !ForReduction. Group them under
  // a single guard so reduction trees skip them with one branch instead of one
  // per check.
  if (!ForReduction) {
    // The single vectorized node, fed by the buildvectors, only repacks the
    // scalars, the gathers and the extracts are paid for the one vector
    // instruction. The vector call replaces the several expensive scalar ones,
    // while the store merges the lanes instead of extracting them back.
    if (TreeSize > 1 && ThresholdNonNegative &&
        (!FrontHasState || (FrontOpcode != Instruction::Call &&
                            FrontOpcode != Instruction::Store)) &&
        count_if(VectorizableTree,
                 [](const std::unique_ptr<TreeEntry> &TE) {
                   return TE->State == TreeEntry::Vectorize;
                 }) == 1 &&
        all_of(VectorizableTree, [](const std::unique_ptr<TreeEntry> &TE) {
          return TE->State == TreeEntry::Vectorize ||
                 (TE->isGather() && !isSplat(TE->Scalars) &&
                  !all_of(TE->Scalars,
                          IsaPred<ExtractElementInst, UndefValue, Constant>));
        }))
      return true;
 
    // If the graph includes only PHI nodes and gathers, it is defnitely not
    // profitable for the vectorization, we can skip it, if the cost threshold
    // is default. The cost of vectorized PHI nodes is almost always 0 + the
    // cost of gathers/buildvectors.
    if (!ThresholdSet &&
        all_of(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
          const bool IsGather = TE->isGather();
          const bool HasState = TE->hasState();
          const unsigned Op = HasState ? TE->getOpcode() : 0u;
          if (IsGather && (!HasState || Op != Instruction::ExtractElement) &&
              static_cast<unsigned>(
                  count_if(TE->Scalars, IsaPred<ExtractElementInst>)) <= Limit)
            return true;
          return HasState && Op == Instruction::PHI;
        }))
      return true;
 
    // Do not vectorize small tree of phis only, if all vector phis are also
    // gathered.
    if (ThresholdSet && TreeSize <= Limit) {
      bool HasVectorPhi = false;
      auto Compatible = [&](const std::unique_ptr<TreeEntry> &TE) {
        const bool IsGather = TE->isGather();
        const bool HasState = TE->hasState();
        const unsigned Op = HasState ? TE->getOpcode() : 0u;
        if (IsGather && (!HasState || Op != Instruction::ExtractElement) &&
            static_cast<unsigned>(
                count_if(TE->Scalars, IsaPred<ExtractElementInst>)) <= Limit)
          return true;
        if (!HasState)
          return false;
        if (Op == Instruction::InsertElement)
          return true;
        if (Op != Instruction::PHI)
          return false;
        if (TE->State == TreeEntry::Vectorize)
          HasVectorPhi = true;
        return all_of(TE->Scalars, [&](Value *V) {
          return isa<PoisonValue>(V) || MustGather.contains(V);
        });
      };
      if (all_of(VectorizableTree, Compatible) && HasVectorPhi)
        return true;
    }
 
    // PHI nodes only and gathers cannot be vectorized, skip.
    if (ThresholdNonNegative) {
      const bool IsLargeTree = TreeSize >= LargeTree;
      bool HasSingleLoad = false;
      if (all_of(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
            const bool IsGather = TE->isGather();
            const bool HasState = TE->hasState();
            const unsigned Op = HasState ? TE->getOpcode() : 0u;
            // HasSingleLoad/PrevLoad are only consulted in the
            // IsLargeTree branch; skip the bookkeeping otherwise.
            if (IsLargeTree) {
              const bool PrevLoad = HasSingleLoad;
              HasSingleLoad |=
                  HasState && !IsGather &&
                  (Op == Instruction::Load || TE->hasCopyableElements()) &&
                  (TE->getVectorFactor() > 2 || TE->ReorderIndices.empty());
              if (HasState) {
                if (Op == Instruction::PHI)
                  return true;
                if (TE->getVectorFactor() <= Limit &&
                    (Op == Instruction::Store ||
                     (Op == Instruction::Load && !PrevLoad)))
                  return true;
              }
            } else if (HasState && Op == Instruction::PHI) {
              return true;
            }
            return IsGather && (!HasState || Op != Instruction::ExtractElement);
          }))
        return true;
 
      // Single non-phi vector node - skip the tree.
      if (TreeSize >= 5 && Front.getVectorFactor() <= 2 &&
          Front.Scalars.front()->getType()->isIntegerTy()) {
        bool VectorNodeFound = false;
        bool AnyNonConst = false;
        if (all_of(VectorizableTree,
                   [&](const std::unique_ptr<TreeEntry> &TE) {
                     if (TE->State == TreeEntry::Vectorize && TE->hasState()) {
                       const unsigned Op = TE->getOpcode();
                       if (Op == Instruction::PHI ||
                           !TE->ReorderIndices.empty())
                         return true;
                       if (VectorNodeFound)
                         return false;
                       VectorNodeFound = true;
                       return true;
                     }
                     // Once AnyNonConst is true, skip the O(n) allConstant
                     // walk for subsequent entries.
                     if (!AnyNonConst)
                       AnyNonConst = !allConstant(TE->Scalars);
                     return TE->isGather() ||
                            TE->State == TreeEntry::SplitVectorize;
                   }) &&
            AnyNonConst)
          return true;
      }
    }
 
    // Common predicate for "phis, buildvectors, split nodes and small nodes
    // with reuses" used by the two checks below. Cheap checks are evaluated
    // before expensive Scalars walks.
    auto IsBenignNode = [&](const TreeEntry &TE) {
      if (TE.State == TreeEntry::SplitVectorize)
        return true;
      const bool IsGather = TE.isGather();
      const bool HasState = TE.hasState();
      if (HasState) {
        const unsigned Op = TE.getOpcode();
        if (Op == Instruction::PHI)
          return true;
        const unsigned ScalarsSize = TE.Scalars.size();
        if (TE.Idx == 0 && ScalarsSize == 2 && Op == Instruction::ICmp &&
            TreeSize > LimitTreeSize)
          return true;
        if (ScalarsSize == 2 &&
            (!TE.ReuseShuffleIndices.empty() || !TE.ReorderIndices.empty() ||
             TE.isAltShuffle()))
          return true;
        if (TE.hasCopyableElements() &&
            static_cast<unsigned>(count_if(
                TE.Scalars, IsaPred<PHINode, Constant>)) >= ScalarsSize / 2)
          return true;
      }
      return IsGather && none_of(TE.Scalars, IsaPred<ExtractElementInst>);
    };
 
    // If the tree contains only phis, buildvectors, split nodes and
    // small nodes with reuses, we can skip it.
    if (!ThresholdSet) {
      SmallVector<const TreeEntry *> StoreLoadNodes;
      unsigned NumGathers = 0;
      if (all_of(VectorizableTree,
                 [&](const std::unique_ptr<TreeEntry> &TE) {
                   const bool IsGather = TE->isGather();
                   if (!IsGather && TE->hasState()) {
                     const unsigned Op = TE->getOpcode();
                     if (Op == Instruction::Load || Op == Instruction::Store) {
                       StoreLoadNodes.push_back(TE.get());
                       return true;
                     }
                   }
                   if (IsGather)
                     ++NumGathers;
                   return IsBenignNode(*TE);
                 }) &&
          (StoreLoadNodes.empty() ||
           (TreeSize > LimitTreeSize * StoreLoadNodes.size() &&
            (NumGathers > 0 ||
             none_of(StoreLoadNodes, [&](const TreeEntry *TE) {
               return TE->getOpcode() == Instruction::Store ||
                      all_of(TE->Scalars, [&](Value *V) {
                        return !isa<LoadInst>(V) ||
                               areAllUsersVectorized(cast<Instruction>(V));
                      });
             })))))
        return true;
    }
 
    // If the tree contains only phis, buildvectors, split nodes and
    // small nodes with reuses, we can skip it.
    if (ThresholdNonNegative && TreeSize > LimitTreeSize) {
      const TreeEntry *VectorNode = nullptr;
      if (all_of(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
            if (!TE->isGather() && TE->hasState() &&
                TE->State != TreeEntry::SplitVectorize &&
                TE->getOpcode() != Instruction::PHI) {
              if (VectorNode)
                return false;
              VectorNode = TE.get();
              return true;
            }
            return IsBenignNode(*TE);
          }))
        return true;
    }
 
    // If the tree contains only buildvector, 2 non-buildvectors (with root
    // user tree node) and other buildvectors, we can skip it.
    if (ThresholdSet && TreeSize >= Limit &&
        Front.State == TreeEntry::SplitVectorize &&
        count_if(ArrayRef(VectorizableTree).drop_front(),
                 [](const std::unique_ptr<TreeEntry> &TE) {
                   return !TE->isGather() && TE->UserTreeIndex.UserTE &&
                          TE->UserTreeIndex.UserTE->Idx == 0;
                 }) == 2)
      return true;
 
    // If the tree contains only vectorization of the phi node from the
    // buildvector - skip it.
    if (ThresholdSet && TreeSize > 2 && Front.State == TreeEntry::Vectorize &&
        FrontOpcode == Instruction::InsertElement &&
        VectorizableTree[1]->State == TreeEntry::Vectorize &&
        VectorizableTree[1]->getOpcode() == Instruction::PHI &&
        all_of(ArrayRef(VectorizableTree).drop_front(2),
               [](const std::unique_ptr<TreeEntry> &TE) {
                 return TE->isGather();
               }))
      return true;
  }
 
  // We can vectorize the tree if its size is greater than or equal to the
  // minimum size specified by the MinTreeSize command line option.
  if (TreeSize >= MinTreeSize)
    return false;
 
  // If we have a tiny tree (a tree whose size is less than MinTreeSize), we
  // can vectorize it if we can prove it fully vectorizable.
  if (isFullyVectorizableTinyTree(ForReduction))
    return false;
 
  // Check if any of the gather node forms an insertelement buildvector
  // somewhere. TreeSize >= 1 is guaranteed, so the multi-node case reduces to
  // a simple TreeSize > 1 short-circuit.
  // A gather with copyable lanes is not a real instruction node; do not let
  // its state qualify it as a buildvector-forming node.
  const bool IsAllowedSingleBVNode =
      TreeSize > 1 ||
      (FrontHasState && !Front.isAltShuffle() && !Front.hasCopyableElements() &&
       FrontOpcode != Instruction::PHI &&
       FrontOpcode != Instruction::GetElementPtr &&
       allSameBlock(Front.Scalars));
  if (any_of(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
        return TE->isGather() && all_of(TE->Scalars, [&](Value *V) {
                 return isa<ExtractElementInst, Constant>(V) ||
                        (IsAllowedSingleBVNode &&
                         !V->hasNUsesOrMore(UsesLimit) &&
                         any_of(V->users(), IsaPred<InsertElementInst>));
               });
      }))
    return false;
 
  const TreeEntry &Back = *VectorizableTree.back();
  if (Back.isGather() && Back.hasState() && Back.isAltShuffle()) {
    const unsigned BackVF = Back.getVectorFactor();
    if (BackVF > 2 && allSameBlock(Back.Scalars) &&
        !Back.Scalars.front()->getType()->isVectorTy() &&
        TTI->getScalarizationOverhead(
            cast<VectorType>(
                getWidenedType(Back.Scalars.front()->getType(), BackVF)),
            APInt::getAllOnes(BackVF),
            /*Insert=*/true, /*Extract=*/false, CostKind) > -SLPCostThreshold)
      return false;
  }
 
  // Otherwise, we can't vectorize the tree. It is both tiny and not fully
  // vectorizable.
  return true;
}
 
bool BoUpSLP::isTreeNotExtendable() const {
  if (getCanonicalGraphSize() != getTreeSize()) {
    constexpr unsigned SmallTree = 3;
    if (getRootNode().isNonPowOf2Vec() &&
        getCanonicalGraphSize() <= SmallTree &&
        count_if(ArrayRef(VectorizableTree).drop_front(getCanonicalGraphSize()),
                 [](const std::unique_ptr<TreeEntry> &TE) {
                   return TE->isGather() && TE->hasState() &&
                          TE->getOpcode() == Instruction::Load &&
                          !allSameBlock(TE->Scalars);
                 }) == 1)
      return true;
    return false;
  }
  bool Res = false;
  for (unsigned Idx : seq<unsigned>(getTreeSize())) {
    TreeEntry &E = *VectorizableTree[Idx];
    if (E.State == TreeEntry::SplitVectorize)
      return false;
    if (!E.isGather())
      continue;
    if ((E.hasState() && E.getOpcode() != Instruction::Load) ||
        (!E.hasState() &&
         all_of(E.Scalars, IsaPred<ExtractElementInst, LoadInst>)) ||
        (isa<ExtractElementInst>(E.Scalars.front()) &&
         getSameOpcode(ArrayRef(E.Scalars).drop_front(), *TLI).valid()))
      return false;
    if (isSplat(E.Scalars) || allConstant(E.Scalars))
      continue;
    Res = true;
  }
  return Res;
}
 
InstructionCost BoUpSLP::getSpillCost() {
  // Walk the vectorizable tree from the root towards its leaves, tracking
  // which vectorized operand values would be live across each tree edge
  // (i.e. between the last instruction of an operand entry and the last
  // instruction of its user entry). When the live range crosses a call
  // instruction that is not part of the vectorized tree, query TTI for the
  // cost of keeping the value live across it (for example, if spills and
  // fills are required).
 
  const TreeEntry *Root = &getRootNode();
  if (Root->isGather())
    return 0;
 
  InstructionCost Cost = 0;
  SmallDenseMap<const TreeEntry *, SmallVector<const TreeEntry *>>
      EntriesToOperands;
  SmallDenseMap<const TreeEntry *, Instruction *> EntriesToLastInstruction;
  SmallPtrSet<const Instruction *, 8> LastInstructions;
  SmallPtrSet<const TreeEntry *, 8> ScalarOrPseudoEntries;
  for (const auto &TEPtr : VectorizableTree) {
    if (TEPtr->CombinedOp == TreeEntry::ReducedBitcast ||
        TEPtr->CombinedOp == TreeEntry::ReducedBitcastBSwap ||
        TEPtr->CombinedOp == TreeEntry::ReducedBitcastLoads ||
        TEPtr->CombinedOp == TreeEntry::ReducedBitcastBSwapLoads ||
        TEPtr->CombinedOp == TreeEntry::ReducedCmpBitcast) {
      ScalarOrPseudoEntries.insert(TEPtr.get());
      continue;
    }
    if (!TEPtr->isGather()) {
      Instruction *LastInst = &getLastInstructionInBundle(TEPtr.get());
      EntriesToLastInstruction.try_emplace(TEPtr.get(), LastInst);
      LastInstructions.insert(LastInst);
    }
    if (TEPtr->UserTreeIndex)
      EntriesToOperands[TEPtr->UserTreeIndex.UserTE].push_back(TEPtr.get());
  }
 
  // Cache NoCallIntrinsicOrDoesNotReturn results - the same intrinsic call may
  // be queried many times during the spill cost scan, and each computation
  // involves two potentially expensive TTI virtual calls.
  SmallDenseMap<const IntrinsicInst *, bool> NoCallIntrinsicCache;
  auto NoCallIntrinsicOrDoesNotReturn = [this, &NoCallIntrinsicCache](
                                            const Instruction *I) {
    const auto *CB = dyn_cast<CallBase>(I);
    if (!CB)
      return false;
    if (CB->doesNotReturn())
      return true;
    const auto *II = dyn_cast<IntrinsicInst>(CB);
    if (!II)
      return false;
    if (II->isAssumeLikeIntrinsic())
      return true;
    auto [It, Inserted] = NoCallIntrinsicCache.try_emplace(II);
    if (!Inserted)
      return It->second;
    IntrinsicCostAttributes ICA(II->getIntrinsicID(), *II);
    InstructionCost IntrCost = TTI->getIntrinsicInstrCost(ICA, CostKind);
    InstructionCost CallCost = TTI->getCallInstrCost(
        nullptr, II->getType(), ICA.getArgTypes(), CostKind);
    bool Res = IntrCost < CallCost;
    It->second = Res;
    return Res;
  };
 
  // Maps last instruction in the entry to the last instruction for the one of
  // operand entries and the flag. If the flag is true, there are no calls in
  // between these instructions.
  SmallDenseMap<const Instruction *, PointerIntPair<const Instruction *, 1>>
      CheckedInstructions;
  unsigned Budget = 0;
  const unsigned BudgetLimit =
      ScheduleRegionSizeBudget / VectorizableTree.size();
  auto CheckForNonVecCallsInSameBlock = [&](Instruction *First,
                                            const Instruction *Last) {
    assert(First->getParent() == Last->getParent() &&
           "Expected instructions in same block.");
    if (auto It = CheckedInstructions.find(Last);
        It != CheckedInstructions.end()) {
      const Instruction *Checked = It->second.getPointer();
      const bool NoCallsInCachedRange = It->second.getInt() != 0;
      if (Checked == First)
        return NoCallsInCachedRange;
      if (Checked->comesBefore(First))
        // In every cached state (full clean scan, call-found, or
        // budget-exhausted) the region strictly above `Checked` up to `Last`
        // was inspected and proved call-free. Since `First` is above
        // `Checked`, the queried range [First, Last] is contained in that
        // call-free region, regardless of whether bit is 0 or 1.
        return true;
      Last = Checked;
    } else if (Last == First || Last->comesBefore(First)) {
      // Empty range.
      return true;
    }
    BasicBlock::const_reverse_iterator InstIt =
                                           ++First->getIterator().getReverse(),
                                       PrevInstIt =
                                           Last->getIterator().getReverse();
    SmallVector<const Instruction *> LastInstsInRange;
    while (InstIt != PrevInstIt && Budget <= BudgetLimit) {
      // Debug information does not impact spill cost.
      // Vectorized calls, represented as vector intrinsics, do not impact spill
      // cost.
      if (const auto *CB = dyn_cast<CallBase>(&*PrevInstIt);
          CB && !NoCallIntrinsicOrDoesNotReturn(CB) && !isVectorized(CB)) {
        for (const Instruction *LastInst : LastInstsInRange)
          CheckedInstructions.try_emplace(LastInst, &*PrevInstIt, 0);
        return false;
      }
      if (LastInstructions.contains(&*PrevInstIt))
        LastInstsInRange.push_back(&*PrevInstIt);
 
      ++PrevInstIt;
      ++Budget;
    }
    // If we reached the scan's lower bound (`PrevInstIt == InstIt`) then the
    // whole [First, Last] range was inspected and found call-free, even if
    // Budget just overflowed at the very last step; do not mislabel such a
    // completed scan as "has call".
    const bool Completed = PrevInstIt == InstIt;
    const bool NoCallsInRange = Completed || Budget <= BudgetLimit;
    for (const Instruction *LastInst : LastInstsInRange)
      CheckedInstructions.try_emplace(
          LastInst, Completed ? First : &*PrevInstIt, NoCallsInRange ? 1 : 0);
    return NoCallsInRange;
  };
  // The spill/reload is executed once per execution of the call, so the
  // cost is scaled by the trip count of the loop containing the call, even
  // for hoisted loop-invariant values defined outside of it.
  auto GetSpillScale = [&](const BasicBlock *BB) {
    return getLoopNestScale(LI->getLoopFor(BB));
  };
  auto AddCosts = [&](const TreeEntry *Op, uint64_t Scale) {
    if (ScalarOrPseudoEntries.contains(Op))
      return;
    Type *ScalarTy = Op->Scalars.front()->getType();
    auto It = MinBWs.find(Op);
    if (It != MinBWs.end())
      ScalarTy = IntegerType::get(ScalarTy->getContext(), It->second.first);
    auto *VecTy = getWidenedType(ScalarTy, Op->getVectorFactor());
    InstructionCost KeepLiveCost = TTI->getCostOfKeepingLiveOverCall(VecTy);
    KeepLiveCost *= Scale;
    Cost += KeepLiveCost;
    if (ScalarTy->isVectorTy()) {
      // Handle revec dead vector instructions.
      Cost -= Op->Scalars.size() * TTI->getCostOfKeepingLiveOverCall(ScalarTy) *
              Scale;
    }
  };
  // Memoize the relationship between blocks, i.e. the spill scale if every
  // path between the blocks crosses a non-vectorized call, 0 if there is (at
  // least one) call-free path. This allows to skip the analysis of the same
  // block paths multiple times.
  SmallDenseMap<std::pair<const BasicBlock *, const BasicBlock *>, uint64_t>
      ParentOpParentToPreds;
  // Memoize whether a basic block contains a non-terminator no-return call.
  // Such blocks are dead-end paths in normal control flow (execution does not
  // exit them past the no-return call), so the block is excluded from the
  // spill cost analysis. Terminator no-return calls (invoke/callbr) are not
  // block-killing because they still have live CFG successors (e.g. the
  // unwind destination of an invoke).
  SmallDenseMap<const BasicBlock *, bool> BlockHasNoReturnCallCache;
  auto BlockHasNoReturnCall = [&](const BasicBlock *BB) {
    auto [It, Inserted] = BlockHasNoReturnCallCache.try_emplace(BB, false);
    if (!Inserted)
      return It->second;
    for (const Instruction &I : *BB) {
      const auto *CB = dyn_cast<CallBase>(&I);
      if (CB && CB->doesNotReturn() && !CB->isTerminator()) {
        It->second = true;
        return true;
      }
    }
    return false;
  };
  // Memoize whether a loop's body (all blocks of the loop, including
  // sub-loops) contains any non-vec call.
  SmallDenseMap<const Loop *, bool> LoopBodyHasNonVecCall;
  auto LoopBodyHasCall = [&](const Loop *L) {
    if (auto It = LoopBodyHasNonVecCall.find(L);
        It != LoopBodyHasNonVecCall.end())
      return It->second;
    for (BasicBlock *BB : L->blocks()) {
      if (isa<CatchSwitchInst>(BB->getTerminator()))
        continue;
      // Blocks containing a no-return call are dead-end paths and never
      // actually flow back through the loop's back-edge, so their calls do
      // not keep loop-invariant vector values live across calls.
      if (BlockHasNoReturnCall(BB))
        continue;
      for (const Instruction &I : *BB) {
        const auto *CB = dyn_cast<CallBase>(&I);
        if (!CB || NoCallIntrinsicOrDoesNotReturn(CB) || isVectorized(CB))
          continue;
        LoopBodyHasNonVecCall.try_emplace(L, true);
        return true;
      }
    }
    LoopBodyHasNonVecCall.try_emplace(L, false);
    return false;
  };
  auto CheckPredecessors = [&](BasicBlock *Root, BasicBlock *Pred,
                               BasicBlock *OpParent) {
    auto Key = std::make_pair(Root, OpParent);
    if (auto It = ParentOpParentToPreds.find(Key);
        It != ParentOpParentToPreds.end())
      return It->second;
    uint64_t Res = 0;
    scope_exit Cleanup([&]() { ParentOpParentToPreds.try_emplace(Key, Res); });
    // If Op is loop-invariant, a call anywhere in the loop body forces a spill,
    // even when a call-free forward path from Root back to OpParent exists on
    // the first iteration. Find the outermost such enclosing loop and reject if
    // its body contains a non-vec call.
    const Loop *L = LI->getLoopFor(Root);
    const Loop *Outermost = nullptr;
    while (L && !L->contains(OpParent)) {
      Outermost = L;
      L = L->getParentLoop();
    }
    if (Outermost && LoopBodyHasCall(Outermost)) {
      Res = getLoopNestScale(Outermost);
      return Res;
    }
    SmallVector<BasicBlock *> Worklist;
    if (Pred)
      Worklist.push_back(Pred);
    else
      Worklist.append(pred_begin(Root), pred_end(Root));
    SmallPtrSet<const BasicBlock *, 16> Visited;
    // With "at least one call-free path" semantics we can only reliably
    // memoize the exact (Root, OpParent) query. Pairs for intermediate
    // blocks that were visited during the BFS are not necessarily
    // call-free-reachable to OpParent themselves - we may have reached
    // OpParent through a *sibling* path that bypassed them.
    // We return 0 (no spill cost) if at least one backward path from
    // some predecessor of Root back to OpParent is call-free. Only when
    // *every* such path goes through a non-vec call do we charge the spill
    // cost: only then is it actually necessary to keep the vectorized value
    // live across a call and therefore spill/reload it.
    //
    // A BB is only explored further (its predecessors added to the worklist)
    // when it is itself call-free and not strictly dominated by Root (blocks
    // dominated by Root are only reachable via loop back-edges - they sit
    // *after* Root in forward execution and must not be counted).
    //
    // If we ever pop OpParent from the worklist, we have reached it through
    // a chain of call-free, non-dominated blocks: a call-free path exists
    // and we return 0. If the worklist is exhausted or the scan budget
    // overflows without reaching OpParent, no call-free path was found and
    // we return the scale of the loop containing Root, so the caller charges
    // the spill cost.
    while (!Worklist.empty()) {
      BasicBlock *BB = Worklist.pop_back_val();
      if (BB == OpParent)
        return Res;
      if (!Visited.insert(BB).second)
        continue;
      // Blocks strictly dominated by Root are reached only *after* Root in
      // forward execution (via loop back-edges); skip them and their
      // dominated predecessors.
      if (DT->properlyDominates(Root, BB))
        continue;
      // A block containing a no-return call cannot reach Root via the
      // forward edge being analyzed: execution does not continue past the
      // no-return call, so the BB -> ... -> Root path is dead. Drop the
      // block from the analysis without following its predecessors.
      if (BlockHasNoReturnCall(BB))
        continue;
      auto Pair = std::make_pair(BB, OpParent);
      if (auto It = ParentOpParentToPreds.find(Pair);
          It != ParentOpParentToPreds.end()) {
        // BB is known to reach OpParent via a call-free path.
        if (It->second == 0)
          return Res;
        // BB is known to be blocked from OpParent by calls; keep checking
        // other paths.
        continue;
      }
      unsigned BlockSize = BB->size();
      if (BlockSize > static_cast<unsigned>(ScheduleRegionSizeBudget))
        continue;
      Budget += BlockSize;
      if (Budget > BudgetLimit)
        break;
      if (!isa<CatchSwitchInst>(BB->getTerminator()) &&
          !CheckForNonVecCallsInSameBlock(&*BB->getFirstNonPHIOrDbgOrAlloca(),
                                          BB->getTerminator()))
        continue;
      Worklist.append(pred_begin(BB), pred_end(BB));
    }
    Res = GetSpillScale(Root);
    return Res;
  };
  SmallVector<const TreeEntry *> LiveEntries(1, Root);
  auto FindNonScalarParentEntry = [&](const TreeEntry *E) -> const TreeEntry * {
    assert(ScalarOrPseudoEntries.contains(E) &&
           "Expected scalar or pseudo entry.");
    const TreeEntry *Entry = E;
    while (Entry->UserTreeIndex) {
      Entry = Entry->UserTreeIndex.UserTE;
      if (!ScalarOrPseudoEntries.contains(Entry))
        return Entry;
    }
    return nullptr;
  };
  while (!LiveEntries.empty()) {
    const TreeEntry *Entry = LiveEntries.pop_back_val();
    const auto OpIt = EntriesToOperands.find(Entry);
    if (OpIt == EntriesToOperands.end())
      continue;
    ArrayRef<const TreeEntry *> Operands = OpIt->second;
    if (Operands.empty())
      continue;
    if (ScalarOrPseudoEntries.contains(Entry)) {
      Entry = FindNonScalarParentEntry(Entry);
      if (!Entry) {
        for (const TreeEntry *Op : Operands) {
          if (!Op->isGather())
            LiveEntries.push_back(Op);
        }
        continue;
      }
    }
    Instruction *LastInst = EntriesToLastInstruction.at(Entry);
    BasicBlock *Parent = LastInst->getParent();
    for (const TreeEntry *Op : Operands) {
      if (!Op->isGather())
        LiveEntries.push_back(Op);
      if (ScalarOrPseudoEntries.contains(Op))
        continue;
      if (Entry->State == TreeEntry::SplitVectorize ||
          (Op->isGather() && allConstant(Op->Scalars)))
        continue;
      // A gather with all loop-invariant lanes is hoisted to the loop
      // preheader by optimizeGatherSequence, so its vector value becomes live
      // across any non-vectorized call in the loop body. Charge it like any
      // other vectorized value live over a call instead of skipping it.
      if (Entry->hasState() && Entry->getOpcode() != Instruction::PHI &&
          Op->isGather()) {
        if (const Loop *L = LI->getLoopFor(Parent);
            L && L->getLoopPreheader() && LoopBodyHasCall(L) &&
            all_of(Op->Scalars, [&](Value *V) {
              return !isa<Instruction>(V) || L->isLoopInvariant(V);
            }))
          AddCosts(Op, GetSpillScale(Parent));
        continue;
      }
      Budget = 0;
      BasicBlock *Pred = nullptr;
      if (auto *Phi = dyn_cast<PHINode>(Entry->getMainOp()))
        Pred = Phi->getIncomingBlock(Op->UserTreeIndex.EdgeIdx);
      BasicBlock *OpParent;
      Instruction *OpLastInst;
      if (Op->isGather()) {
        assert(Entry->getOpcode() == Instruction::PHI &&
               "Expected phi node only.");
        OpParent = cast<PHINode>(Entry->getMainOp())
                       ->getIncomingBlock(Op->UserTreeIndex.EdgeIdx);
        OpLastInst = OpParent->getTerminator();
        for (Value *V : Op->Scalars) {
          auto *Inst = dyn_cast<Instruction>(V);
          if (!Inst)
            continue;
          if (isVectorized(V)) {
            OpParent = Inst->getParent();
            OpLastInst = Inst;
            break;
          }
        }
      } else {
        OpLastInst = EntriesToLastInstruction.at(Op);
        OpParent = OpLastInst->getParent();
      }
      // Check the call instructions within the same basic blocks.
      if (OpParent == Parent) {
        if (Entry->getOpcode() == Instruction::PHI) {
          if (!CheckForNonVecCallsInSameBlock(LastInst, OpLastInst))
            AddCosts(Op, GetSpillScale(Parent));
          continue;
        }
        if (!CheckForNonVecCallsInSameBlock(OpLastInst, LastInst))
          AddCosts(Op, GetSpillScale(Parent));
        continue;
      }
      // Check for call instruction in between blocks.
      // 1. Check entry's block to the head.
      if (Entry->getOpcode() != Instruction::PHI &&
          !CheckForNonVecCallsInSameBlock(
              &*Parent->getFirstNonPHIOrDbgOrAlloca(), LastInst)) {
        AddCosts(Op, GetSpillScale(Parent));
        continue;
      }
      // 2. Check op's block from the end.
      if (!CheckForNonVecCallsInSameBlock(OpLastInst,
                                          OpParent->getTerminator())) {
        AddCosts(Op, GetSpillScale(OpParent));
        continue;
      }
      // 3. Check the predecessors of entry's block till op's block.
      if (uint64_t Scale = CheckPredecessors(Parent, Pred, OpParent))
        AddCosts(Op, Scale);
    }
  }
 
  return Cost;
}
 
/// Checks if the \p IE1 instructions is followed by \p IE2 instruction in the
/// buildvector sequence.
static bool isFirstInsertElement(const InsertElementInst *IE1,
                                 const InsertElementInst *IE2) {
  if (IE1 == IE2)
    return false;
  const auto *I1 = IE1;
  const auto *I2 = IE2;
  const InsertElementInst *PrevI1;
  const InsertElementInst *PrevI2;
  unsigned Idx1 = *getElementIndex(IE1);
  unsigned Idx2 = *getElementIndex(IE2);
  do {
    if (I2 == IE1)
      return true;
    if (I1 == IE2)
      return false;
    PrevI1 = I1;
    PrevI2 = I2;
    if (I1 && (I1 == IE1 || I1->hasOneUse()) &&
        getElementIndex(I1).value_or(Idx2) != Idx2)
      I1 = dyn_cast<InsertElementInst>(I1->getOperand(0));
    if (I2 && ((I2 == IE2 || I2->hasOneUse())) &&
        getElementIndex(I2).value_or(Idx1) != Idx1)
      I2 = dyn_cast<InsertElementInst>(I2->getOperand(0));
  } while ((I1 && PrevI1 != I1) || (I2 && PrevI2 != I2));
  llvm_unreachable("Two different buildvectors not expected.");
}
 
namespace {
/// Returns incoming Value *, if the requested type is Value * too, or a default
/// value, otherwise.
struct ValueSelect {
  template <typename U>
  static std::enable_if_t<std::is_same_v<Value *, U>, Value *> get(Value *V) {
    return V;
  }
  template <typename U>
  static std::enable_if_t<!std::is_same_v<Value *, U>, U> get(Value *) {
    return U();
  }
};
} // namespace
 
/// Does the analysis of the provided shuffle masks and performs the requested
/// actions on the vectors with the given shuffle masks. It tries to do it in
/// several steps.
/// 1. If the Base vector is not undef vector, resizing the very first mask to
/// have common VF and perform action for 2 input vectors (including non-undef
/// Base). Other shuffle masks are combined with the resulting after the 1 stage
/// and processed as a shuffle of 2 elements.
/// 2. If the Base is undef vector and have only 1 shuffle mask, perform the
/// action only for 1 vector with the given mask, if it is not the identity
/// mask.
/// 3. If > 2 masks are used, perform the remaining shuffle actions for 2
/// vectors, combing the masks properly between the steps.
template <typename T>
static T *performExtractsShuffleAction(
    MutableArrayRef<std::pair<T *, SmallVector<int>>> ShuffleMask, Value *Base,
    function_ref<unsigned(T *)> GetVF,
    function_ref<std::pair<T *, bool>(T *, ArrayRef<int>, bool)> ResizeAction,
    function_ref<T *(ArrayRef<int>, ArrayRef<T *>)> Action) {
  assert(!ShuffleMask.empty() && "Empty list of shuffles for inserts.");
  SmallVector<int> Mask(ShuffleMask.begin()->second);
  auto VMIt = std::next(ShuffleMask.begin());
  T *Prev = nullptr;
  SmallBitVector UseMask =
      buildUseMask(Mask.size(), Mask, UseMask::UndefsAsMask);
  SmallBitVector IsBaseUndef = isUndefVector(Base, UseMask);
  if (!IsBaseUndef.all()) {
    // Base is not undef, need to combine it with the next subvectors.
    std::pair<T *, bool> Res =
        ResizeAction(ShuffleMask.begin()->first, Mask, /*ForSingleMask=*/false);
    SmallBitVector IsBasePoison = isUndefVector<true>(Base, UseMask);
    for (unsigned Idx = 0, VF = Mask.size(); Idx < VF; ++Idx) {
      if (Mask[Idx] == PoisonMaskElem)
        Mask[Idx] = IsBasePoison.test(Idx) ? PoisonMaskElem : Idx;
      else
        Mask[Idx] = (Res.second ? Idx : Mask[Idx]) + VF;
    }
    [[maybe_unused]] auto *V = ValueSelect::get<T *>(Base);
    assert((!V || GetVF(V) == Mask.size()) &&
           "Expected base vector of VF number of elements.");
    Prev = Action(Mask, {nullptr, Res.first});
  } else if (ShuffleMask.size() == 1) {
    // Base is undef and only 1 vector is shuffled - perform the action only for
    // single vector, if the mask is not the identity mask.
    std::pair<T *, bool> Res = ResizeAction(ShuffleMask.begin()->first, Mask,
                                            /*ForSingleMask=*/true);
    if (Res.second)
      // Identity mask is found.
      Prev = Res.first;
    else
      Prev = Action(Mask, {ShuffleMask.begin()->first});
  } else {
    // Base is undef and at least 2 input vectors shuffled - perform 2 vectors
    // shuffles step by step, combining shuffle between the steps.
    unsigned Vec1VF = GetVF(ShuffleMask.begin()->first);
    unsigned Vec2VF = GetVF(VMIt->first);
    if (Vec1VF == Vec2VF) {
      // No need to resize the input vectors since they are of the same size, we
      // can shuffle them directly.
      ArrayRef<int> SecMask = VMIt->second;
      for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) {
        if (SecMask[I] != PoisonMaskElem) {
          assert(Mask[I] == PoisonMaskElem && "Multiple uses of scalars.");
          Mask[I] = SecMask[I] + Vec1VF;
        }
      }
      Prev = Action(Mask, {ShuffleMask.begin()->first, VMIt->first});
    } else {
      // Vectors of different sizes - resize and reshuffle.
      std::pair<T *, bool> Res1 = ResizeAction(ShuffleMask.begin()->first, Mask,
                                               /*ForSingleMask=*/false);
      std::pair<T *, bool> Res2 =
          ResizeAction(VMIt->first, VMIt->second, /*ForSingleMask=*/false);
      ArrayRef<int> SecMask = VMIt->second;
      for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) {
        if (Mask[I] != PoisonMaskElem) {
          assert(SecMask[I] == PoisonMaskElem && "Multiple uses of scalars.");
          if (Res1.second)
            Mask[I] = I;
        } else if (SecMask[I] != PoisonMaskElem) {
          assert(Mask[I] == PoisonMaskElem && "Multiple uses of scalars.");
          Mask[I] = (Res2.second ? I : SecMask[I]) + VF;
        }
      }
      Prev = Action(Mask, {Res1.first, Res2.first});
    }
    VMIt = std::next(VMIt);
  }
  [[maybe_unused]] bool IsBaseNotUndef = !IsBaseUndef.all();
  // Perform requested actions for the remaining masks/vectors.
  for (auto E = ShuffleMask.end(); VMIt != E; ++VMIt) {
    // Shuffle other input vectors, if any.
    std::pair<T *, bool> Res =
        ResizeAction(VMIt->first, VMIt->second, /*ForSingleMask=*/false);
    ArrayRef<int> SecMask = VMIt->second;
    for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) {
      if (SecMask[I] != PoisonMaskElem) {
        assert((Mask[I] == PoisonMaskElem || IsBaseNotUndef) &&
               "Multiple uses of scalars.");
        Mask[I] = (Res.second ? I : SecMask[I]) + VF;
      } else if (Mask[I] != PoisonMaskElem) {
        Mask[I] = I;
      }
    }
    Prev = Action(Mask, {Prev, Res.first});
  }
  return Prev;
}
 
InstructionCost
BoUpSLP::calculateTreeCostAndTrimNonProfitable(ArrayRef<Value *> VectorizedVals,
                                               Instruction *RdxRoot) {
  // FIXME: support buildvector of the gather nodes with struct types.
  if (any_of(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
        return TE->isGather() &&
               isa<StructType>(TE->Scalars.front()->getType());
      })) {
    LLVM_DEBUG(
        dbgs() << "SLP: rejecting tree with buildvector struct values of size "
               << VectorizableTree.size() << ".\n");
    return InstructionCost::getInvalid();
  }
 
  // The tree is priced, so its minimal vector nodes are analyzed, including the
  // ones, trimmed below as non-profitable. The verdict of the tree itself does
  // not matter: the retry with the single-user seeds rebuilds the same narrow
  // nodes, and their cost without the tree is not better. The wider nodes are
  // not registered: their rejection does not prevent the members from forming
  // the profitable narrower node. Same for the trees, rejected before the cost
  // estimation, and for the buildvector trees, rejected by the insert/extract
  // overhead rather than by the members themselves.
  const TreeEntry &Root = getRootNode();
  if (!Root.hasState() ||
      !isa<InsertElementInst, InsertValueInst>(Root.getMainOp())) {
    for (const std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
      if (TE->isGather() || TE->Scalars.size() > 2)
        continue;
      for (Value *V : TE->Scalars)
        if (isa_and_present<Instruction>(V) && !TE->isCopyableElement(V))
          AnalyzedScalars.insert(V);
    }
  }
 
  SmallDenseMap<const TreeEntry *, InstructionCost> NodesCosts;
  SmallPtrSet<Value *, 4> CheckedExtracts;
  SmallSetVector<TreeEntry *, 4> GatheredLoadsNodes;
  SmallDenseMap<const TreeEntry *, InstructionCost> ExtractCosts;
  LLVM_DEBUG(dbgs() << "SLP: Calculating cost for tree of size "
                    << VectorizableTree.size() << ".\n");
  // The V-only-dependent part of the predicate. Same V is commonly seen in
  // multiple TEs (shared scalars), so cache the result across calls.
  // DeletedNodes is read-only during this cost loop, so caching is safe.
  const size_t NumVectScalars = ScalarToTreeEntries.size() + 1;
  SmallDenseMap<Value *, bool> ExternalUseVCache;
  auto IsExternallyUsedV = [&](Value *V) {
    auto [It, Inserted] = ExternalUseVCache.try_emplace(V);
    if (!Inserted)
      return It->second;
    bool Res = false;
    if (V->hasOneUse() || V->getType()->isVoidTy()) {
      // Res stays false.
    } else if (V->hasNUsesOrMore(NumVectScalars)) {
      Res = true;
    } else if (auto *I = dyn_cast<Instruction>(V)) {
      Res = any_of(I->users(), [&](const User *U) {
        // store/insertelt v, [cast]U will likely be vectorized.
        if (match(U,
                  m_InsertElt(m_Value(), m_OneUse(m_CastOrSelf(m_Specific(I))),
                              m_ConstantInt())))
          return false;
        if (match(U, m_InsertElt(m_Value(), m_Specific(I), m_ConstantInt())))
          return false;
        if (match(U, m_Store(m_OneUse(m_CastOrSelf(m_Specific(I))), m_Value())))
          return false;
        if (match(U, m_Store(m_Specific(I), m_Value())))
          return false;
        ArrayRef<TreeEntry *> Entries = getTreeEntries(U);
        if (Entries.empty() && !MustGather.contains(U))
          return true;
        if (any_of(Entries,
                   [&](TreeEntry *TE) { return DeletedNodes.contains(TE); }))
          return true;
        return any_of(ValueToGatherNodes.lookup(U), [&](const TreeEntry *TE) {
          return DeletedNodes.contains(TE);
        });
      });
    }
    It->second = Res;
    return Res;
  };
  auto IsExternallyUsed = [&](const TreeEntry &TE, Value *V) {
    assert(TE.hasState() && !TE.isGather() &&
           TE.State != TreeEntry::SplitVectorize && "Expected vector node.");
    if (TE.hasCopyableElements() && TE.isCopyableElement(V))
      return false;
    return IsExternallyUsedV(V);
  };
  InstructionCost Cost = 0;
  SmallDenseMap<const TreeEntry *, uint64_t> EntryToScale;
  uint64_t PrevScale = 0;
  BasicBlock *PrevVecParent = nullptr;
  for (const std::unique_ptr<TreeEntry> &Ptr : VectorizableTree) {
    TreeEntry &TE = *Ptr;
    // No need to count the cost for combined entries, they are combined and
    // just skip their cost.
    if (TE.State == TreeEntry::CombinedVectorize) {
      LLVM_DEBUG(
          dbgs() << "SLP: Skipping cost for combined node that starts with "
                 << *TE.Scalars[0] << ".\n";
          TE.dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n");
      NodesCosts.try_emplace(&TE);
      continue;
    }
    if (TE.hasState() &&
        (TE.isGather() || TE.State == TreeEntry::SplitVectorize)) {
      if (const TreeEntry *E =
              getSameValuesTreeEntry(TE.getMainOp(), TE.Scalars);
          E && E->getVectorFactor() == TE.getVectorFactor()) {
        // Some gather nodes might be absolutely the same as some vectorizable
        // nodes after reordering, need to handle it.
        LLVM_DEBUG(dbgs() << "SLP: Adding cost 0 for bundle "
                          << shortBundleName(TE.Scalars, TE.Idx) << ".\n"
                          << "SLP: Current total cost = " << Cost << "\n");
        NodesCosts.try_emplace(&TE);
        continue;
      }
    }
 
    // Exclude cost of gather loads nodes which are not used. These nodes were
    // built as part of the final attempt to vectorize gathered loads.
    assert((!TE.isGather() || TE.Idx == 0 || TE.UserTreeIndex) &&
           "Expected gather nodes with users only.");
 
    InstructionCost C = getEntryCost(&TE, VectorizedVals, CheckedExtracts);
    uint64_t Scale = 0;
    bool CostIsFree = C == 0;
    // For gather/buildvector (and split-vectorize) entries, prefer the
    // per-lane refined scale that accounts for LICM-hoistable insertelements
    // when an operand is invariant in the current loop nest but defined in
    // an outer loop. This prevents over-costing cross-loop-nest buildvectors.
    if (!CostIsFree && !TE.isGather() && TE.hasState()) {
      if (PrevVecParent == TE.getMainOp()->getParent()) {
        Scale = PrevScale;
        C *= Scale;
        EntryToScale.try_emplace(&TE, Scale);
      }
    }
    if (!CostIsFree && !Scale) {
      Scale = getEntryEffectiveScale(TE, TE.Idx == 0 ? RdxRoot : nullptr);
      C *= Scale;
      EntryToScale.try_emplace(&TE, Scale);
      if (!TE.isGather() && TE.hasState()) {
        PrevVecParent = TE.getMainOp()->getParent();
        PrevScale = Scale;
      }
    }
    Cost += C;
    NodesCosts.try_emplace(&TE, C);
    LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C << " for bundle "
                      << shortBundleName(TE.Scalars, TE.Idx) << ".\n"
                      << "SLP: Current total cost = " << Cost << "\n");
    // Add gathered loads nodes to the set for later processing.
    if (TE.Idx > 0 && !TE.UserTreeIndex && TE.hasState() &&
        TE.getOpcode() == Instruction::Load)
      GatheredLoadsNodes.insert(&TE);
    if (!TE.isGather() && TE.State != TreeEntry::SplitVectorize &&
        !(TE.Idx == 0 && (TE.getOpcode() == Instruction::InsertElement ||
                          TE.getOpcode() == Instruction::InsertValue ||
                          TE.getOpcode() == Instruction::Store)) &&
        !isa<StructType>(getValueType(TE.Scalars.front(), SLPReVec))) {
      // Calculate costs of external uses.
      APInt DemandedElts = APInt::getZero(TE.getVectorFactor());
      for (Value *V : TE.Scalars) {
        if (IsExternallyUsed(TE, V))
          DemandedElts.setBit(TE.findLaneForValue(V));
      }
      if (!DemandedElts.isZero()) {
        Type *ScalarTy = TE.Scalars.front()->getType();
        auto It = MinBWs.find(&TE);
        if (It != MinBWs.end())
          ScalarTy = IntegerType::get(ScalarTy->getContext(), It->second.first);
        auto *VecTy = getWidenedType(ScalarTy, TE.getVectorFactor());
        InstructionCost ExtCost = getScalarizationOverhead(
            *TTI, SLPReVec, ScalarTy, cast<VectorType>(VecTy), DemandedElts,
            /*Insert=*/false,
            /*Extract=*/true, CostKind);
        if (ExtCost.isValid() && ExtCost != 0) {
          if (!Scale)
            Scale = getScaleToLoopIterations(TE);
          ExtCost *= Scale;
          EntryToScale.try_emplace(&TE, Scale);
        }
        ExtractCosts.try_emplace(&TE, ExtCost);
      }
    }
  }
  // Bail out if the cost threshold is negative and cost already below it.
  // The splat subtrees may still force extracts of their scalars on top of
  // the node cost and have to be trimmed, so do not bail out if there are
  // any.
  if (SLPCostThreshold.getNumOccurrences() > 0 && SLPCostThreshold < 0 &&
      Cost < -SLPCostThreshold && SplatGatheredScalarsRoots.empty())
    return Cost;
  // The narrow non-profitable tree in loop? Skip, may cause regressions.
  constexpr unsigned PartLimit = 2;
  // Measure the narrowness by the width the tree is vectorized at.
  const TreeEntry *RootTE = &getRootNode();
  unsigned Sz;
  if (auto It = MinBWs.find(RootTE); It != MinBWs.end()) {
    Sz = It->second.first;
  } else {
    Sz = std::max<unsigned>(
        getVectorElementSize(RootTE->Scalars.front()),
        DL->getTypeSizeInBits(
            getValueType(RootTE->Scalars.front(), SLPReVec)->getScalarType()));
  }
  const unsigned MinVF = getMinVF(Sz);
  if (Cost >= -SLPCostThreshold &&
      getRootNodeScalars().size() * PartLimit <= MinVF &&
      (!getRootNode().hasState() ||
       (getRootNode().getOpcode() != Instruction::Store &&
        LI->getLoopFor(getRootNode().getMainOp()->getParent()))))
    return Cost;
  // Store the cost + external uses estimation as the first element of the
  // tuple, just the cost as the second element of the tuple. Required to return
  // correct cost estimation for the tree, extracts are calculated separately.
  // Extracts, calculated here, are just quick estimations.
  SmallVector<
      std::tuple<InstructionCost, InstructionCost, SmallVector<unsigned>>>
      SubtreeCosts(VectorizableTree.size());
  auto UpdateParentNodes =
      [&](const TreeEntry *UserTE, const TreeEntry *TE,
          InstructionCost TotalCost, InstructionCost Cost,
          SmallDenseSet<std::pair<const TreeEntry *, const TreeEntry *>, 4>
              &VisitedUser,
          bool AddToList = true) {
        while (UserTE &&
               VisitedUser.insert(std::make_pair(TE, UserTE)).second) {
          std::get<0>(SubtreeCosts[UserTE->Idx]) += TotalCost;
          std::get<1>(SubtreeCosts[UserTE->Idx]) += Cost;
          if (AddToList)
            std::get<2>(SubtreeCosts[UserTE->Idx]).push_back(TE->Idx);
          UserTE = UserTE->UserTreeIndex.UserTE;
        }
      };
  for (const std::unique_ptr<TreeEntry> &Ptr : VectorizableTree) {
    TreeEntry &TE = *Ptr;
    // Combined subnodes are not costed on their own (their cost is 0), but
    // must be included into the ancestors' subtree node lists, so that they
    // get deleted together with the trimmed combined root.
    InstructionCost C = NodesCosts.at(&TE);
    InstructionCost ExtractCost = ExtractCosts.lookup(&TE);
    std::get<0>(SubtreeCosts[TE.Idx]) += C + ExtractCost;
    std::get<1>(SubtreeCosts[TE.Idx]) += C;
    if (const TreeEntry *UserTE = TE.UserTreeIndex.UserTE) {
      SmallDenseSet<std::pair<const TreeEntry *, const TreeEntry *>, 4>
          VisitedUser;
      UpdateParentNodes(UserTE, &TE, C + ExtractCost, C, VisitedUser);
    }
  }
  SmallDenseSet<std::pair<const TreeEntry *, const TreeEntry *>, 4> Visited;
  for (TreeEntry *TE : GatheredLoadsNodes) {
    InstructionCost TotalCost = std::get<0>(SubtreeCosts[TE->Idx]);
    InstructionCost Cost = std::get<1>(SubtreeCosts[TE->Idx]);
    for (Value *V : TE->Scalars) {
      for (const TreeEntry *BVTE : ValueToGatherNodes.lookup(V))
        UpdateParentNodes(BVTE, TE, TotalCost, Cost, Visited,
                          /*AddToList=*/false);
    }
  }
  Visited.clear();
  using CostIndicesTy =
      std::pair<TreeEntry *, std::tuple<InstructionCost, InstructionCost,
                                        SmallVector<unsigned>>>;
  struct FirstGreater {
    bool operator()(const CostIndicesTy &LHS, const CostIndicesTy &RHS) const {
      return std::get<0>(LHS.second) < std::get<0>(RHS.second) ||
             (std::get<0>(LHS.second) == std::get<0>(RHS.second) &&
              LHS.first->Idx < RHS.first->Idx);
    }
  };
  PriorityQueue<CostIndicesTy, SmallVector<CostIndicesTy>, FirstGreater>
      Worklist;
  for (const auto [Idx, P] : enumerate(SubtreeCosts)) {
    // Combined subnodes are not trimmed on their own, only as a whole combined
    // node, so only the root nodes are checked and included into the worklist.
    if (VectorizableTree[Idx]->State == TreeEntry::CombinedVectorize)
      continue;
    Worklist.emplace(VectorizableTree[Idx].get(), P);
  }
 
  // Narrow store trees with non-profitable immediate values - exit.
  if (!UserIgnoreList && getRootNode().getVectorFactor() < MinVF &&
      getRootNode().hasState() &&
      getRootNode().getOpcode() == Instruction::Store &&
      (Worklist.top().first->Idx == 0 || Worklist.top().first->Idx == 1))
    return Cost;
 
  bool Changed = false;
  bool PreferTrimmedTree = false;
  while (!Worklist.empty() && std::get<0>(Worklist.top().second) > 0) {
    TreeEntry *TE = Worklist.top().first;
    if (TE->isGather() || TE->Idx == 0 || DeletedNodes.contains(TE) ||
        isa<StructType>(getValueType(TE->Scalars.front(), SLPReVec)) ||
        // Exit early if the parent node is split node and any of scalars is
        // used in other split nodes.
        (TE->UserTreeIndex &&
         TE->UserTreeIndex.UserTE->State == TreeEntry::SplitVectorize &&
         any_of(TE->Scalars, [&](Value *V) {
           ArrayRef<TreeEntry *> Entries = getSplitTreeEntries(V);
           return Entries.size() > 1;
         }))) {
      Worklist.pop();
      continue;
    }
    // Skip inversed compare nodes, they cannot be transformed to buildvectors.
    if (TE->State == TreeEntry::Vectorize && !TE->isAltShuffle() &&
        (TE->getOpcode() == Instruction::ICmp ||
         TE->getOpcode() == Instruction::FCmp) &&
        any_of(TE->Scalars, [&](Value *V) {
          auto *I = dyn_cast<CmpInst>(V);
          if (!I)
            return false;
          return I->getPredicate() !=
                 cast<CmpInst>(TE->getMainOp())->getPredicate();
        })) {
      Worklist.pop();
      continue;
    }
 
    // Calculate the gather cost of the root node.
    InstructionCost TotalSubtreeCost = std::get<0>(Worklist.top().second);
    InstructionCost SubtreeCost = std::get<1>(Worklist.top().second);
    if (TotalSubtreeCost < TE->Scalars.size()) {
      Worklist.pop();
      continue;
    }
    if (!TransformedToGatherNodes.empty()) {
      for (unsigned Idx : std::get<2>(Worklist.top().second)) {
        auto It = TransformedToGatherNodes.find(VectorizableTree[Idx].get());
        if (It != TransformedToGatherNodes.end()) {
          TotalSubtreeCost -= std::get<0>(SubtreeCosts[Idx]);
          SubtreeCost -= std::get<1>(SubtreeCosts[Idx]);
          TotalSubtreeCost += It->second;
          SubtreeCost += It->second;
        }
      }
    }
    if (TotalSubtreeCost < 0 || TotalSubtreeCost < TE->Scalars.size()) {
      Worklist.pop();
      continue;
    }
    const unsigned EntryVF = TE->getVectorFactor();
    APInt DemandedElts = APInt::getZero(EntryVF);
    for (auto [Idx, V] : enumerate(TE->Scalars)) {
      if (!isConstant(V))
        DemandedElts.setBit(Idx);
    }
 
    Type *ScalarTy = getValueType(TE->Scalars.front(), SLPReVec);
    auto It = MinBWs.find(TE);
    if (It != MinBWs.end())
      ScalarTy = IntegerType::get(ScalarTy->getContext(), It->second.first);
    auto *VecTy = getWidenedType(ScalarTy, EntryVF);
    InstructionCost GatherCost = getScalarizationOverhead(
        *TTI, SLPReVec, ScalarTy, cast<VectorType>(VecTy), DemandedElts,
        /*Insert=*/true, /*Extract=*/false, CostKind);
    SmallVector<int> Mask;
    if (!TE->ReorderIndices.empty() &&
        TE->State != TreeEntry::CompressVectorize &&
        (TE->State != TreeEntry::StridedVectorize ||
         !isReverseOrder(TE->ReorderIndices))) {
      SmallVector<int> NewMask;
      if (TE->getOpcode() == Instruction::Store) {
        // For stores the order is actually a mask.
        NewMask.resize(TE->ReorderIndices.size());
        copy(TE->ReorderIndices, NewMask.begin());
      } else {
        inversePermutation(TE->ReorderIndices, NewMask);
      }
      addMask(Mask, NewMask);
    }
    if (!TE->ReuseShuffleIndices.empty())
      addMask(Mask, TE->ReuseShuffleIndices);
    if (!Mask.empty() && !ShuffleVectorInst::isIdentityMask(Mask, EntryVF))
      GatherCost += getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc,
                                   cast<VectorType>(VecTy), CostKind, Mask);
    // If all scalars are reused in gather node(s) or other vector nodes, there
    // might be extra cost for inserting them.
    if ((!TE->hasState() || !TE->isAltShuffle()) &&
        all_of(TE->Scalars, [&](Value *V) {
          return (TE->hasCopyableElements() && TE->isCopyableElement(V)) ||
                 isConstant(V) || isGathered(V) || getTreeEntries(V).size() > 1;
        }))
      GatherCost *= 2;
    // Erase subtree if it is non-profitable.
    ArrayRef<unsigned> Nodes = std::get<2>(Worklist.top().second);
    // Prefer trimming equal-cost alternate-shuffle subtrees rooted at binary
    // ops: alt-shuffles introduce runtime shuffle overhead that the cost model
    // may underestimate. Skip if the subtree contains ExtractElement nodes,
    // since those operate on already-materialized vectors where the cost model
    // is more accurate.
    auto IsEqualCostAltShuffleToTrim = [&]() {
      return TotalSubtreeCost == GatherCost && TE->hasState() &&
             TE->isAltShuffle() && Instruction::isBinaryOp(TE->getOpcode()) &&
             none_of(Nodes, [&](unsigned Idx) {
               return VectorizableTree[Idx]->hasState() &&
                      VectorizableTree[Idx]->getOpcode() ==
                          Instruction::ExtractElement;
             });
    };
    // Non-power-of-2 entries may have inflated costs - add a margin of 1
    // before trimming to avoid over-pruning.
    bool HasNonPowerOf2 = any_of(Nodes, [&](unsigned Idx) {
      return !has_single_bit(VectorizableTree[Idx]->Scalars.size());
    });
    InstructionCost TrimMargin = HasNonPowerOf2 ? 1 : 0;
    if (TotalSubtreeCost > GatherCost + TrimMargin ||
        IsEqualCostAltShuffleToTrim()) {
      PreferTrimmedTree |= TotalSubtreeCost == GatherCost;
      // If the remaining tree is just a buildvector - exit, it will cause
      // endless attempts to vectorize. When the tree is already profitable,
      // skip trimming this node and let the post-loop logic (including
      // gathered loads processing) decide.
      if (getRootNode().hasState() &&
          getRootNode().getOpcode() == Instruction::InsertElement &&
          TE->Idx == 1) {
        if (Cost < -SLPCostThreshold) {
          LLVM_DEBUG(dbgs() << "SLP: Skipping trim of node " << TE->Idx
                            << " - tree already profitable with cost " << Cost
                            << ".\n");
          Worklist.pop();
          continue;
        }
        return InstructionCost::getInvalid();
      }
 
      LLVM_DEBUG(dbgs() << "SLP: Trimming unprofitable subtree at node "
                        << TE->Idx << " with cost "
                        << std::get<0>(Worklist.top().second)
                        << " and gather cost " << GatherCost << ".\n");
      if (TE->UserTreeIndex) {
        TransformedToGatherNodes.try_emplace(TE, GatherCost);
        NodesCosts.erase(TE);
      } else {
        DeletedNodes.insert(TE);
        TransformedToGatherNodes.erase(TE);
        NodesCosts.erase(TE);
      }
      for (unsigned Idx : Nodes) {
        TreeEntry &ChildTE = *VectorizableTree[Idx];
        DeletedNodes.insert(&ChildTE);
        TransformedToGatherNodes.erase(&ChildTE);
        NodesCosts.erase(&ChildTE);
      }
      Changed = true;
    }
    Worklist.pop();
  }
  if (!Changed) {
    // The splat subtrees are not linked to the tree root, so their cost is
    // not included in the root's subtree cost; add it explicitly.
    InstructionCost TotalCost = std::get<1>(SubtreeCosts.front());
    for (const TreeEntry *TE : SplatGatheredScalarsRoots)
      TotalCost += std::get<1>(SubtreeCosts[TE->Idx]);
    return TotalCost;
  }
 
  SmallPtrSet<TreeEntry *, 4> SubtreesToDelete;
  SmallPtrSet<TreeEntry *, 4> DroppedSplatSubtrees;
  InstructionCost LoadsExtractsCost = 0;
  using ValuesToInsertTy =
      SmallDenseMap<const TreeEntry *, SmallVector<Value *>>;
  auto GetScalarTy = [&](const TreeEntry *TE) {
    Type *ScalarTy = TE->Scalars.front()->getType();
    auto It = MinBWs.find(TE);
    if (It != MinBWs.end())
      ScalarTy = IntegerType::get(ScalarTy->getContext(), It->second.first);
    return ScalarTy;
  };
  // Lanes of the subtree scalars used by the surviving gather nodes, and the
  // values to materialize in those gathers if the subtree is deleted.
  auto FindDemandedElts = [&](TreeEntry *TE, ValuesToInsertTy &ValuesToInsert) {
    APInt DemandedElts = APInt::getZero(TE->getVectorFactor());
    for (Value *V : TE->Scalars) {
      unsigned Pos = TE->findLaneForValue(V);
      for (const TreeEntry *BVE : ValueToGatherNodes.lookup(V)) {
        if (DeletedNodes.contains(BVE))
          continue;
        DemandedElts.setBit(Pos);
        ValuesToInsert.try_emplace(BVE).first->second.push_back(V);
      }
    }
    return DemandedElts;
  };
  // Cost of materializing the values directly in the surviving gather nodes
  // that use them.
  auto GetGatherInsertCost = [&](Type *ScalarTy,
                                 const ValuesToInsertTy &ValuesToInsert) {
    InstructionCost BVCost = 0;
    for (const auto &[BVE, Values] : ValuesToInsert) {
      APInt BVDemandedElts = APInt::getZero(BVE->getVectorFactor());
      SmallVector<Value *> BVValues(BVE->getVectorFactor(),
                                    PoisonValue::get(ScalarTy));
      for (Value *V : Values) {
        unsigned Pos = BVE->findLaneForValue(V);
        BVValues[Pos] = V;
        BVDemandedElts.setBit(Pos);
      }
      BVCost += getScalarizationOverhead(
          *TTI, SLPReVec, ScalarTy,
          cast<VectorType>(getWidenedType(ScalarTy, BVE->getVectorFactor())),
          BVDemandedElts, /*Insert=*/true, /*Extract=*/false, CostKind,
          BVDemandedElts.isAllOnes(), BVValues);
    }
    return BVCost;
  };
  // Check if all loads of gathered loads nodes are marked for deletion. In this
  // case the whole gathered loads subtree must be deleted.
  // Also, try to account for extracts, which might be required, if only part of
  // gathered load must be vectorized. Keep partially vectorized nodes, if
  // extracts are cheaper than gathers.
  for (TreeEntry *TE : GatheredLoadsNodes) {
    if (DeletedNodes.contains(TE) || TransformedToGatherNodes.contains(TE))
      continue;
    SubtreesToDelete.insert(TE);
    // All loads are removed from gathered? Need to delete the subtree.
    ValuesToInsertTy ValuesToInsert;
    APInt DemandedElts = FindDemandedElts(TE, ValuesToInsert);
    if (!DemandedElts.isZero()) {
      Type *ScalarTy = GetScalarTy(TE);
      auto *VecTy = getWidenedType(ScalarTy, TE->getVectorFactor());
      InstructionCost ExtractsCost = getScalarizationOverhead(
          *TTI, SLPReVec, ScalarTy, cast<VectorType>(VecTy), DemandedElts,
          /*Insert=*/false, /*Extract=*/true, CostKind);
      InstructionCost BVCost = GetGatherInsertCost(ScalarTy, ValuesToInsert);
      if (ExtractsCost < BVCost) {
        LoadsExtractsCost += ExtractsCost;
        SubtreesToDelete.erase(TE);
        continue;
      }
      LoadsExtractsCost += BVCost;
    }
    NodesCosts.erase(TE);
  }
 
  // Check if all gather nodes that reuse the splat subtrees are marked for
  // deletion. In this case the whole splat subtree must be deleted. If only
  // some of the gathers are trimmed, keeping the subtree still costs its full
  // price plus the extracts of the scalars used by the remaining scalar code,
  // while the surviving gathers can materialize the splatted scalars
  // directly. Drop the subtree if it does not pay off.
  for (TreeEntry *TE : SplatGatheredScalarsRoots) {
    if (DeletedNodes.contains(TE))
      continue;
    ValuesToInsertTy ValuesToInsert;
    APInt DemandedElts = FindDemandedElts(TE, ValuesToInsert);
    if (!DemandedElts.isZero()) {
      Type *ScalarTy = GetScalarTy(TE);
      // Lanes of the subtree scalars still used by the remaining scalar code
      // must be extracted if the subtree is kept.
      APInt ExtractElts = APInt::getZero(TE->getVectorFactor());
      for (Value *V : TE->Scalars) {
        if (!isa<Instruction>(V) || TE->isCopyableElement(V))
          continue;
        // Too many users - the scalar is extracted anyway.
        if (V->hasNUsesOrMore(UsesLimit) || any_of(V->users(), [&](User *U) {
              return none_of(getTreeEntries(U), [&](const TreeEntry *UseTE) {
                return !DeletedNodes.contains(UseTE) &&
                       !TransformedToGatherNodes.contains(UseTE);
              });
            }))
          ExtractElts.setBit(TE->findLaneForValue(V));
      }
      InstructionCost KeepCost = getScalarizationOverhead(
          *TTI, SLPReVec, ScalarTy,
          cast<VectorType>(getWidenedType(ScalarTy, TE->getVectorFactor())),
          ExtractElts, /*Insert=*/false, /*Extract=*/true, CostKind);
      // Add the cost of the subtree itself, computed before any trimming:
      // trimming of the subtree's own nodes would otherwise make it look
      // artificially cheap.
      KeepCost += std::get<1>(SubtreeCosts[TE->Idx]);
      InstructionCost DropCost = GetGatherInsertCost(ScalarTy, ValuesToInsert);
      if (KeepCost <= DropCost)
        continue;
      // Dropped as unprofitable: exclude its cost from the reference cost, so
      // the trimming of the remaining tree is not reverted because of it, and
      // keep it deleted even if the trimming is reverted.
      DroppedSplatSubtrees.insert(TE);
      for (unsigned Idx : std::get<2>(SubtreeCosts[TE->Idx]))
        DroppedSplatSubtrees.insert(VectorizableTree[Idx].get());
      Cost -= std::get<1>(SubtreeCosts[TE->Idx]);
    }
    // Not used by the surviving gathers or not profitable to keep.
    SubtreesToDelete.insert(TE);
    NodesCosts.erase(TE);
  }
 
  // Deleted all subtrees rooted at gathered loads nodes or splat subtrees.
  for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
    if (TE->UserTreeIndex &&
        SubtreesToDelete.contains(TE->UserTreeIndex.UserTE)) {
      DeletedNodes.insert(TE.get());
      NodesCosts.erase(TE.get());
      SubtreesToDelete.insert(TE.get());
    }
    if (SubtreesToDelete.contains(TE.get()))
      DeletedNodes.insert(TE.get());
  }
 
  for (std::unique_ptr<TreeEntry> &TE : VectorizableTree) {
    if (!TE->UserTreeIndex && TransformedToGatherNodes.contains(TE.get())) {
      assert(TE->getOpcode() == Instruction::Load && "Expected load only.");
      continue;
    }
    if (DeletedNodes.contains(TE.get()))
      continue;
    // Gather costs depend on the set of vectorized nodes available for
    // reuse, which changes during trimming, so recalculate them for all
    // gather nodes, not just for the transformed ones.
    if (TE->isGather() || !NodesCosts.contains(TE.get())) {
      InstructionCost C =
          getEntryCost(TE.get(), VectorizedVals, CheckedExtracts);
      if (!C.isValid() || C == 0) {
        NodesCosts[TE.get()] = C;
        continue;
      }
      uint64_t Scale = EntryToScale.lookup(TE.get());
      if (!Scale)
        Scale = getEntryEffectiveScale(*TE);
      C *= Scale;
      NodesCosts[TE.get()] = C;
    }
  }
 
  LLVM_DEBUG(dbgs() << "SLP: Recalculate costs after tree trimming.\n");
  InstructionCost NewCost = 0;
  for (const auto &P : NodesCosts) {
    NewCost += P.second;
    LLVM_DEBUG(dbgs() << "SLP: Adding cost " << P.second << " for bundle "
                      << shortBundleName(P.first->Scalars, P.first->Idx)
                      << ".\n"
                      << "SLP: Current total cost = " << NewCost << "\n");
  }
  if (NewCost + LoadsExtractsCost > Cost ||
      (!PreferTrimmedTree && NewCost + LoadsExtractsCost == Cost)) {
    DeletedNodes.clear();
    TransformedToGatherNodes.clear();
    // The dropped splat subtrees stay deleted: they were excluded from the
    // reference cost and must not be resurrected by the revert.
    DeletedNodes.insert(DroppedSplatSubtrees.begin(),
                        DroppedSplatSubtrees.end());
    NewCost = Cost;
  } else {
    // If the remaining tree is just a buildvector - exit, it will cause
    // endless attempts to vectorize.
    if (VectorizableTree.size() >= 2 && getRootNode().hasState() &&
        getRootNode().getOpcode() == Instruction::InsertElement &&
        TransformedToGatherNodes.contains(VectorizableTree[1].get()))
      return InstructionCost::getInvalid();
    if (VectorizableTree.size() >= 3 && getRootNode().hasState() &&
        getRootNode().getOpcode() == Instruction::InsertElement &&
        VectorizableTree[1]->hasState() &&
        VectorizableTree[1]->State == TreeEntry::Vectorize &&
        (VectorizableTree[1]->getOpcode() == Instruction::ZExt ||
         VectorizableTree[1]->getOpcode() == Instruction::SExt ||
         VectorizableTree[1]->getOpcode() == Instruction::Trunc) &&
        TransformedToGatherNodes.contains(VectorizableTree[2].get()))
      return InstructionCost::getInvalid();
  }
  return NewCost;
}
 
namespace {
/// Data type for handling buildvector sequences with the reused scalars from
/// other tree entries.
template <typename T> struct ShuffledInsertData {
  /// List of insertelements to be replaced by shuffles.
  SmallVector<InsertElementInst *> InsertElements;
  /// The parent vectors and shuffle mask for the given list of inserts.
  MapVector<T, SmallVector<int>> ValueMasks;
};
} // namespace
 
InstructionCost BoUpSLP::getTreeCost(InstructionCost TreeCost,
                                     ArrayRef<Value *> VectorizedVals,
                                     InstructionCost ReductionCost,
                                     Instruction *RdxRoot) {
  // A poor-throughput entry's real vector-vs-scalar savings (fdiv/frem/fsqrt)
  // are already folded into TreeCost like any other entry, including all
  // shuffle/insert/extract overhead elsewhere in the tree. So bypassing the
  // instruction-count veto below just has to trust that already-complete
  // TreeCost instead of guessing at extra savings: heavy gather/shuffle
  // overhead raises TreeCost too, and will still block the bypass.
  auto BypassesInstCountCheck = [&]() {
    if (!VectorizePoorThroughput || TreeCost >= -SLPCostThreshold)
      return false;
    PoorThroughputOpCache Cache;
    return any_of(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &Ptr) {
      const TreeEntry &TE = *Ptr;
      return TE.hasState() && !DeletedNodes.contains(&TE) && !TE.isGather() &&
             !TransformedToGatherNodes.contains(&TE) &&
             TE.State != TreeEntry::CombinedVectorize &&
             isPoorThroughputOp(TE.getMainOp(), *TTI, *TLI, Cache, CostKind);
    });
  };
  // Reject vectorization if the vector code would produce more instructions
  // than the scalar code. The cost model may underestimate overhead from
  // shuffles, inserts, and extracts.
  // FIXME: remove this as soon as correct fractional model is landed for all
  // targets.
  if (CostKind != TTI::TCK_CodeSize && SLPInstCountCheck &&
      TTI->preferSLPInstCountCheck() && getRootNode().getVectorFactor() == 2 &&
      SLPCostThreshold == 0 &&
      (!SLPReVec ||
       !isa<VectorType>(getRootNodeScalars().front()->getType()))) {
    // Loop containing the tree root; null for flat code or disabled
    // loop-aware modeling. Shared by both calls below.
    const Loop *TreeLoop = nullptr;
    if (LoopAwareTripCount != 0 && getRootNode().hasState())
      TreeLoop = LI->getLoopFor(getRootNode().getMainOp()->getParent());
    uint64_t NumScalar = getNumScalarInsts(TreeLoop);
    uint64_t NumVector = getNumVectorInsts(TreeLoop);
    LLVM_DEBUG(dbgs() << "SLP: Inst count check: vector=" << NumVector
                      << " scalar=" << NumScalar << "\n");
    if (NumVector > NumScalar && !BypassesInstCountCheck()) {
      LLVM_DEBUG(dbgs() << "SLP: Rejecting tree: vector inst count "
                        << NumVector << " > scalar inst count " << NumScalar
                        << ".\n");
      return InstructionCost::getInvalid();
    }
  }
  InstructionCost Cost = TreeCost;
 
  SmallDenseMap<std::tuple<const TreeEntry *, Value *, Instruction *>, uint64_t>
      EntryToScale;
  auto ScaleCost = [&](InstructionCost C, const TreeEntry &TE,
                       Value *Scalar = nullptr, Instruction *U = nullptr) {
    if (!C.isValid() || C == 0)
      return C;
    uint64_t &Scale =
        EntryToScale.try_emplace(std::make_tuple(&TE, Scalar, U), 0)
            .first->getSecond();
    if (!Scale)
      Scale = getScaleToLoopIterations(TE, Scalar, U);
    LLVM_DEBUG(dbgs() << "Scale " << Scale << " For entry " << TE.Idx << "\n");
    return C * Scale;
  };
  Instruction *ReductionRoot = RdxRoot;
  if (UserIgnoreList) {
    // Scale reduction cost to the factor of the loop nest trip count.
    ReductionCost = ScaleCost(ReductionCost, getRootNode(), /*Scalar=*/nullptr,
                              ReductionRoot);
  }
 
  // Add the cost for reduction.
  Cost += ReductionCost;
 
  if (any_of(ExternalUses, [](const ExternalUser &EU) {
        return isa<StructType>(EU.Scalar->getType()) &&
               (EU.E.Idx != 0 || EU.E.State != TreeEntry::Vectorize ||
                EU.E.getOpcode() != Instruction::InsertValue) &&
               !isa_and_nonnull<ExtractValueInst>(EU.User);
      }))
    return InstructionCost::getInvalid();
 
  // Skip trees, which are non-profitable even if there are insertelements with
  // external uses.
  constexpr unsigned CostLimit = 100;
  if (Cost >= -SLPCostThreshold + CostLimit &&
      (VectorizableTree.size() - DeletedNodes.size()) *
              getRootNode().getVectorFactor() <
          CostLimit)
    return Cost;
 
  if (Cost >= -SLPCostThreshold &&
      none_of(ExternalUses, [](const ExternalUser &EU) {
        return isa_and_nonnull<InsertElementInst>(EU.User);
      }))
    return Cost;
 
  SmallDenseSet<std::pair<Value *, SmallVector<unsigned, 1>>, 16>
      ExtractCostCalculated;
  InstructionCost ExtractCost = 0;
  SmallVector<ShuffledInsertData<const TreeEntry *>> ShuffledInserts;
  SmallVector<APInt> DemandedElts;
  SmallDenseSet<Value *, 4> UsedInserts;
  DenseSet<std::pair<const TreeEntry *, Type *>> VectorCasts;
  std::optional<DenseMap<Value *, unsigned>> ValueToExtUses;
  DenseMap<const TreeEntry *, DenseSet<Value *>> ExtractsCount;
  SmallPtrSet<Value *, 4> ScalarOpsFromCasts;
  // Keep track {Scalar, Index, User} tuple.
  // On AArch64, this helps in fusing a mov instruction, associated with
  // extractelement, with fmul in the backend so that extractelement is free.
  SmallVector<std::tuple<Value *, User *, int>, 4> ScalarUserAndIdx;
  // Record every external use: a missing entry is indistinguishable from
  // lane 0 and is priced as a free extract by the extract-fusion cost model.
  for (ExternalUser &EU : ExternalUses)
    ScalarUserAndIdx.emplace_back(EU.Scalar, EU.User, EU.Lane);
  // Detect external uses that drive address computations: the scalar (through
  // an optional single-use index-promotion cast) is used as a GEP index.
  bool AllUsersGEPSWithStoresLoads = true;
  SmallVector<const Value *> Pointers;
  Type *UserScalarTy = nullptr;
  for (ExternalUser &EU : ExternalUses) {
    Value *Usr = EU.User;
    if (Usr && match(Usr, m_OneUse(m_ZExtOrSExt(m_Value()))))
      Usr = cast<Instruction>(Usr)->user_back();
    auto *User = dyn_cast_if_present<GetElementPtrInst>(Usr);
    // Only a GEP that feeds a single load/store of a fixed access type drives
    // a real memory address computation.
    Type *AccessTy = nullptr;
    if (User && User->hasOneUse() &&
        isa<LoadInst, StoreInst>(User->user_back()))
      AccessTy = getValueType(User->user_back(), SLPReVec);
    if (AccessTy && !isa<ScalableVectorType>(AccessTy) &&
        (!UserScalarTy || UserScalarTy == AccessTy)) {
      UserScalarTy = AccessTy;
      Pointers.push_back(User);
    } else {
      AllUsersGEPSWithStoresLoads = false;
      break;
    }
  }
 
  // Pre-pass: for each externally-used scalar, find the basic block at which
  // the extractelement will be placed by codegen. This mirrors what
  // vectorizeTree does: the extract is placed at the nearest common dominator
  // of all effective use sites. For a non-PHI user the effective site is the
  // user's own block; for a PHI user it is the incoming block for the scalar
  // operand (the predecessor of the PHI on the edge that carries the scalar).
  // Using the NCD of all effective sites rather than the first-encountered
  // user's block makes the extract-cost scale order-independent and correct
  // even when users live in different loop nests.
  SmallDenseMap<Value *, BasicBlock *> ScalarToExtractBlock;
  for (const ExternalUser &EU : ExternalUses) {
    if (!EU.User || isa<InsertElementInst>(EU.User))
      continue;
    if (EphValues.count(EU.User))
      continue;
    BasicBlock *UserParent = cast<Instruction>(EU.User)->getParent();
    if (!DT->isReachableFromEntry(UserParent) || UserParent->isEHPad() ||
        isa_and_present<UnreachableInst>(UserParent->getTerminator()))
      continue;
    BasicBlock *UseBlock = nullptr;
    if (auto *PHI = dyn_cast<PHINode>(EU.User)) {
      // When the PHI itself is inside a loop, the extractelement is placed
      // in the incoming block for the scalar operand (the predecessor edge),
      // not in the PHI's own block.  This applies to LCSSA phis at an inner-
      // loop exit that are still inside an outer loop: the incoming block is
      // in the inner loop while the PHI block is in the outer loop.
      // When the PHI is outside all loops (a true loop-exit phi), codegen
      // uses a vector phi at the exit block and the extract stays there
      // (scale = 1), so we keep the PHI's own block as the effective site.
      if (LI->getLoopFor(PHI->getParent())) {
        for (unsigned Idx : seq<unsigned>(PHI->getNumIncomingValues())) {
          if (PHI->getIncomingValue(Idx) != EU.Scalar)
            continue;
          BasicBlock *InBB = PHI->getIncomingBlock(Idx);
          UseBlock =
              UseBlock ? DT->findNearestCommonDominator(UseBlock, InBB) : InBB;
        }
      }
      if (!UseBlock)
        UseBlock = cast<Instruction>(EU.User)->getParent();
    } else {
      UseBlock = cast<Instruction>(EU.User)->getParent();
    }
    auto [It, Inserted] = ScalarToExtractBlock.try_emplace(EU.Scalar, UseBlock);
    if (!Inserted && It->second && UseBlock)
      It->second = DT->findNearestCommonDominator(It->second, UseBlock);
  }
 
  SmallDenseSet<std::pair<Value *, Value *>, 8> CheckedScalarUser;
  for (ExternalUser &EU : ExternalUses) {
    LLVM_DEBUG(dbgs() << "SLP: Computing cost for external use of TreeEntry "
                      << EU.E.Idx << " in lane " << EU.Lane << "\n");
    LLVM_DEBUG(if (EU.User) dbgs() << "  User:" << *EU.User << "\n";
               else dbgs() << "  User: nullptr\n");
    LLVM_DEBUG(dbgs() << "  Use: " << EU.Scalar->getNameOrAsOperand() << "\n");
 
    // Uses by ephemeral values are free (because the ephemeral value will be
    // removed prior to code generation, and so the extraction will be
    // removed as well).
    if (EphValues.count(EU.User))
      continue;
 
    // Check if the scalar for the given user or all users is accounted already.
    if (!CheckedScalarUser.insert(std::make_pair(EU.Scalar, EU.User)).second ||
        (EU.User &&
         CheckedScalarUser.contains(std::make_pair(EU.Scalar, nullptr))))
      continue;
 
    // Used in unreachable blocks or in EH pads (rarely executed) or is
    // terminated with unreachable instruction.
    if (BasicBlock *UserParent =
            EU.User ? cast<Instruction>(EU.User)->getParent() : nullptr;
        UserParent &&
        (!DT->isReachableFromEntry(UserParent) || UserParent->isEHPad() ||
         isa_and_present<UnreachableInst>(UserParent->getTerminator())))
      continue;
 
    // No extract cost for vector "scalar" if REVEC is disabled.
    if (isVectorizedTy(EU.Scalar->getType()) &&
        (!SLPReVec ||
         (EU.E.hasState() && EU.E.getOpcode() == Instruction::InsertElement)))
      continue;
 
    if (isa<InsertValueInst>(EU.Scalar))
      continue;
 
    // If found user is an insertelement, do not calculate extract cost but try
    // to detect it as a final shuffled/identity match.
    // TODO: what if a user is insertvalue when REVEC is enabled?
    if (auto *VU = dyn_cast_or_null<InsertElementInst>(EU.User);
        VU && VU->getOperand(1) == EU.Scalar) {
      if (auto *FTy = dyn_cast<FixedVectorType>(VU->getType())) {
        if (!UsedInserts.insert(VU).second)
          continue;
        std::optional<unsigned> InsertIdx = getElementIndex(VU);
        if (InsertIdx) {
          const TreeEntry *ScalarTE = &EU.E;
          auto *It = find_if(
              ShuffledInserts,
              [this, VU](const ShuffledInsertData<const TreeEntry *> &Data) {
                // Checks if 2 insertelements are from the same buildvector.
                InsertElementInst *VecInsert = Data.InsertElements.front();
                return areTwoInsertFromSameBuildVector(
                    VU, VecInsert, [this](InsertElementInst *II) -> Value * {
                      Value *Op0 = II->getOperand(0);
                      if (isVectorized(II) && !isVectorized(Op0))
                        return nullptr;
                      return Op0;
                    });
              });
          int VecId = -1;
          if (It == ShuffledInserts.end()) {
            auto &Data = ShuffledInserts.emplace_back();
            Data.InsertElements.emplace_back(VU);
            DemandedElts.push_back(APInt::getZero(FTy->getNumElements()));
            VecId = ShuffledInserts.size() - 1;
            auto It = MinBWs.find(ScalarTE);
            if (It != MinBWs.end() &&
                VectorCasts
                    .insert(std::make_pair(ScalarTE, FTy->getElementType()))
                    .second) {
              unsigned BWSz = It->second.first;
              unsigned DstBWSz = DL->getTypeSizeInBits(FTy->getElementType());
              unsigned VecOpcode;
              if (DstBWSz < BWSz)
                VecOpcode = Instruction::Trunc;
              else
                VecOpcode =
                    It->second.second ? Instruction::SExt : Instruction::ZExt;
              InstructionCost C = TTI->getCastInstrCost(
                  VecOpcode, FTy,
                  getWidenedType(IntegerType::get(FTy->getContext(), BWSz),
                                 FTy->getNumElements()),
                  TTI::CastContextHint::None, CostKind);
              LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C
                                << " for extending externally used vector with "
                                   "non-equal minimum bitwidth.\n");
              Cost += C;
            }
          } else {
            if (isFirstInsertElement(VU, It->InsertElements.front()))
              It->InsertElements.front() = VU;
            VecId = std::distance(ShuffledInserts.begin(), It);
          }
          int InIdx = *InsertIdx;
          SmallVectorImpl<int> &Mask =
              ShuffledInserts[VecId].ValueMasks[ScalarTE];
          if (Mask.empty())
            Mask.assign(FTy->getNumElements(), PoisonMaskElem);
          Mask[InIdx] = EU.Lane;
          DemandedElts[VecId].setBit(InIdx);
          continue;
        }
      }
    }
 
    // If we plan to rewrite the tree in a smaller type, we will need to sign
    // extend the extracted value back to the original type. Here, we account
    // for the extract and the added cost of the sign extend if needed.
    InstructionCost ExtraCost = TTI::TCC_Free;
    auto *ScalarTy = EU.Scalar->getType();
    const unsigned BundleWidth = EU.E.getVectorFactor();
    assert(EU.Lane < BundleWidth && "Extracted lane out of bounds.");
    auto *VecTy = getWidenedType(ScalarTy, BundleWidth);
    const TreeEntry *Entry = &EU.E;
    auto It = MinBWs.find(Entry);
    SmallVector<unsigned, 1> Indices;
    if (isa<StructType>(VecTy)) {
      assert(EU.User && "Expected user for struct extract");
      const auto *EV = cast<ExtractValueInst>(EU.User);
      Indices.assign(EV->getIndices());
    }
    // We only add extract cost once for the same scalar and struct field.
    auto ExtractKey = std::make_pair(EU.Scalar, Indices);
    if (!isa_and_nonnull<InsertElementInst>(EU.User) &&
        !ExtractCostCalculated.insert(ExtractKey).second)
      continue;
    if (It != MinBWs.end()) {
      Type *MinTy = IntegerType::get(F->getContext(), It->second.first);
      if (auto *VecTy = dyn_cast<FixedVectorType>(ScalarTy))
        MinTy = getWidenedType(MinTy, VecTy->getNumElements());
      unsigned Extend = isKnownNonNegative(EU.Scalar, SimplifyQuery(*DL))
                            ? Instruction::ZExt
                            : Instruction::SExt;
      VecTy = getWidenedType(MinTy, BundleWidth);
      ExtraCost =
          getExtractWithExtendCost(*TTI, SLPReVec, Extend, ScalarTy,
                                   cast<VectorType>(VecTy), EU.Lane, CostKind);
      LLVM_DEBUG(dbgs() << "  ExtractExtend or ExtractSubvec cost: "
                        << ExtraCost << "\n");
    } else {
      Type *ExtractTy = VecTy;
      if (auto *ST = dyn_cast<StructType>(VecTy)) {
        ExtractTy = ExtractValueInst::getIndexedType(ST, Indices);
      }
      ExtraCost = getVectorInstrCost(
          *TTI, SLPReVec, ScalarTy, Instruction::ExtractElement, ExtractTy,
          CostKind, EU.Lane, EU.Scalar, ScalarUserAndIdx);
      LLVM_DEBUG(dbgs() << "  ExtractElement cost for " << *ScalarTy << " from "
                        << *VecTy << ": " << ExtraCost << "\n");
    }
    // Leave the scalar instructions as is if they are cheaper than extracts.
    if (Entry->Idx != 0 || Entry->getOpcode() == Instruction::GetElementPtr ||
        Entry->getOpcode() == Instruction::Load) {
      // Checks if the user of the external scalar is phi in loop body.
      auto IsPhiInLoop = [&](const ExternalUser &U) {
        if (auto *Phi = dyn_cast_if_present<PHINode>(U.User)) {
          auto *I = cast<Instruction>(U.Scalar);
          const Loop *L = LI->getLoopFor(Phi->getParent());
          return L && (Phi->getParent() == I->getParent() ||
                       L == LI->getLoopFor(I->getParent()));
        }
        return false;
      };
      if (!ValueToExtUses) {
        ValueToExtUses.emplace();
        for (const auto &P : enumerate(ExternalUses)) {
          // Ignore phis in loops.
          if (IsPhiInLoop(P.value()))
            continue;
 
          ValueToExtUses->try_emplace(P.value().Scalar, P.index());
        }
      }
      // Can use original instruction, if no operands vectorized or they are
      // marked as externally used already.
      auto *Inst = cast<Instruction>(EU.Scalar);
      InstructionCost ScalarCost = TTI->getInstructionCost(Inst, CostKind);
      auto OperandIsScalar = [&](Value *V) {
        if (!isVectorized(V)) {
          // Some extractelements might be not vectorized, but
          // transformed into shuffle and removed from the function,
          // consider it here.
          if (auto *EE = dyn_cast<ExtractElementInst>(V))
            return !EE->hasOneUse() || !MustGather.contains(EE);
          return true;
        }
        if (isa<StructType>(V->getType()))
          return false;
        return ValueToExtUses->contains(V);
      };
      bool CanBeUsedAsScalar = all_of(Inst->operands(), OperandIsScalar);
      bool CanBeUsedAsScalarCast = false;
      if (auto *CI = dyn_cast<CastInst>(Inst); CI && !CanBeUsedAsScalar) {
        if (auto *Op = dyn_cast<Instruction>(CI->getOperand(0));
            Op && all_of(Op->operands(), OperandIsScalar)) {
          InstructionCost OpCost =
              (isVectorized(Op) && !ValueToExtUses->contains(Op))
                  ? TTI->getInstructionCost(Op, CostKind)
                  : 0;
          if (ScalarCost + OpCost <= ExtraCost) {
            CanBeUsedAsScalar = CanBeUsedAsScalarCast = true;
            ScalarCost += OpCost;
          }
        }
      }
      if (CanBeUsedAsScalar) {
        bool KeepScalar = ScalarCost <= ExtraCost;
        // Try to keep original scalar if the user is the phi node from the same
        // block as the root phis, currently vectorized. It allows to keep
        // better ordering info of PHIs, being vectorized currently.
        bool IsProfitablePHIUser =
            (KeepScalar || (ScalarCost - ExtraCost <= TTI::TCC_Basic &&
                            getRootNodeScalars().size() > 2)) &&
            getRootNode().hasState() &&
            getRootNode().getOpcode() == Instruction::PHI &&
            !Inst->hasNUsesOrMore(UsesLimit) &&
            none_of(Inst->users(),
                    [&](User *U) {
                      auto *PHIUser = dyn_cast<PHINode>(U);
                      return (!PHIUser ||
                              PHIUser->getParent() !=
                                  getRootNode().getMainOp()->getParent()) &&
                             !isVectorized(U);
                    }) &&
            count_if(Entry->Scalars, [&](Value *V) {
              return ValueToExtUses->contains(V);
            }) <= 2;
        if (IsProfitablePHIUser) {
          KeepScalar = true;
        } else if (KeepScalar && ScalarCost != TTI::TCC_Free &&
                   ExtraCost - ScalarCost <= TTI::TCC_Basic &&
                   (!GatheredLoadsEntriesFirst.has_value() ||
                    Entry->Idx < *GatheredLoadsEntriesFirst)) {
          unsigned ScalarUsesCount = count_if(Entry->Scalars, [&](Value *V) {
            return ValueToExtUses->contains(V);
          });
          auto It = ExtractsCount.find(Entry);
          if (It != ExtractsCount.end()) {
            assert(ScalarUsesCount >= It->getSecond().size() &&
                   "Expected total number of external uses not less than "
                   "number of scalar uses.");
            ScalarUsesCount -= It->getSecond().size();
          }
          // Keep original scalar if number of externally used instructions in
          // the same entry is not power of 2. It may help to do some extra
          // vectorization for now.
          KeepScalar = ScalarUsesCount <= 1 || !has_single_bit(ScalarUsesCount);
        }
        if (KeepScalar) {
          ExternalUsesAsOriginalScalar.insert(EU.Scalar);
          for (Value *V : Inst->operands()) {
            // Struct operands cannot be rebuilt by the !User extraction
            // path (it has no insertvalue chain), so leave their existing
            // ExtractValueInst user in place.
            if (isa<StructType>(V->getType()))
              continue;
            auto It = ValueToExtUses->find(V);
            if (It != ValueToExtUses->end()) {
              // Replace all uses to avoid compiler crash.
              ExternalUses[It->second].User = nullptr;
            }
          }
          ExtraCost = ScalarCost;
          if (!IsPhiInLoop(EU))
            ExtractsCount[Entry].insert(Inst);
          if (CanBeUsedAsScalarCast) {
            ScalarOpsFromCasts.insert(Inst->getOperand(0));
            // Update the users of the operands of the cast operand to avoid
            // compiler crash.
            if (auto *IOp = dyn_cast<Instruction>(Inst->getOperand(0))) {
              for (Value *V : IOp->operands()) {
                if (isa<StructType>(V->getType()))
                  continue;
                auto It = ValueToExtUses->find(V);
                if (It != ValueToExtUses->end()) {
                  // Replace all uses to avoid compiler crash.
                  ExternalUses[It->second].User = nullptr;
                }
              }
            }
          }
        }
      }
    }
 
    // Scale the extract cost by the execution frequency of the block where
    // codegen will place the extractelement. That block is the nearest common
    // dominator of all effective use sites (precomputed in ScalarToExtractBlock
    // above), which is order-independent. For scalars kept as originals the
    // existing ScaleCost path (user-block based) remains correct, since the
    // scalar instruction executes at its definition site's frequency.
    if (!ExternalUsesAsOriginalScalar.contains(EU.Scalar)) {
      if (ExtraCost.isValid() && ExtraCost != 0) {
        if (!EU.User) {
          // No external user instruction is recorded (User == nullptr): the
          // scalar stays live in vectorized instructions or is used as an
          // extra arg, and is not present in ScalarToExtractBlock (the
          // pre-pass only records sites of real users). vectorizeTree() then
          // places the extractelement right after the vectorized instruction
          // (in the entry's block) and replaces the scalar uses with it, so
          // scale by the entry block's execution frequency to match that
          // placement.
          ExtraCost = ScaleCost(ExtraCost, *Entry, EU.Scalar, /*U=*/nullptr);
        } else {
          BasicBlock *ExtractBB = ScalarToExtractBlock.lookup(EU.Scalar);
          if (const Loop *L = ExtractBB ? LI->getLoopFor(ExtractBB) : nullptr) {
            uint64_t Scale = getLoopNestScale(
                findInnermostNonInvariantLoop(L, ArrayRef<Value *>(EU.Scalar)));
            LLVM_DEBUG(dbgs()
                       << "SLP: Extract scale " << Scale << " (NCD block) for "
                       << EU.Scalar->getNameOrAsOperand() << "\n");
            ExtraCost *= Scale;
          }
        }
      }
    } else {
      ExtraCost = ScaleCost(ExtraCost, *Entry, EU.Scalar,
                            cast_or_null<Instruction>(EU.User));
    }
 
    ExtractCost += ExtraCost;
  }
  // Charge the pointer-chain cost difference once when every escaped scalar
  // is used only to drive an address computation (see the detection loop
  // above). Vectorizing the tree in this pattern forces lane extracts (or a
  // vector GEP with unknown stride) to drive the address computation, which is
  // typically more expensive than keeping the indices scalar in a unit-stride
  // address chain. Add the delta once rather than per external use.
  if (AllUsersGEPSWithStoresLoads && !Pointers.empty()) {
    const TreeEntry &RootEntry = getRootNode();
    const Value *CommonBase = getUnderlyingObject(Pointers.front());
    if (all_of(Pointers, [CommonBase](const Value *P) {
          return getUnderlyingObject(P) == CommonBase;
        })) {
      auto *VecTy = getWidenedType(UserScalarTy, RootEntry.Scalars.size());
      InstructionCost ScalarGEPCost = TTI->getPointersChainCost(
          Pointers, CommonBase, TTI::PointersChainInfo::getUnitStride(),
          UserScalarTy, CostKind);
      InstructionCost VectorGEPCost = TTI->getPointersChainCost(
          Pointers, CommonBase, TTI::PointersChainInfo::getUnknownStride(),
          VecTy, CostKind);
      ExtractCost += ScaleCost(VectorGEPCost - ScalarGEPCost, RootEntry);
    }
  }
  // Insert externals for extract of operands of casts to be emitted as scalars
  // instead of extractelement.
  for (Value *V : ScalarOpsFromCasts) {
    ExternalUsesAsOriginalScalar.insert(V);
    if (ArrayRef<TreeEntry *> TEs = getTreeEntries(V); !TEs.empty()) {
      const auto *It = find_if_not(TEs, [&](TreeEntry *TE) {
        return TransformedToGatherNodes.contains(TE) ||
               DeletedNodes.contains(TE);
      });
      if (It != TEs.end()) {
        const TreeEntry *UserTE = *It;
        ExternalUses.emplace_back(V, nullptr, *UserTE,
                                  UserTE->findLaneForValue(V));
      }
    }
  }
  // Add reduced value cost, if resized.
  if (!VectorizedVals.empty()) {
    const TreeEntry &Root = getRootNode();
    auto BWIt = MinBWs.find(&Root);
    if (BWIt != MinBWs.end()) {
      Type *DstTy = Root.Scalars.front()->getType();
      unsigned OriginalSz = DL->getTypeSizeInBits(DstTy->getScalarType());
      unsigned SrcSz =
          ReductionBitWidth == 0 ? BWIt->second.first : ReductionBitWidth;
      if (OriginalSz != SrcSz) {
        unsigned Opcode = Instruction::Trunc;
        if (OriginalSz > SrcSz)
          Opcode = BWIt->second.second ? Instruction::SExt : Instruction::ZExt;
        Type *SrcTy = IntegerType::get(DstTy->getContext(), SrcSz);
        if (auto *VecTy = dyn_cast<FixedVectorType>(DstTy)) {
          assert(SLPReVec && "Only supported by REVEC.");
          SrcTy = getWidenedType(SrcTy, VecTy->getNumElements());
        }
        InstructionCost CastCost = TTI->getCastInstrCost(
            Opcode, DstTy, SrcTy, TTI::CastContextHint::None, CostKind);
        CastCost = ScaleCost(CastCost, Root, /*Scalar=*/nullptr, ReductionRoot);
        Cost += CastCost;
      }
    }
  }
 
  // Buildvector with externally used scalars, which should remain as scalars,
  // should not be vectorized, the compiler may hang.
  if (SLPCostThreshold < 0 && VectorizableTree.size() > 1 &&
      isa<InsertElementInst>(VectorizableTree[0]->Scalars[0]) &&
      VectorizableTree[1]->hasState() &&
      VectorizableTree[1]->State == TreeEntry::Vectorize &&
      all_of(VectorizableTree[1]->Scalars, [&](Value *V) {
        return ExternalUsesAsOriginalScalar.contains(V);
      }))
    return InstructionCost::getInvalid();
 
  Cost += ExtractCost;
  auto &&ResizeToVF = [this, &Cost](const TreeEntry *TE, ArrayRef<int> Mask,
                                    bool ForSingleMask) {
    InstructionCost C = 0;
    unsigned VF = Mask.size();
    unsigned VecVF = TE->getVectorFactor();
    bool HasLargeIndex =
        any_of(Mask, [VF](int Idx) { return Idx >= static_cast<int>(VF); });
    if ((VF != VecVF && HasLargeIndex) ||
        !ShuffleVectorInst::isIdentityMask(Mask, VF)) {
 
      if (HasLargeIndex) {
        SmallVector<int> OrigMask(VecVF, PoisonMaskElem);
        std::copy(Mask.begin(), std::next(Mask.begin(), std::min(VF, VecVF)),
                  OrigMask.begin());
        C = getShuffleCost(
            *TTI, TTI::SK_PermuteSingleSrc,
            cast<VectorType>(getWidenedType(TE->getMainOp()->getType(), VecVF)),
            CostKind, OrigMask);
        LLVM_DEBUG(
            dbgs() << "SLP: Adding cost " << C
                   << " for final shuffle of insertelement external users.\n";
            TE->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n");
        Cost += C;
        return std::make_pair(TE, true);
      }
 
      if (!ForSingleMask) {
        SmallVector<int> ResizeMask(VF, PoisonMaskElem);
        for (unsigned I = 0; I < VF; ++I) {
          if (Mask[I] != PoisonMaskElem)
            ResizeMask[Mask[I]] = Mask[I];
        }
        if (!ShuffleVectorInst::isIdentityMask(ResizeMask, VF))
          C = getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc,
                             cast<VectorType>(getWidenedType(
                                 TE->getMainOp()->getType(), VecVF)),
                             CostKind, ResizeMask);
        LLVM_DEBUG(
            dbgs() << "SLP: Adding cost " << C
                   << " for final shuffle of insertelement external users.\n";
            TE->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n");
 
        Cost += C;
      }
    }
    return std::make_pair(TE, false);
  };
  // Calculate the cost of the reshuffled vectors, if any.
  for (int I = 0, E = ShuffledInserts.size(); I < E; ++I) {
    Value *Base = ShuffledInserts[I].InsertElements.front()->getOperand(0);
    auto Vector = ShuffledInserts[I].ValueMasks.takeVector();
    unsigned VF = 0;
    auto EstimateShufflesCost = [&](ArrayRef<int> Mask,
                                    ArrayRef<const TreeEntry *> TEs) {
      assert((TEs.size() == 1 || TEs.size() == 2) &&
             "Expected exactly 1 or 2 tree entries.");
      if (TEs.size() == 1) {
        if (VF == 0)
          VF = TEs.front()->getVectorFactor();
        auto *FTy = getWidenedType(TEs.back()->Scalars.front()->getType(), VF);
        if (!ShuffleVectorInst::isIdentityMask(Mask, VF) &&
            !all_of(enumerate(Mask), [=](const auto &Data) {
              return Data.value() == PoisonMaskElem ||
                     (Data.index() < VF &&
                      static_cast<int>(Data.index()) == Data.value());
            })) {
          InstructionCost C =
              getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc,
                             cast<VectorType>(FTy), CostKind, Mask);
          C = ScaleCost(C, *TEs.front());
          LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C
                            << " for final shuffle of insertelement "
                               "external users.\n";
                     TEs.front()->dump();
                     dbgs() << "SLP: Current total cost = " << Cost << "\n");
          Cost += C;
        }
      } else {
        if (VF == 0) {
          if (TEs.front() &&
              TEs.front()->getVectorFactor() == TEs.back()->getVectorFactor())
            VF = TEs.front()->getVectorFactor();
          else
            VF = Mask.size();
        }
        auto *FTy = getWidenedType(TEs.back()->Scalars.front()->getType(), VF);
        InstructionCost C = getShuffleCost(
            *TTI, TTI::SK_PermuteTwoSrc, cast<VectorType>(FTy), CostKind, Mask);
        C = ScaleCost(C, *TEs.back());
        LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C
                          << " for final shuffle of vector node and external "
                             "insertelement users.\n";
                   if (TEs.front()) { TEs.front()->dump(); } TEs.back()->dump();
                   dbgs() << "SLP: Current total cost = " << Cost << "\n");
        Cost += C;
      }
      VF = Mask.size();
      return TEs.back();
    };
    (void)performExtractsShuffleAction<const TreeEntry>(
        MutableArrayRef(Vector.data(), Vector.size()), Base,
        [](const TreeEntry *E) { return E->getVectorFactor(); }, ResizeToVF,
        EstimateShufflesCost);
    InstructionCost InsertCost = TTI->getScalarizationOverhead(
        cast<FixedVectorType>(
            ShuffledInserts[I].InsertElements.front()->getType()),
        DemandedElts[I],
        /*Insert*/ true, /*Extract*/ false, CostKind);
    Cost -= InsertCost;
  }
 
  // Add the cost for reduced value resize (if required).
  if (ReductionBitWidth != 0) {
    assert(UserIgnoreList && "Expected reduction tree.");
    const TreeEntry &E = getRootNode();
    auto It = MinBWs.find(&E);
    if (It != MinBWs.end() && It->second.first != ReductionBitWidth) {
      unsigned SrcSize = It->second.first;
      unsigned DstSize = ReductionBitWidth;
      unsigned Opcode = Instruction::Trunc;
      if (SrcSize < DstSize) {
        bool IsArithmeticExtendedReduction =
            all_of(*UserIgnoreList, [](Value *V) {
              auto *I = cast<Instruction>(V);
              return is_contained({Instruction::Add, Instruction::FAdd,
                                   Instruction::Mul, Instruction::FMul,
                                   Instruction::And, Instruction::Or,
                                   Instruction::Xor},
                                  I->getOpcode());
            });
        if (IsArithmeticExtendedReduction)
          Opcode =
              Instruction::BitCast; // Handle it by getExtendedReductionCost
        else
          Opcode = It->second.second ? Instruction::SExt : Instruction::ZExt;
      }
      if (Opcode != Instruction::BitCast) {
        auto *SrcVecTy =
            getWidenedType(Builder.getIntNTy(SrcSize), E.getVectorFactor());
        auto *DstVecTy =
            getWidenedType(Builder.getIntNTy(DstSize), E.getVectorFactor());
        TTI::CastContextHint CCH = getCastContextHint(E);
        switch (E.getOpcode()) {
        case Instruction::SExt:
        case Instruction::ZExt:
        case Instruction::Trunc: {
          const TreeEntry *OpTE = getOperandEntry(&E, 0);
          CCH = getCastContextHint(*OpTE);
          break;
        }
        default:
          break;
        }
        InstructionCost CastCost =
            TTI->getCastInstrCost(Opcode, DstVecTy, SrcVecTy, CCH, CostKind);
        CastCost = ScaleCost(CastCost, getRootNode(), /*Scalar=*/nullptr,
                             ReductionRoot);
        Cost += CastCost;
        LLVM_DEBUG(dbgs() << "SLP: Adding cost " << CastCost
                          << " for final resize for reduction from " << SrcVecTy
                          << " to " << DstVecTy << "\n";
                   dbgs() << "SLP: Current total cost = " << Cost << "\n");
      }
    }
  }
 
  std::optional<InstructionCost> SpillCost;
  if (Cost < -SLPCostThreshold) {
    SpillCost = getSpillCost();
    Cost += *SpillCost;
  }
#ifndef NDEBUG
  SmallString<256> Str;
  {
    raw_svector_ostream OS(Str);
    OS << "SLP: Spill Cost = ";
    if (SpillCost)
      OS << *SpillCost;
    else
      OS << "<skipped>";
    OS << ".\nSLP: Extract Cost = " << ExtractCost << ".\n";
    if (ReductionRoot)
      OS << "SLP:  Reduction Cost = " << ReductionCost << ".\n";
    OS << "SLP: Total Cost = " << Cost << ".\n";
  }
  LLVM_DEBUG(dbgs() << Str);
  if (ViewSLPTree)
    ViewGraph(this, "SLP" + F->getName(), false, Str);
#endif
 
  return Cost;
}
 
/// Tries to find extractelement instructions with constant indices from fixed
/// vector type and gather such instructions into a bunch, which highly likely
/// might be detected as a shuffle of 1 or 2 input vectors. If this attempt was
/// successful, the matched scalars are replaced by poison values in \p VL for
/// future analysis.
std::optional<TTI::ShuffleKind>
BoUpSLP::tryToGatherSingleRegisterExtractElements(
    MutableArrayRef<Value *> VL, SmallVectorImpl<int> &Mask) const {
  // Scan list of gathered scalars for extractelements that can be represented
  // as shuffles.
  MapVector<Value *, SmallVector<int>> VectorOpToIdx;
  SmallVector<int> UndefVectorExtracts;
  for (int I = 0, E = VL.size(); I < E; ++I) {
    auto *EI = dyn_cast<ExtractElementInst>(VL[I]);
    if (!EI) {
      if (isa<UndefValue>(VL[I]))
        UndefVectorExtracts.push_back(I);
      continue;
    }
    auto *VecTy = dyn_cast<FixedVectorType>(EI->getVectorOperandType());
    if (!VecTy || !isa<ConstantInt, UndefValue>(EI->getIndexOperand()))
      continue;
    std::optional<unsigned> Idx = getExtractIndex(EI);
    // Undefined index.
    if (!Idx) {
      UndefVectorExtracts.push_back(I);
      continue;
    }
    if (Idx >= VecTy->getNumElements()) {
      UndefVectorExtracts.push_back(I);
      continue;
    }
    SmallBitVector ExtractMask(VecTy->getNumElements(), true);
    ExtractMask.reset(*Idx);
    if (isUndefVector</*IsPoisonOnly=*/true>(EI->getVectorOperand(),
                                             ExtractMask)
            .all()) {
      UndefVectorExtracts.push_back(I);
      continue;
    }
    VectorOpToIdx[EI->getVectorOperand()].push_back(I);
  }
  // Sort the vector operands by the maximum number of uses in extractelements.
  SmallVector<std::pair<Value *, SmallVector<int>>> Vectors =
      VectorOpToIdx.takeVector();
  stable_sort(Vectors, [](const auto &P1, const auto &P2) {
    return P1.second.size() > P2.second.size();
  });
  // Find the best pair of the vectors or a single vector.
  const int UndefSz = UndefVectorExtracts.size();
  unsigned SingleMax = 0;
  unsigned PairMax = 0;
  if (!Vectors.empty()) {
    SingleMax = Vectors.front().second.size() + UndefSz;
    if (Vectors.size() > 1) {
      auto *ItNext = std::next(Vectors.begin());
      PairMax = SingleMax + ItNext->second.size();
    }
  }
  if (SingleMax == 0 && PairMax == 0 && UndefSz == 0)
    return std::nullopt;
  // Check if better to perform a shuffle of 2 vectors or just of a single
  // vector.
  SmallVector<Value *> SavedVL(VL.begin(), VL.end());
  SmallVector<Value *> GatheredExtracts(
      VL.size(), PoisonValue::get(VL.front()->getType()));
  if (SingleMax >= PairMax && SingleMax) {
    for (int Idx : Vectors.front().second)
      std::swap(GatheredExtracts[Idx], VL[Idx]);
  } else if (!Vectors.empty()) {
    for (unsigned Idx : {0, 1})
      for (int Idx : Vectors[Idx].second)
        std::swap(GatheredExtracts[Idx], VL[Idx]);
  }
  // Add extracts from undefs too.
  for (int Idx : UndefVectorExtracts)
    std::swap(GatheredExtracts[Idx], VL[Idx]);
  // Check that gather of extractelements can be represented as just a
  // shuffle of a single/two vectors the scalars are extracted from.
  std::optional<TTI::ShuffleKind> Res =
      isFixedVectorShuffle(GatheredExtracts, Mask, AC);
  if (!Res || all_of(Mask, equal_to(PoisonMaskElem))) {
    // TODO: try to check other subsets if possible.
    // Restore the original VL if attempt was not successful.
    copy(SavedVL, VL.begin());
    return std::nullopt;
  }
  // Restore unused scalars from mask, if some of the extractelements were not
  // selected for shuffle.
  for (int I = 0, E = GatheredExtracts.size(); I < E; ++I) {
    if (Mask[I] == PoisonMaskElem && !isa<PoisonValue>(GatheredExtracts[I]) &&
        isa<UndefValue>(GatheredExtracts[I])) {
      std::swap(VL[I], GatheredExtracts[I]);
      continue;
    }
    auto *EI = dyn_cast<ExtractElementInst>(VL[I]);
    if (!EI || !isa<FixedVectorType>(EI->getVectorOperandType()) ||
        !isa<ConstantInt, UndefValue>(EI->getIndexOperand()) ||
        is_contained(UndefVectorExtracts, I))
      continue;
  }
  return Res;
}
 
/// Tries to find extractelement instructions with constant indices from fixed
/// vector type and gather such instructions into a bunch, which highly likely
/// might be detected as a shuffle of 1 or 2 input vectors. If this attempt was
/// successful, the matched scalars are replaced by poison values in \p VL for
/// future analysis.
SmallVector<std::optional<TTI::ShuffleKind>>
BoUpSLP::tryToGatherExtractElements(SmallVectorImpl<Value *> &VL,
                                    SmallVectorImpl<int> &Mask,
                                    unsigned NumParts) const {
  assert(NumParts > 0 && "NumParts expected be greater than or equal to 1.");
  SmallVector<std::optional<TTI::ShuffleKind>> ShufflesRes(NumParts);
  Mask.assign(VL.size(), PoisonMaskElem);
  unsigned SliceSize = getPartNumElems(VL.size(), NumParts);
  for (unsigned Part : seq<unsigned>(NumParts)) {
    // Scan list of gathered scalars for extractelements that can be represented
    // as shuffles.
    const unsigned PartOffset = Part * SliceSize;
    const unsigned PartSize = getNumElems(VL.size(), SliceSize, Part);
    // It may happen in case of revec, need to check no access out of bounds.
    if (PartOffset + PartSize > VL.size())
      break;
    MutableArrayRef<Value *> SubVL =
        MutableArrayRef(VL).slice(PartOffset, PartSize);
    SmallVector<int> SubMask;
    std::optional<TTI::ShuffleKind> Res =
        tryToGatherSingleRegisterExtractElements(SubVL, SubMask);
    ShufflesRes[Part] = Res;
    copy(SubMask, std::next(Mask.begin(), Part * SliceSize));
    if (SubVL.size() != SliceSize)
      break;
  }
  if (none_of(ShufflesRes, [](const std::optional<TTI::ShuffleKind> &Res) {
        return Res.has_value();
      }))
    ShufflesRes.clear();
  return ShufflesRes;
}
 
std::optional<TargetTransformInfo::ShuffleKind>
BoUpSLP::isGatherShuffledSingleRegisterEntry(
    const TreeEntry *TE, ArrayRef<Value *> VL, MutableArrayRef<int> Mask,
    SmallVectorImpl<const TreeEntry *> &Entries, unsigned Part, bool ForOrder,
    unsigned SliceSize) {
  Entries.clear();
  if (TE->Idx == 0)
    return std::nullopt;
  const unsigned MaskBase = Part * SliceSize;
  // TODO: currently checking only for Scalars in the tree entry, need to count
  // reused elements too for better cost estimation.
  auto GetUserEntry = [&](const TreeEntry *TE) {
    while (TE->UserTreeIndex && TE->UserTreeIndex.EdgeIdx == UINT_MAX)
      TE = TE->UserTreeIndex.UserTE;
    if (TE == &getRootNode())
      return EdgeInfo(const_cast<TreeEntry *>(TE), 0);
    return TE->UserTreeIndex;
  };
  auto HasGatherUser = [&](const TreeEntry *TE) {
    while (TE->Idx != 0 && TE->UserTreeIndex) {
      if (TE->UserTreeIndex.EdgeIdx == UINT_MAX)
        return true;
      TE = TE->UserTreeIndex.UserTE;
    }
    return false;
  };
  const EdgeInfo TEUseEI = GetUserEntry(TE);
  if (!TEUseEI || (TEUseEI.UserTE->Idx == 0 && TEUseEI.UserTE->isGather() &&
                   !TEUseEI.UserTE->hasState()))
    return std::nullopt;
  const Instruction *TEInsertPt = &getLastInstructionInBundle(TEUseEI.UserTE);
  const BasicBlock *TEInsertBlock = nullptr;
  // Main node of PHI entries keeps the correct order of operands/incoming
  // blocks.
  if (auto *PHI = dyn_cast_or_null<PHINode>(
          TEUseEI.UserTE->hasState() ? TEUseEI.UserTE->getMainOp() : nullptr);
      PHI && TEUseEI.UserTE->State != TreeEntry::SplitVectorize) {
    TEInsertBlock = PHI->getIncomingBlock(TEUseEI.EdgeIdx);
    TEInsertPt = TEInsertBlock->getTerminator();
  } else {
    TEInsertBlock = TEInsertPt->getParent();
  }
  if (!DT->isReachableFromEntry(TEInsertBlock))
    return std::nullopt;
  auto *NodeUI = DT->getNode(TEInsertBlock);
  assert(NodeUI && "Should only process reachable instructions");
  SmallPtrSet<Value *, 4> GatheredScalars(llvm::from_range, VL);
  auto CheckOrdering = [&](const Instruction *InsertPt) {
    // Argument InsertPt is an instruction where vector code for some other
    // tree entry (one that shares one or more scalars with TE) is going to be
    // generated. This lambda returns true if insertion point of vector code
    // for the TE dominates that point (otherwise dependency is the other way
    // around). The other node is not limited to be of a gather kind. Gather
    // nodes are not scheduled and their vector code is inserted before their
    // first user. If user is PHI, that is supposed to be at the end of a
    // predecessor block. Otherwise it is the last instruction among scalars of
    // the user node. So, instead of checking dependency between instructions
    // themselves, we check dependency between their insertion points for vector
    // code (since each scalar instruction ends up as a lane of a vector
    // instruction).
    const BasicBlock *InsertBlock = InsertPt->getParent();
    auto *NodeEUI = DT->getNode(InsertBlock);
    if (!NodeEUI)
      return false;
    assert((NodeUI == NodeEUI) ==
               (NodeUI->getDFSNumIn() == NodeEUI->getDFSNumIn()) &&
           "Different nodes should have different DFS numbers");
    // Check the order of the gather nodes users.
    if (TEInsertPt->getParent() != InsertBlock &&
        (DT->dominates(NodeUI, NodeEUI) || !DT->dominates(NodeEUI, NodeUI)))
      return false;
    if (TEInsertPt->getParent() == InsertBlock &&
        TEInsertPt->comesBefore(InsertPt))
      return false;
    return true;
  };
  // Find all tree entries used by the gathered values. If no common entries
  // found - not a shuffle.
  // Here we build a set of tree nodes for each gathered value and trying to
  // find the intersection between these sets. If we have at least one common
  // tree node for each gathered value - we have just a permutation of the
  // single vector. If we have 2 different sets, we're in situation where we
  // have a permutation of 2 input vectors.
  SmallVector<SmallPtrSet<const TreeEntry *, 4>> UsedTEs;
  SmallDenseMap<Value *, int> UsedValuesEntry;
  SmallPtrSet<const Value *, 16> VisitedValue;
  bool IsReusedNodeFound = false;
  auto CheckAndUseSameNode = [&](const TreeEntry *TEPtr) {
    // The node is reused - exit.
    if (IsReusedNodeFound)
      return false;
    if ((TEPtr->getVectorFactor() != VL.size() &&
         TEPtr->Scalars.size() != VL.size()) ||
        (!TEPtr->isSame(VL) && !TEPtr->isSame(TE->Scalars)))
      return false;
    IsReusedNodeFound =
        equal(TE->Scalars, TEPtr->Scalars) &&
        equal(TE->ReorderIndices, TEPtr->ReorderIndices) &&
        equal(TE->ReuseShuffleIndices, TEPtr->ReuseShuffleIndices);
    UsedTEs.clear();
    UsedTEs.emplace_back().insert(TEPtr);
    for (Value *V : VL) {
      if (isConstant(V))
        continue;
      UsedValuesEntry.try_emplace(V, 0);
    }
    return true;
  };
  auto CheckParentNodes = [&](const TreeEntry *User1, const TreeEntry *User2,
                              unsigned EdgeIdx) {
    const TreeEntry *Ptr1 = User1;
    const TreeEntry *Ptr2 = User2;
    SmallDenseMap<const TreeEntry *, unsigned> PtrToIdx;
    while (Ptr2) {
      PtrToIdx.try_emplace(Ptr2, EdgeIdx);
      EdgeIdx = Ptr2->UserTreeIndex.EdgeIdx;
      Ptr2 = Ptr2->UserTreeIndex.UserTE;
    }
    while (Ptr1) {
      unsigned Idx = Ptr1->UserTreeIndex.EdgeIdx;
      Ptr1 = Ptr1->UserTreeIndex.UserTE;
      if (auto It = PtrToIdx.find(Ptr1); It != PtrToIdx.end())
        return Idx < It->second;
    }
    return false;
  };
  // Cache `isUsedOutsideBlock(TEInsertPt)` - TEInsertPt is loop-invariant and
  // the function walks the instruction's user list.
  std::optional<bool> TEInsertPtUsedOutsideBlock;
  auto IsTEInsertPtUsedOutsideBlock = [&] {
    if (!TEInsertPtUsedOutsideBlock)
      TEInsertPtUsedOutsideBlock =
          isUsedOutsideBlock(const_cast<Instruction *>(TEInsertPt));
    return *TEInsertPtUsedOutsideBlock;
  };
  // Cache the TEUseEI/TEInsertPt-only prefix of the per-call lambda predicate
  // below - all of these depend only on outer-scope state, not the lambda's
  // arguments.
  const bool TEUseEIInsertPtUsedOutside =
      TEUseEI && TEUseEI.UserTE && TEUseEI.UserTE->hasCopyableElements() &&
      !TEUseEI.UserTE->isCopyableElement(
          const_cast<Instruction *>(TEInsertPt)) &&
      IsTEInsertPtUsedOutsideBlock();
  auto CheckNonSchedulableOrdering = [&](const TreeEntry *E,
                                         Instruction *InsertPt) {
    return TEUseEIInsertPtUsedOutside &&
           InsertPt->getNextNode() == TEInsertPt &&
           (!E->hasCopyableElements() || !E->isCopyableElement(InsertPt) ||
            !isUsedOutsideBlock(InsertPt));
  };
  // Cache the TEUseEI.UserTE-dependent predicate - it is invariant across the
  // double loop below. all_of with isUsedOutsideBlock walks each scalar's
  // users and is the expensive component.
  const bool TEUserNeedsEmitFirst =
      TEUseEI.UserTE->State == TreeEntry::Vectorize &&
      TEUseEI.UserTE->hasState() &&
      (TEUseEI.UserTE->getOpcode() != Instruction::PHI ||
       TEUseEI.UserTE->isAltShuffle()) &&
      all_of(TEUseEI.UserTE->Scalars, isUsedOutsideBlock);
  // Cache `all_of(UserTE->Scalars, isUsedOutsideBlock)` per UserTE - the
  // same UserTE may be encountered for many TEPtr values inside the loop.
  SmallDenseMap<const TreeEntry *, bool> ScalarsUsedOutsideBlockCache;
  auto AllScalarsUsedOutsideBlock = [&](const TreeEntry *UserTE) {
    auto [It, Inserted] = ScalarsUsedOutsideBlockCache.try_emplace(UserTE);
    if (!Inserted)
      return It->second;
    bool Res = all_of(UserTE->Scalars, isUsedOutsideBlock);
    It->second = Res;
    return Res;
  };
  for (Value *V : VL) {
    if (isConstant(V) || !VisitedValue.insert(V).second)
      continue;
    // Build a list of tree entries where V is used.
    SmallPtrSet<const TreeEntry *, 4> VToTEs;
    SmallVector<const TreeEntry *> GatherNodes(
        ValueToGatherNodes.lookup(V).takeVector());
    if (TransformedToGatherNodes.contains(TE)) {
      for (TreeEntry *E : getSplitTreeEntries(V)) {
        if (TE == E || !TransformedToGatherNodes.contains(E) ||
            !E->UserTreeIndex || E->UserTreeIndex.UserTE->isGather())
          continue;
        GatherNodes.push_back(E);
      }
      for (TreeEntry *E : getTreeEntries(V)) {
        if (TE == E || !TransformedToGatherNodes.contains(E) ||
            !E->UserTreeIndex || E->UserTreeIndex.UserTE->isGather())
          continue;
        GatherNodes.push_back(E);
      }
    }
    for (const TreeEntry *TEPtr : GatherNodes) {
      if (TEPtr == TE || TEPtr->Idx == 0 || DeletedNodes.contains(TEPtr))
        continue;
      assert(any_of(TEPtr->Scalars,
                    [&](Value *V) { return GatheredScalars.contains(V); }) &&
             "Must contain at least single gathered value.");
      assert(TEPtr->UserTreeIndex &&
             "Expected only single user of a gather node.");
      if (any_of(TEPtr->CombinedEntriesWithIndices,
                 [&](const auto &P) { return P.first == TE->Idx; }))
        continue;
      const EdgeInfo &UseEI = TEPtr->UserTreeIndex;
 
      PHINode *UserPHI = (UseEI.UserTE->State != TreeEntry::SplitVectorize &&
                          UseEI.UserTE->hasState())
                             ? dyn_cast<PHINode>(UseEI.UserTE->getMainOp())
                             : nullptr;
      Instruction *InsertPt =
          UserPHI ? UserPHI->getIncomingBlock(UseEI.EdgeIdx)->getTerminator()
                  : &getLastInstructionInBundle(UseEI.UserTE);
      if (TEInsertPt == InsertPt) {
        // Check nodes, which might be emitted first.
        if (TEUserNeedsEmitFirst) {
          if (UseEI.UserTE->State != TreeEntry::Vectorize ||
              (UseEI.UserTE->hasState() &&
               UseEI.UserTE->getOpcode() == Instruction::PHI &&
               !UseEI.UserTE->isAltShuffle()) ||
              !AllScalarsUsedOutsideBlock(UseEI.UserTE))
            continue;
        }
 
        // If the schedulable insertion point is used in multiple entries - just
        // exit, no known ordering at this point, available only after real
        // scheduling.
        if (!doesNotNeedToBeScheduled(InsertPt) &&
            (TEUseEI.UserTE != UseEI.UserTE || TEUseEI.EdgeIdx < UseEI.EdgeIdx))
          continue;
        // If the users are the PHI nodes with the same incoming blocks - skip.
        if (TEUseEI.UserTE->State == TreeEntry::Vectorize &&
            TEUseEI.UserTE->getOpcode() == Instruction::PHI &&
            UseEI.UserTE->State == TreeEntry::Vectorize &&
            UseEI.UserTE->getOpcode() == Instruction::PHI &&
            TEUseEI.UserTE != UseEI.UserTE)
          continue;
        // If 2 gathers are operands of the same entry (regardless of whether
        // user is PHI or else), compare operands indices, use the earlier one
        // as the base.
        if (TEUseEI.UserTE == UseEI.UserTE && TEUseEI.EdgeIdx < UseEI.EdgeIdx)
          continue;
        // If the user instruction is used for some reason in different
        // vectorized nodes - make it depend on index.
        if (TEUseEI.UserTE != UseEI.UserTE &&
            (TEUseEI.UserTE->Idx < UseEI.UserTE->Idx ||
             HasGatherUser(TEUseEI.UserTE)))
          continue;
        // If the user node is the operand of the other user node - skip.
        if (CheckParentNodes(TEUseEI.UserTE, UseEI.UserTE, UseEI.EdgeIdx))
          continue;
      }
 
      if (!TEUseEI.UserTE->isGather() && !UserPHI &&
          TEUseEI.UserTE->doesNotNeedToSchedule() !=
              UseEI.UserTE->doesNotNeedToSchedule() &&
          is_contained(UseEI.UserTE->Scalars, TEInsertPt))
        continue;
      // Check if the user node of the TE comes after user node of TEPtr,
      // otherwise TEPtr depends on TE.
      if ((TEInsertBlock != InsertPt->getParent() ||
           TEUseEI.EdgeIdx < UseEI.EdgeIdx || TEUseEI.UserTE != UseEI.UserTE) &&
          (!CheckOrdering(InsertPt) ||
           (UseEI.UserTE->hasCopyableElements() &&
            IsTEInsertPtUsedOutsideBlock() &&
            is_contained(UseEI.UserTE->Scalars, TEInsertPt))))
        continue;
      // The node is reused - exit.
      if (CheckAndUseSameNode(TEPtr))
        break;
      // The parent node is copyable with last inst used outside? And the last
      // inst is the next inst for the lastinst of TEPtr? Exit, if yes, to
      // preserve def-use chain.
      if (CheckNonSchedulableOrdering(UseEI.UserTE, InsertPt))
        continue;
      VToTEs.insert(TEPtr);
    }
    if (ArrayRef<TreeEntry *> VTEs = getSplitTreeEntries(V); !VTEs.empty()) {
      const auto *It = find_if(VTEs, [&](const TreeEntry *MTE) {
        return MTE != TE && MTE != TEUseEI.UserTE &&
               !DeletedNodes.contains(MTE) &&
               !TransformedToGatherNodes.contains(MTE);
      });
      if (It != VTEs.end()) {
        const TreeEntry *VTE = *It;
        if (none_of(TE->CombinedEntriesWithIndices,
                    [&](const auto &P) { return P.first == VTE->Idx; })) {
          Instruction &LastBundleInst = getLastInstructionInBundle(VTE);
          if (&LastBundleInst == TEInsertPt || !CheckOrdering(&LastBundleInst))
            continue;
        }
        // The node is reused - exit.
        if (CheckAndUseSameNode(VTE))
          break;
        VToTEs.insert(VTE);
      }
    }
    if (ArrayRef<TreeEntry *> VTEs = getTreeEntries(V); !VTEs.empty()) {
      const auto *It = find_if(VTEs, [&, MainTE = TE](const TreeEntry *TE) {
        return TE != MainTE && !DeletedNodes.contains(TE) &&
               !TransformedToGatherNodes.contains(TE);
      });
      if (It != VTEs.end()) {
        const TreeEntry *VTE = *It;
        if (ForOrder && VTE->Idx < GatheredLoadsEntriesFirst.value_or(0) &&
            VTEs.size() > 1 && VTE->State != TreeEntry::Vectorize) {
          VTEs = VTEs.drop_front();
          // Iterate through all vectorized nodes.
          const auto *MIt = find_if(VTEs, [](const TreeEntry *MTE) {
            return MTE->State == TreeEntry::Vectorize;
          });
          if (MIt == VTEs.end())
            continue;
          VTE = *MIt;
        }
        if (none_of(TE->CombinedEntriesWithIndices,
                    [&](const auto &P) { return P.first == VTE->Idx; })) {
          Instruction &LastBundleInst = getLastInstructionInBundle(VTE);
          if (&LastBundleInst == TEInsertPt ||
              !CheckOrdering(&LastBundleInst) ||
              CheckNonSchedulableOrdering(VTE, &LastBundleInst))
            continue;
        }
        // The node is reused - exit.
        if (CheckAndUseSameNode(VTE))
          break;
        VToTEs.insert(VTE);
      }
    }
    if (IsReusedNodeFound)
      break;
    if (VToTEs.empty())
      continue;
    if (UsedTEs.empty()) {
      // The first iteration, just insert the list of nodes to vector.
      UsedTEs.push_back(VToTEs);
      UsedValuesEntry.try_emplace(V, 0);
    } else {
      // Need to check if there are any previously used tree nodes which use V.
      // If there are no such nodes, consider that we have another one input
      // vector.
      SmallPtrSet<const TreeEntry *, 4> SavedVToTEs(VToTEs);
      unsigned Idx = 0;
      for (SmallPtrSet<const TreeEntry *, 4> &Set : UsedTEs) {
        // Do we have a non-empty intersection of previously listed tree entries
        // and tree entries using current V?
        set_intersect(VToTEs, Set);
        if (!VToTEs.empty()) {
          // Yes, write the new subset and continue analysis for the next
          // scalar.
          Set.swap(VToTEs);
          break;
        }
        VToTEs = SavedVToTEs;
        ++Idx;
      }
      // No non-empty intersection found - need to add a second set of possible
      // source vectors.
      if (Idx == UsedTEs.size()) {
        // If the number of input vectors is greater than 2 - not a permutation,
        // fallback to the regular gather.
        // TODO: support multiple reshuffled nodes.
        if (UsedTEs.size() == 2)
          continue;
        UsedTEs.push_back(SavedVToTEs);
        Idx = UsedTEs.size() - 1;
      }
      UsedValuesEntry.try_emplace(V, Idx);
    }
  }
 
  if (UsedTEs.empty()) {
    Entries.clear();
    return std::nullopt;
  }
 
  unsigned VF = 0;
  if (UsedTEs.size() == 1) {
    // Keep the order to avoid non-determinism.
    SmallVector<const TreeEntry *> FirstEntries(UsedTEs.front().begin(),
                                                UsedTEs.front().end());
    sort(FirstEntries, [](const TreeEntry *TE1, const TreeEntry *TE2) {
      return TE1->Idx < TE2->Idx;
    });
    // Try to find the perfect match in another gather node at first.
    auto *It = find_if(FirstEntries, [=](const TreeEntry *EntryPtr) {
      return (EntryPtr->getVectorFactor() == TE->Scalars.size() &&
              EntryPtr->isSame(TE->Scalars)) ||
             EntryPtr->isSame(VL);
    });
    if (It != FirstEntries.end() &&
        (IsReusedNodeFound || (*It)->getVectorFactor() == VL.size() ||
         ((*It)->getVectorFactor() == TE->Scalars.size() &&
          TE->ReuseShuffleIndices.size() == VL.size() &&
          (*It)->isSame(TE->Scalars)))) {
      Entries.push_back(*It);
      if (IsReusedNodeFound || (*It)->getVectorFactor() == VL.size()) {
        std::iota(std::next(Mask.begin(), MaskBase),
                  std::next(Mask.begin(), MaskBase + VL.size()), 0);
      } else {
        SmallVector<int> CommonMask = TE->getCommonMask();
        copy(CommonMask, Mask.begin());
      }
      // Clear undef scalars.
      for (unsigned I : seq<unsigned>(VL.size()))
        if (isa<PoisonValue>(VL[I]))
          Mask[MaskBase + I] = PoisonMaskElem;
      return TargetTransformInfo::SK_PermuteSingleSrc;
    }
    // No perfect match, just shuffle, so choose the first tree node from the
    // tree.
    Entries.push_back(FirstEntries.front());
    // Update mapping between values and corresponding tree entries.
    for (auto &P : UsedValuesEntry)
      P.second = 0;
    VF = FirstEntries.front()->getVectorFactor();
  } else {
    // Try to find nodes with the same vector factor.
    assert(UsedTEs.size() == 2 && "Expected at max 2 permuted entries.");
    // Keep the order of tree nodes to avoid non-determinism.
    DenseMap<int, const TreeEntry *> VFToTE;
    for (const TreeEntry *TE : UsedTEs.front()) {
      unsigned VF = TE->getVectorFactor();
      auto It = VFToTE.find(VF);
      if (It != VFToTE.end()) {
        if (It->second->Idx > TE->Idx)
          It->getSecond() = TE;
        continue;
      }
      VFToTE.try_emplace(VF, TE);
    }
    // Same, keep the order to avoid non-determinism.
    SmallVector<const TreeEntry *> SecondEntries(UsedTEs.back().begin(),
                                                 UsedTEs.back().end());
    sort(SecondEntries, [](const TreeEntry *TE1, const TreeEntry *TE2) {
      return TE1->Idx < TE2->Idx;
    });
    for (const TreeEntry *TE : SecondEntries) {
      auto It = VFToTE.find(TE->getVectorFactor());
      if (It != VFToTE.end()) {
        VF = It->first;
        Entries.push_back(It->second);
        Entries.push_back(TE);
        break;
      }
    }
    // No 2 source vectors with the same vector factor - just choose 2 with max
    // index.
    if (Entries.empty()) {
      Entries.push_back(*llvm::max_element(
          UsedTEs.front(), [](const TreeEntry *TE1, const TreeEntry *TE2) {
            return TE1->Idx < TE2->Idx;
          }));
      Entries.push_back(SecondEntries.front());
      VF = std::max(Entries.front()->getVectorFactor(),
                    Entries.back()->getVectorFactor());
    } else {
      VF = Entries.front()->getVectorFactor();
    }
    SmallVector<SmallPtrSet<Value *, 8>> ValuesToEntries;
    for (const TreeEntry *E : Entries)
      ValuesToEntries.emplace_back().insert(E->Scalars.begin(),
                                            E->Scalars.end());
    // Update mapping between values and corresponding tree entries.
    for (auto &P : UsedValuesEntry) {
      for (unsigned Idx : seq<unsigned>(ValuesToEntries.size()))
        if (ValuesToEntries[Idx].contains(P.first)) {
          P.second = Idx;
          break;
        }
    }
  }
 
  bool IsSplatOrUndefs = isSplat(VL) || all_of(VL, IsaPred<UndefValue>);
  // Checks if the 2 PHIs are compatible in terms of high possibility to be
  // vectorized.
  auto AreCompatiblePHIs = [&](Value *V, Value *V1) {
    auto *PHI = cast<PHINode>(V);
    auto *PHI1 = cast<PHINode>(V1);
    // Check that all incoming values are compatible/from same parent (if they
    // are instructions).
    // The incoming values are compatible if they all are constants, or
    // instruction with the same/alternate opcodes from the same basic block.
    for (int I = 0, E = PHI->getNumIncomingValues(); I < E; ++I) {
      Value *In = PHI->getIncomingValue(I);
      Value *In1 = PHI1->getIncomingValue(I);
      if (isConstant(In) && isConstant(In1))
        continue;
      if (!getSameOpcode({In, In1}, *TLI))
        return false;
      if (cast<Instruction>(In)->getParent() !=
          cast<Instruction>(In1)->getParent())
        return false;
    }
    return true;
  };
  // Check if the value can be ignored during analysis for shuffled gathers.
  // We suppose it is better to ignore instruction, which do not form splats,
  // are not vectorized/not extractelements (these instructions will be handled
  // by extractelements processing) or may form vector node in future.
  // Cache results - each V in VL is queried up to 3 times (direct +
  // NeighborMightBeIgnored from both neighbors), and areAllUsersVectorized
  // walks each instruction's user list.
  SmallDenseMap<Value *, bool> MightBeIgnoredCache;
  auto MightBeIgnored = [=, &MightBeIgnoredCache](Value *V) {
    auto [It, Inserted] = MightBeIgnoredCache.try_emplace(V);
    if (!Inserted)
      return It->second;
    auto *I = dyn_cast<Instruction>(V);
    bool Res = I && !IsSplatOrUndefs && !isVectorized(I) &&
               !isVectorLikeInstWithConstOps(I) &&
               !areAllUsersVectorized(I, UserIgnoreList) && isSimple(I);
    It->second = Res;
    return Res;
  };
  // Check that the neighbor instruction may form a full vector node with the
  // current instruction V. It is possible, if they have same/alternate opcode
  // and same parent basic block.
  auto NeighborMightBeIgnored = [&](Value *V, int Idx) {
    Value *V1 = VL[Idx];
    bool UsedInSameVTE = false;
    auto It = UsedValuesEntry.find(V1);
    if (It != UsedValuesEntry.end())
      UsedInSameVTE = It->second == UsedValuesEntry.find(V)->second;
    return V != V1 && MightBeIgnored(V1) && !UsedInSameVTE &&
           getSameOpcode({V, V1}, *TLI) &&
           cast<Instruction>(V)->getParent() ==
               cast<Instruction>(V1)->getParent() &&
           (!isa<PHINode>(V1) || AreCompatiblePHIs(V, V1));
  };
  // Build a shuffle mask for better cost estimation and vector emission.
  SmallBitVector UsedIdxs(Entries.size());
  SmallVector<std::pair<unsigned, int>> EntryLanes;
  for (int I = 0, E = VL.size(); I < E; ++I) {
    Value *V = VL[I];
    auto It = UsedValuesEntry.find(V);
    if (It == UsedValuesEntry.end())
      continue;
    // Do not try to shuffle scalars, if they are constants, or instructions
    // that can be vectorized as a result of the following vector build
    // vectorization.
    if (isConstant(V) || (MightBeIgnored(V) &&
                          ((I > 0 && NeighborMightBeIgnored(V, I - 1)) ||
                           (I != E - 1 && NeighborMightBeIgnored(V, I + 1)))))
      continue;
    unsigned Idx = It->second;
    EntryLanes.emplace_back(Idx, I);
    UsedIdxs.set(Idx);
  }
  // Iterate through all shuffled scalars and select entries, which can be used
  // for final shuffle.
  SmallVector<const TreeEntry *> TempEntries;
  for (unsigned I = 0, Sz = Entries.size(); I < Sz; ++I) {
    if (!UsedIdxs.test(I))
      continue;
    // Fix the entry number for the given scalar. If it is the first entry, set
    // Pair.first to 0, otherwise to 1 (currently select at max 2 nodes).
    // These indices are used when calculating final shuffle mask as the vector
    // offset.
    for (std::pair<unsigned, int> &Pair : EntryLanes)
      if (Pair.first == I)
        Pair.first = TempEntries.size();
    TempEntries.push_back(Entries[I]);
  }
  Entries.swap(TempEntries);
  if (EntryLanes.size() == Entries.size() &&
      !VL.equals(ArrayRef(TE->Scalars)
                     .slice(MaskBase, getNumElems(TE->Scalars.size(), SliceSize,
                                                  Part)))) {
    // We may have here 1 or 2 entries only. If the number of scalars is equal
    // to the number of entries, no need to do the analysis, it is not very
    // profitable. Since VL is not the same as TE->Scalars, it means we already
    // have some shuffles before. Cut off not profitable case.
    Entries.clear();
    return std::nullopt;
  }
  // Build the final mask, check for the identity shuffle, if possible.
  bool IsIdentity = Entries.size() == 1;
  // Pair.first is the offset to the vector, while Pair.second is the index of
  // scalar in the list.
  for (const std::pair<unsigned, int> &Pair : EntryLanes) {
    unsigned Idx = MaskBase + Pair.second;
    Mask[Idx] =
        Pair.first * VF +
        (ForOrder ? std::distance(
                        Entries[Pair.first]->Scalars.begin(),
                        find(Entries[Pair.first]->Scalars, VL[Pair.second]))
                  : Entries[Pair.first]->findLaneForValue(VL[Pair.second]));
    IsIdentity &= Mask[Idx] == Pair.second;
  }
  if (ForOrder || IsIdentity || Entries.empty()) {
    switch (Entries.size()) {
    case 1:
      if (IsIdentity || EntryLanes.size() > 1 || VL.size() <= 2)
        return TargetTransformInfo::SK_PermuteSingleSrc;
      break;
    case 2:
      if (EntryLanes.size() > 2 || VL.size() <= 2)
        return TargetTransformInfo::SK_PermuteTwoSrc;
      break;
    default:
      break;
    }
  } else if (!isa<VectorType>(VL.front()->getType()) &&
             (EntryLanes.size() > Entries.size() || VL.size() <= 2)) {
    // Do the cost estimation if shuffle beneficial than buildvector.
    SmallVector<int> SubMask(std::next(Mask.begin(), MaskBase),
                             std::next(Mask.begin(), MaskBase + VL.size()));
    int MinElement = SubMask.front(), MaxElement = SubMask.front();
    for (int Idx : SubMask) {
      if (Idx == PoisonMaskElem)
        continue;
      if (MinElement == PoisonMaskElem || MinElement % VF > Idx % VF)
        MinElement = Idx;
      if (MaxElement == PoisonMaskElem || MaxElement % VF < Idx % VF)
        MaxElement = Idx;
    }
    assert(MaxElement >= 0 && MinElement >= 0 &&
           MaxElement % VF >= MinElement % VF &&
           "Expected at least single element.");
    // If the leading [0, MinIdx) range sits in its own register part(s),
    // skip those whole parts when sizing the destination - everything below
    // the register-aligned floor is unused and never indexed.
    unsigned Offset = 0;
    unsigned MinIdx = MinElement % VF;
    if (MinIdx > 1) {
      unsigned RegFloor = getFloorFullVectorNumberOfElements(
          *TTI, VL.front()->getType(), MinIdx, SLPReVec);
      auto *RegFloorTy = getWidenedType(VL.front()->getType(), RegFloor);
      unsigned RegFloorParts =
          getNumberOfParts(RegFloorTy, VL.front()->getType(), RegFloor);
      if (RegFloorParts > 1)
        Offset = RegFloor;
    }
    unsigned NewVF =
        std::max<unsigned>(VL.size(), (MaxElement % VF) - Offset + 1);
    if (NewVF < VF) {
      for (int &Idx : SubMask) {
        if (Idx == PoisonMaskElem)
          continue;
        Idx = (Idx % VF) - Offset + (Idx >= static_cast<int>(VF) ? NewVF : 0);
      }
    } else {
      NewVF = VF;
    }
 
    auto *VecTy =
        cast<VectorType>(getWidenedType(VL.front()->getType(), NewVF));
    auto *MaskVecTy =
        cast<VectorType>(getWidenedType(VL.front()->getType(), SubMask.size()));
    auto GetShuffleCost = [&,
                           &TTI = *TTI](ArrayRef<int> Mask,
                                        ArrayRef<const TreeEntry *> Entries,
                                        VectorType *VecTy) -> InstructionCost {
      if (Entries.size() == 1 && Entries.front()->getInterleaveFactor() > 0 &&
          ShuffleVectorInst::isDeInterleaveMaskOfFactor(
              Mask, Entries.front()->getInterleaveFactor()))
        return TTI::TCC_Free;
      return getShuffleCost(TTI,
                            Entries.size() > 1 ? TTI::SK_PermuteTwoSrc
                                               : TTI::SK_PermuteSingleSrc,
                            VecTy, CostKind, Mask);
    };
    InstructionCost ShuffleCost = GetShuffleCost(SubMask, Entries, VecTy);
    InstructionCost FirstShuffleCost = 0;
    SmallVector<int> FirstMask(SubMask.begin(), SubMask.end());
    if (Entries.size() == 1 || !Entries[0]->isGather()) {
      FirstShuffleCost = ShuffleCost;
    } else {
      // Transform mask to include only first entry.
      APInt DemandedElts = APInt::getAllOnes(SubMask.size());
      bool IsIdentity = true;
      for (auto [I, Idx] : enumerate(FirstMask)) {
        if (Idx >= static_cast<int>(NewVF)) {
          Idx = PoisonMaskElem;
        } else {
          DemandedElts.clearBit(I);
          if (Idx != PoisonMaskElem)
            IsIdentity &= static_cast<int>(I) == Idx;
        }
      }
      if (!IsIdentity)
        FirstShuffleCost = GetShuffleCost(FirstMask, Entries.front(), VecTy);
      FirstShuffleCost +=
          getScalarizationOverhead(*TTI, SLPReVec, VL.front()->getType(),
                                   MaskVecTy, DemandedElts, /*Insert=*/true,
                                   /*Extract=*/false, CostKind);
    }
    InstructionCost SecondShuffleCost = 0;
    SmallVector<int> SecondMask(SubMask.begin(), SubMask.end());
    if (Entries.size() == 1 || !Entries[1]->isGather()) {
      SecondShuffleCost = ShuffleCost;
    } else {
      // Transform mask to include only first entry.
      APInt DemandedElts = APInt::getAllOnes(SubMask.size());
      bool IsIdentity = true;
      for (auto [I, Idx] : enumerate(SecondMask)) {
        if (Idx < static_cast<int>(NewVF) && Idx >= 0) {
          Idx = PoisonMaskElem;
        } else {
          DemandedElts.clearBit(I);
          if (Idx != PoisonMaskElem) {
            Idx -= NewVF;
            IsIdentity &= static_cast<int>(I) == Idx;
          }
        }
      }
      if (!IsIdentity)
        SecondShuffleCost = GetShuffleCost(SecondMask, Entries[1], VecTy);
      SecondShuffleCost +=
          getScalarizationOverhead(*TTI, SLPReVec, VL.front()->getType(),
                                   MaskVecTy, DemandedElts, /*Insert=*/true,
                                   /*Extract=*/false, CostKind);
    }
    APInt DemandedElts = APInt::getAllOnes(SubMask.size());
    for (auto [I, Idx] : enumerate(SubMask))
      if (Idx == PoisonMaskElem)
        DemandedElts.clearBit(I);
    InstructionCost BuildVectorCost =
        getScalarizationOverhead(*TTI, SLPReVec, VL.front()->getType(),
                                 MaskVecTy, DemandedElts, /*Insert=*/true,
                                 /*Extract=*/false, CostKind);
    const TreeEntry *BestEntry = nullptr;
    auto MaskSlice = MutableArrayRef(Mask).slice(MaskBase, VL.size());
    if (FirstShuffleCost < ShuffleCost) {
      for (int &Idx : MaskSlice)
        if (Idx >= static_cast<int>(VF))
          Idx = PoisonMaskElem;
      BestEntry = Entries.front();
      ShuffleCost = FirstShuffleCost;
    }
    if (SecondShuffleCost < ShuffleCost) {
      for (int &Idx : MaskSlice) {
        if (Idx < static_cast<int>(VF))
          Idx = PoisonMaskElem;
        else
          Idx -= VF;
      }
      BestEntry = Entries[1];
      ShuffleCost = SecondShuffleCost;
    }
    if (BuildVectorCost >= ShuffleCost) {
      if (BestEntry) {
        Entries.clear();
        Entries.push_back(BestEntry);
      }
      return Entries.size() > 1 ? TargetTransformInfo::SK_PermuteTwoSrc
                                : TargetTransformInfo::SK_PermuteSingleSrc;
    }
  }
  Entries.clear();
  // Clear the corresponding mask elements.
  std::fill(std::next(Mask.begin(), MaskBase),
            std::next(Mask.begin(), MaskBase + VL.size()), PoisonMaskElem);
  return std::nullopt;
}
 
SmallVector<std::optional<TargetTransformInfo::ShuffleKind>>
BoUpSLP::isGatherShuffledEntry(
    const TreeEntry *TE, ArrayRef<Value *> VL, SmallVectorImpl<int> &Mask,
    SmallVectorImpl<SmallVector<const TreeEntry *>> &Entries, unsigned NumParts,
    bool ForOrder) {
  assert(NumParts > 0 && NumParts < VL.size() &&
         "Expected positive number of registers.");
  Entries.clear();
  // No need to check for the topmost gather node.
  if (TE == &getRootNode() &&
      (!GatheredLoadsEntriesFirst.has_value() ||
       none_of(ArrayRef(VectorizableTree).drop_front(),
               [](const std::unique_ptr<TreeEntry> &TE) {
                 return !TE->isGather();
               })))
    return {};
  Mask.assign(VL.size(), PoisonMaskElem);
  assert((TE->UserTreeIndex || TE == &getRootNode()) &&
         "Expected only single user of the gather node.");
  unsigned PWSz = getFullVectorNumberOfElements(*TTI, VL.front()->getType(),
                                                VL.size(), SLPReVec);
  if (TE->UserTreeIndex && TE->UserTreeIndex.UserTE->isGather() &&
      TE->UserTreeIndex.EdgeIdx == UINT_MAX &&
      (TE->Idx == 0 ||
       (TE->hasState() && TE->getOpcode() == Instruction::ExtractElement) ||
       isSplat(TE->Scalars) ||
       (TE->hasState() &&
        getSameValuesTreeEntry(TE->getMainOp(), TE->Scalars))))
    return {};
  unsigned SliceSize = getPartNumElems(PWSz, NumParts);
  SmallVector<std::optional<TTI::ShuffleKind>> Res;
  for (unsigned Part : seq<unsigned>(NumParts)) {
    if (Part * SliceSize >= VL.size())
      break;
    ArrayRef<Value *> SubVL =
        VL.slice(Part * SliceSize, getNumElems(VL.size(), SliceSize, Part));
    SmallVectorImpl<const TreeEntry *> &SubEntries = Entries.emplace_back();
    std::optional<TTI::ShuffleKind> SubRes =
        isGatherShuffledSingleRegisterEntry(TE, SubVL, Mask, SubEntries, Part,
                                            ForOrder, SliceSize);
    if (!SubRes)
      SubEntries.clear();
    Res.push_back(SubRes);
    if (SubEntries.size() == 1 && *SubRes == TTI::SK_PermuteSingleSrc &&
        SubEntries.front()->getVectorFactor() == VL.size() &&
        (SubEntries.front()->isSame(TE->Scalars) ||
         SubEntries.front()->isSame(VL))) {
      SmallVector<const TreeEntry *> LocalSubEntries;
      LocalSubEntries.swap(SubEntries);
      Entries.clear();
      Res.clear();
      std::iota(Mask.begin(), Mask.end(), 0);
      // Clear undef scalars.
      for (int I = 0, Sz = VL.size(); I < Sz; ++I)
        if (isa<PoisonValue>(VL[I]))
          Mask[I] = PoisonMaskElem;
      Entries.emplace_back(1, LocalSubEntries.front());
      Res.push_back(TargetTransformInfo::SK_PermuteSingleSrc);
      return Res;
    }
  }
  if (all_of(Res,
             [](const std::optional<TTI::ShuffleKind> &SK) { return !SK; })) {
    Entries.clear();
    return {};
  }
  return Res;
}
 
InstructionCost BoUpSLP::getGatherCost(ArrayRef<Value *> VL, bool ForPoisonSrc,
                                       Type *ScalarTy) const {
  const unsigned VF = VL.size();
  auto *VecTy = getWidenedType(ScalarTy, VF);
  // Find the cost of inserting/extracting values from the vector.
  // Check if the same elements are inserted several times and count them as
  // shuffle candidates.
  APInt DemandedElements = APInt::getZero(VF);
  InstructionCost Cost;
  auto EstimateInsertCost = [&](unsigned I, Value *V) {
    DemandedElements.setBit(I);
    if (V->getType() != ScalarTy)
      Cost += TTI->getCastInstrCost(Instruction::Trunc, ScalarTy, V->getType(),
                                    TTI::CastContextHint::None, CostKind);
  };
  SmallVector<int> ConstantShuffleMask(VF, PoisonMaskElem);
  std::iota(ConstantShuffleMask.begin(), ConstantShuffleMask.end(), 0);
  for (auto [I, V] : enumerate(VL)) {
    // No need to shuffle duplicates for constants.
    if ((ForPoisonSrc && isConstant(V)) || isa<UndefValue>(V))
      continue;
 
    if (isConstant(V)) {
      ConstantShuffleMask[I] = I + VF;
      continue;
    }
    EstimateInsertCost(I, V);
  }
  // FIXME: add a cost for constant vector materialization.
  bool IsAnyNonUndefConst =
      any_of(VL, [](Value *V) { return !isa<UndefValue>(V) && isConstant(V); });
  // 1. Shuffle input source vector and constant vector.
  if (!ForPoisonSrc && IsAnyNonUndefConst) {
    Cost +=
        getShuffleCost(*TTI, TargetTransformInfo::SK_PermuteTwoSrc,
                       cast<VectorType>(VecTy), CostKind, ConstantShuffleMask);
  }
 
  // 2. Insert unique non-constants.
  if (!DemandedElements.isZero())
    Cost += getScalarizationOverhead(
        *TTI, SLPReVec, ScalarTy, cast<VectorType>(VecTy), DemandedElements,
        /*Insert=*/true,
        /*Extract=*/false, CostKind, ForPoisonSrc && !IsAnyNonUndefConst, VL);
  return Cost;
}
 
Instruction &BoUpSLP::getLastInstructionInBundle(const TreeEntry *E) {
  auto It = EntryToLastInstruction.find(E);
  if (It != EntryToLastInstruction.end())
    return *cast<Instruction>(It->second);
  Instruction *Res = nullptr;
  // Get the basic block this bundle is in. All instructions in the bundle
  // should be in this block (except for extractelement-like instructions with
  // constant indices or gathered loads or copyables).
  Instruction *Front;
  unsigned Opcode;
  if (E->hasState()) {
    Front = E->getMainOp();
    Opcode = E->getOpcode();
  } else {
    Front = cast<Instruction>(*find_if(E->Scalars, IsaPred<Instruction>));
    Opcode = Front->getOpcode();
  }
  auto *BB = Front->getParent();
  assert(
      ((GatheredLoadsEntriesFirst.has_value() && Opcode == Instruction::Load &&
        E->isGather() && E->Idx < *GatheredLoadsEntriesFirst) ||
       E->State == TreeEntry::SplitVectorize || E->hasCopyableElements() ||
       all_of(E->Scalars,
              [=](Value *V) -> bool {
                if (Opcode == Instruction::GetElementPtr &&
                    !isa<GetElementPtrInst>(V))
                  return true;
                auto *I = dyn_cast<Instruction>(V);
                return !I || !E->getMatchingMainOpOrAltOp(I) ||
                       I->getParent() == BB || isVectorLikeInstWithConstOps(I);
              })) &&
      "Expected gathered loads or GEPs or instructions from same basic "
      "block.");
 
  auto FindLastInst = [&]() {
    Instruction *LastInst = Front;
    for (Value *V : E->Scalars) {
      auto *I = dyn_cast<Instruction>(V);
      if (!I)
        continue;
      if (E->isCopyableElement(I))
        continue;
      if (LastInst->getParent() == I->getParent()) {
        if (LastInst->comesBefore(I))
          LastInst = I;
        continue;
      }
      assert(((Opcode == Instruction::GetElementPtr &&
               !isa<GetElementPtrInst>(I)) ||
              E->State == TreeEntry::SplitVectorize ||
              (isVectorLikeInstWithConstOps(LastInst) &&
               isVectorLikeInstWithConstOps(I)) ||
              (GatheredLoadsEntriesFirst.has_value() &&
               Opcode == Instruction::Load && E->isGather() &&
               E->Idx < *GatheredLoadsEntriesFirst)) &&
             "Expected vector-like or non-GEP in GEP node insts only.");
      if (!DT->isReachableFromEntry(LastInst->getParent())) {
        LastInst = I;
        continue;
      }
      if (!DT->isReachableFromEntry(I->getParent()))
        continue;
      auto *NodeA = DT->getNode(LastInst->getParent());
      auto *NodeB = DT->getNode(I->getParent());
      assert(NodeA && "Should only process reachable instructions");
      assert(NodeB && "Should only process reachable instructions");
      assert((NodeA == NodeB) ==
                 (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) &&
             "Different nodes should have different DFS numbers");
      if (NodeA->getDFSNumIn() < NodeB->getDFSNumIn())
        LastInst = I;
    }
    BB = LastInst->getParent();
    return LastInst;
  };
 
  auto FindFirstInst = [&]() {
    Instruction *FirstInst = Front;
    for (Value *V : E->Scalars) {
      auto *I = dyn_cast<Instruction>(V);
      if (!I)
        continue;
      if (E->isCopyableElement(I))
        continue;
      if (FirstInst->getParent() == I->getParent()) {
        if (I->comesBefore(FirstInst))
          FirstInst = I;
        continue;
      }
      assert(((Opcode == Instruction::GetElementPtr &&
               !isa<GetElementPtrInst>(I)) ||
              (isVectorLikeInstWithConstOps(FirstInst) &&
               isVectorLikeInstWithConstOps(I))) &&
             "Expected vector-like or non-GEP in GEP node insts only.");
      if (!DT->isReachableFromEntry(FirstInst->getParent())) {
        FirstInst = I;
        continue;
      }
      if (!DT->isReachableFromEntry(I->getParent()))
        continue;
      auto *NodeA = DT->getNode(FirstInst->getParent());
      auto *NodeB = DT->getNode(I->getParent());
      assert(NodeA && "Should only process reachable instructions");
      assert(NodeB && "Should only process reachable instructions");
      assert((NodeA == NodeB) ==
                 (NodeA->getDFSNumIn() == NodeB->getDFSNumIn()) &&
             "Different nodes should have different DFS numbers");
      if (NodeA->getDFSNumIn() > NodeB->getDFSNumIn())
        FirstInst = I;
    }
    return FirstInst;
  };
 
  if (E->State == TreeEntry::SplitVectorize) {
    Res = FindLastInst();
    if (ArrayRef<TreeEntry *> Entries = getTreeEntries(Res); !Entries.empty()) {
      for (auto *E : Entries) {
        auto *I = dyn_cast_or_null<Instruction>(E->VectorizedValue);
        if (!I)
          I = &getLastInstructionInBundle(E);
        if (Res->getParent() == I->getParent() && Res->comesBefore(I))
          Res = I;
      }
    }
    EntryToLastInstruction.try_emplace(E, Res);
    return *Res;
  }
 
  // Set insertpoint for gathered loads to the very first load.
  if (GatheredLoadsEntriesFirst.has_value() &&
      E->Idx >= *GatheredLoadsEntriesFirst && !E->isGather() &&
      Opcode == Instruction::Load) {
    Res = FindFirstInst();
    EntryToLastInstruction.try_emplace(E, Res);
    return *Res;
  }
 
  // Set the insert point to the beginning of the basic block if the entry
  // should not be scheduled.
  auto FindScheduleBundle = [&](const TreeEntry *E) -> const ScheduleBundle * {
    if (E->isGather())
      return nullptr;
    // Found previously that the instruction do not need to be scheduled.
    const auto *It = BlocksSchedules.find(BB);
    if (It == BlocksSchedules.end())
      return nullptr;
    for (Value *V : E->Scalars) {
      auto *I = dyn_cast<Instruction>(V);
      if (!I || isa<PHINode>(I) ||
          (!E->isCopyableElement(I) && doesNotNeedToBeScheduled(I)))
        continue;
      ArrayRef<ScheduleBundle *> Bundles = It->second->getScheduleBundles(I);
      if (Bundles.empty())
        continue;
      const auto *It = find_if(
          Bundles, [&](ScheduleBundle *B) { return B->getTreeEntry() == E; });
      if (It != Bundles.end())
        return *It;
    }
    return nullptr;
  };
  const ScheduleBundle *Bundle = FindScheduleBundle(E);
  if (!E->isGather() && !Bundle) {
    if ((Opcode == Instruction::GetElementPtr &&
         any_of(E->Scalars,
                [](Value *V) {
                  return !isa<GetElementPtrInst>(V) && isa<Instruction>(V);
                })) ||
        (all_of(E->Scalars,
                [&](Value *V) {
                  return isa<PoisonValue>(V) ||
                         (E->Idx == 0 &&
                          isa<InsertElementInst, InsertValueInst>(V)) ||
                         E->isCopyableElement(V) ||
                         (!isVectorLikeInstWithConstOps(V) &&
                          isUsedOutsideBlock(V));
                }) &&
         (!E->doesNotNeedToSchedule() ||
          any_of(E->Scalars,
                 [&](Value *V) {
                   if (!isa<Instruction>(V) ||
                       (E->hasCopyableElements() && E->isCopyableElement(V)))
                     return false;
                   return !areAllOperandsNonInsts(V);
                 }) ||
          none_of(E->Scalars, [&](Value *V) {
            if (!isa<Instruction>(V) ||
                (E->hasCopyableElements() && E->isCopyableElement(V)))
              return false;
            return MustGather.contains(V);
          }))))
      Res = FindLastInst();
    else
      Res = FindFirstInst();
    EntryToLastInstruction.try_emplace(E, Res);
    return *Res;
  }
 
  // Find the last instruction. The common case should be that BB has been
  // scheduled, and the last instruction is VL.back(). So we start with
  // VL.back() and iterate over schedule data until we reach the end of the
  // bundle. The end of the bundle is marked by null ScheduleData.
  if (Bundle) {
    assert(!E->isGather() && "Gathered instructions should not be scheduled");
    Res = Bundle->getBundle().back()->getInst();
    EntryToLastInstruction.try_emplace(E, Res);
    return *Res;
  }
 
  // LastInst can still be null at this point if there's either not an entry
  // for BB in BlocksSchedules or there's no ScheduleData available for
  // VL.back(). This can be the case if buildTreeRec aborts for various
  // reasons (e.g., the maximum recursion depth is reached, the maximum region
  // size is reached, etc.). ScheduleData is initialized in the scheduling
  // "dry-run".
  //
  // If this happens, we can still find the last instruction by brute force. We
  // iterate forwards from Front (inclusive) until we either see all
  // instructions in the bundle or reach the end of the block. If Front is the
  // last instruction in program order, LastInst will be set to Front, and we
  // will visit all the remaining instructions in the block.
  //
  // One of the reasons we exit early from buildTreeRec is to place an upper
  // bound on compile-time. Thus, taking an additional compile-time hit here is
  // not ideal. However, this should be exceedingly rare since it requires that
  // we both exit early from buildTreeRec and that the bundle be out-of-order
  // (causing us to iterate all the way to the end of the block).
  if (!Res)
    Res = FindLastInst();
  assert(Res && "Failed to find last instruction in bundle");
  EntryToLastInstruction.try_emplace(E, Res);
  return *Res;
}
 
void BoUpSLP::setInsertPointAfterBundle(const TreeEntry *E) {
  auto *Front = E->getMainOp();
  Instruction *LastInst = &getLastInstructionInBundle(E);
  assert(LastInst && "Failed to find last instruction in bundle");
  BasicBlock::iterator LastInstIt = LastInst->getIterator();
  // If the instruction is PHI, set the insert point after all the PHIs.
  bool IsPHI = isa<PHINode>(LastInst);
  if (IsPHI) {
    LastInstIt = LastInst->getParent()->getFirstNonPHIIt();
    if (LastInstIt != LastInst->getParent()->end() &&
        LastInstIt->getParent()->isLandingPad())
      LastInstIt = std::next(LastInstIt);
  }
  if (IsPHI ||
      (!E->isGather() && E->State != TreeEntry::SplitVectorize &&
       (E->doesNotNeedToSchedule() ||
        (E->hasCopyableElements() && !E->isCopyableElement(LastInst) &&
         isUsedOutsideBlock(LastInst)))) ||
      (GatheredLoadsEntriesFirst.has_value() &&
       E->Idx >= *GatheredLoadsEntriesFirst && !E->isGather() &&
       E->getOpcode() == Instruction::Load)) {
    Builder.SetInsertPoint(LastInst->getParent(), LastInstIt);
  } else {
    // Set the insertion point after the last instruction in the bundle. Set the
    // debug location to Front.
    Builder.SetInsertPoint(
        LastInst->getParent(),
        LastInst->getNextNode()->getIterator());
    if (Instruction *Res = LastInstructionToPos.lookup(LastInst)) {
      Builder.SetInsertPoint(LastInst->getParent(), Res->getIterator());
    } else {
      Res = Builder.CreateAlignedLoad(Builder.getPtrTy(),
                                      PoisonValue::get(Builder.getPtrTy()),
                                      MaybeAlign());
      Builder.SetInsertPoint(LastInst->getParent(), Res->getIterator());
      eraseInstruction(Res);
      if (E->State != TreeEntry::SplitVectorize)
        LastInstructionToPos.try_emplace(LastInst, Res);
    }
  }
  Builder.SetCurrentDebugLocation(Front->getDebugLoc());
}
 
Value *BoUpSLP::gather(
    ArrayRef<Value *> VL, Value *Root, Type *ScalarTy,
    function_ref<Value *(Value *, Value *, ArrayRef<int>)> CreateShuffle) {
  // List of instructions/lanes from current block and/or the blocks which are
  // part of the current loop. These instructions will be inserted at the end to
  // make it possible to optimize loops and hoist invariant instructions out of
  // the loops body with better chances for success.
  SmallVector<std::pair<Value *, unsigned>, 4> PostponedInsts;
  SmallSet<int, 4> PostponedIndices;
  Loop *L = LI->getLoopFor(Builder.GetInsertBlock());
  auto &&CheckPredecessor = [](BasicBlock *InstBB, BasicBlock *InsertBB) {
    SmallPtrSet<BasicBlock *, 4> Visited;
    while (InsertBB && InsertBB != InstBB && Visited.insert(InsertBB).second)
      InsertBB = InsertBB->getSinglePredecessor();
    return InsertBB && InsertBB == InstBB;
  };
  for (int I = 0, E = VL.size(); I < E; ++I) {
    if (auto *Inst = dyn_cast<Instruction>(VL[I]))
      if ((CheckPredecessor(Inst->getParent(), Builder.GetInsertBlock()) ||
           isVectorized(Inst) ||
           (L && (!Root || L->isLoopInvariant(Root)) && L->contains(Inst))) &&
          PostponedIndices.insert(I).second)
        PostponedInsts.emplace_back(Inst, I);
  }
 
  auto &&CreateInsertElement = [this](Value *Vec, Value *V, unsigned Pos,
                                      Type *Ty) {
    Value *Scalar = V;
    // Drop NUW from trunc to avoid incorrect codegen.
    Value *Trunced;
    if (match(Scalar, m_NUWTrunc(m_Value(Trunced))))
      cast<TruncInst>(Scalar)->setHasNoUnsignedWrap(/*B=*/false);
    if (Scalar->getType() != Ty) {
      assert(Scalar->getType()->isIntOrIntVectorTy() &&
             Ty->isIntOrIntVectorTy() && "Expected integer types only.");
      Value *V = Scalar;
      if (auto *CI = dyn_cast<CastInst>(Scalar);
          isa_and_nonnull<SExtInst, ZExtInst>(CI)) {
        Value *Op = CI->getOperand(0);
        if (auto *IOp = dyn_cast<Instruction>(Op);
            !IOp || !(isDeleted(IOp) || isVectorized(IOp)))
          V = Op;
      }
      Scalar = Builder.CreateIntCast(
          V, Ty, !isKnownNonNegative(Scalar, SimplifyQuery(*DL)));
    }
 
    Instruction *InsElt;
    if (auto *VecTy = dyn_cast<FixedVectorType>(Scalar->getType())) {
      assert(SLPReVec && "FixedVectorType is not expected.");
      Vec =
          createInsertVector(Builder, Vec, Scalar, Pos * getNumElements(VecTy));
      auto *II = dyn_cast<Instruction>(Vec);
      if (!II)
        return Vec;
      InsElt = II;
    } else {
      Vec = Builder.CreateInsertElement(Vec, Scalar, Pos);
      InsElt = dyn_cast<InsertElementInst>(Vec);
      if (!InsElt)
        return Vec;
    }
    GatherShuffleExtractSeq.insert(InsElt);
    CSEBlocks.insert(InsElt->getParent());
    // Add to our 'need-to-extract' list.
    if (isa<Instruction>(V)) {
      ArrayRef<TreeEntry *> Entries = getTreeEntries(V);
      const auto *It = find_if(Entries, [&](const TreeEntry *E) {
        return !TransformedToGatherNodes.contains(E) &&
               !DeletedNodes.contains(E);
      });
      if (It != Entries.end()) {
        // Find which lane we need to extract.
        User *UserOp = nullptr;
        if (Scalar != V) {
          if (auto *SI = dyn_cast<Instruction>(Scalar))
            UserOp = SI;
        } else {
          if (V->getType()->isVectorTy()) {
            if (auto *SV = dyn_cast<ShuffleVectorInst>(InsElt);
                SV && SV->getOperand(0) != V && SV->getOperand(1) != V) {
              // Find shufflevector, caused by resize.
              auto FindOperand = [](Value *Vec, Value *V) -> Instruction * {
                if (auto *SV = dyn_cast<ShuffleVectorInst>(Vec)) {
                  if (SV->getOperand(0) == V)
                    return SV;
                  if (SV->getOperand(1) == V)
                    return SV;
                }
                return nullptr;
              };
              InsElt = nullptr;
              if (Instruction *User = FindOperand(SV->getOperand(0), V))
                InsElt = User;
              else if (Instruction *User = FindOperand(SV->getOperand(1), V))
                InsElt = User;
              assert(InsElt &&
                     "Failed to find shufflevector, caused by resize.");
            } else if (SLPReVec && isa<ShuffleVectorInst>(InsElt)) {
              // ReVec gather used V directly as a shufflevector operand.
              // Register a nullptr-User external use so all remaining
              // in-IR uses of V get rewritten via replaceAllUsesWith,
              // and track V in ExternalUsesWithNonUsers to match the
              // bookkeeping done by buildExternalUses.
              unsigned FoundLane = (*It)->findLaneForValue(V);
              ExternalUses.emplace_back(V, nullptr, **It, FoundLane);
              ExternalUsesWithNonUsers.insert(V);
            }
          }
          UserOp = InsElt;
        }
        if (UserOp) {
          unsigned FoundLane = (*It)->findLaneForValue(V);
          ExternalUses.emplace_back(V, UserOp, **It, FoundLane);
        }
      }
    }
    return Vec;
  };
  auto *VecTy = getWidenedType(ScalarTy, VL.size());
  Value *Vec = PoisonValue::get(VecTy);
  SmallVector<int> NonConsts;
  SmallVector<int> Mask(VL.size());
  std::iota(Mask.begin(), Mask.end(), 0);
  Value *OriginalRoot = Root;
  if (auto *SV = dyn_cast_or_null<ShuffleVectorInst>(Root);
      SV && isa<PoisonValue>(SV->getOperand(1)) &&
      SV->getOperand(0)->getType() == VecTy) {
    Root = SV->getOperand(0);
    Mask.assign(SV->getShuffleMask().begin(), SV->getShuffleMask().end());
  }
  // Insert constant values at first.
  for (int I = 0, E = VL.size(); I < E; ++I) {
    if (PostponedIndices.contains(I))
      continue;
    if (!isConstant(VL[I])) {
      NonConsts.push_back(I);
      continue;
    }
    if (isa<PoisonValue>(VL[I]))
      continue;
    Vec = CreateInsertElement(Vec, VL[I], I, ScalarTy);
    Mask[I] = I + E;
  }
  if (Root) {
    if (isa<PoisonValue>(Vec)) {
      Vec = OriginalRoot;
    } else {
      Vec = CreateShuffle(Root, Vec, Mask);
      if (auto *OI = dyn_cast<Instruction>(OriginalRoot);
          OI && OI->use_empty() &&
          none_of(VectorizableTree, [&](const std::unique_ptr<TreeEntry> &TE) {
            return TE->VectorizedValue == OI;
          }))
        eraseInstruction(OI);
    }
  }
  // Insert non-constant values.
  for (int I : NonConsts)
    Vec = CreateInsertElement(Vec, VL[I], I, ScalarTy);
  // Append instructions, which are/may be part of the loop, in the end to make
  // it possible to hoist non-loop-based instructions.
  for (const std::pair<Value *, unsigned> &Pair : PostponedInsts)
    Vec = CreateInsertElement(Vec, Pair.first, Pair.second, ScalarTy);
 
  return Vec;
}
 
/// Merges shuffle masks and emits final shuffle instruction, if required. It
/// supports shuffling of 2 input vectors. It implements lazy shuffles emission,
/// when the actual shuffle instruction is generated only if this is actually
/// required. Otherwise, the shuffle instruction emission is delayed till the
/// end of the process, to reduce the number of emitted instructions and further
/// analysis/transformations.
/// The class also will look through the previously emitted shuffle instructions
/// and properly mark indices in mask as undef.
/// For example, given the code
/// \code
/// %s1 = shufflevector <2 x ty> %0, poison, <1, 0>
/// %s2 = shufflevector <2 x ty> %1, poison, <1, 0>
/// \endcode
/// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 3, 2>, it will
/// look through %s1 and %s2 and emit
/// \code
/// %res = shufflevector <2 x ty> %0, %1, <0, 1, 2, 3>
/// \endcode
/// instead.
/// If 2 operands are of different size, the smallest one will be resized and
/// the mask recalculated properly.
/// For example, given the code
/// \code
/// %s1 = shufflevector <2 x ty> %0, poison, <1, 0, 1, 0>
/// %s2 = shufflevector <2 x ty> %1, poison, <1, 0, 1, 0>
/// \endcode
/// and if need to emit shuffle of %s1 and %s2 with mask <1, 0, 5, 4>, it will
/// look through %s1 and %s2 and emit
/// \code
/// %res = shufflevector <2 x ty> %0, %1, <0, 1, 2, 3>
/// \endcode
/// instead.
class BoUpSLP::ShuffleInstructionBuilder final : public BaseShuffleAnalysis {
  bool IsFinalized = false;
  /// Combined mask for all applied operands and masks. It is built during
  /// analysis and actual emission of shuffle vector instructions.
  SmallVector<int> CommonMask;
  /// List of operands for the shuffle vector instruction. It hold at max 2
  /// operands, if the 3rd is going to be added, the first 2 are combined into
  /// shuffle with \p CommonMask mask, the first operand sets to be the
  /// resulting shuffle and the second operand sets to be the newly added
  /// operand. The \p CommonMask is transformed in the proper way after that.
  SmallVector<Value *, 2> InVectors;
  IRBuilderBase &Builder;
  BoUpSLP &R;
 
  class ShuffleIRBuilder {
    IRBuilderBase &Builder;
    /// Holds all of the instructions that we gathered.
    SetVector<Instruction *> &GatherShuffleExtractSeq;
    /// A list of blocks that we are going to CSE.
    DenseSet<BasicBlock *> &CSEBlocks;
    /// Data layout.
    const DataLayout &DL;
 
  public:
    ShuffleIRBuilder(IRBuilderBase &Builder,
                     SetVector<Instruction *> &GatherShuffleExtractSeq,
                     DenseSet<BasicBlock *> &CSEBlocks, const DataLayout &DL)
        : Builder(Builder), GatherShuffleExtractSeq(GatherShuffleExtractSeq),
          CSEBlocks(CSEBlocks), DL(DL) {}
    ~ShuffleIRBuilder() = default;
    /// Creates shufflevector for the 2 operands with the given mask.
    Value *createShuffleVector(Value *V1, Value *V2, ArrayRef<int> Mask) {
      if (V1->getType() != V2->getType()) {
        assert(V1->getType()->isIntOrIntVectorTy() &&
               V1->getType()->isIntOrIntVectorTy() &&
               "Expected integer vector types only.");
        if (V1->getType() != V2->getType()) {
          if (cast<VectorType>(V2->getType())
                  ->getElementType()
                  ->getIntegerBitWidth() < cast<VectorType>(V1->getType())
                                               ->getElementType()
                                               ->getIntegerBitWidth())
            V2 = Builder.CreateIntCast(
                V2, V1->getType(), !isKnownNonNegative(V2, SimplifyQuery(DL)));
          else
            V1 = Builder.CreateIntCast(
                V1, V2->getType(), !isKnownNonNegative(V1, SimplifyQuery(DL)));
        }
      }
      Value *Vec = Builder.CreateShuffleVector(V1, V2, Mask);
      if (auto *I = dyn_cast<Instruction>(Vec)) {
        GatherShuffleExtractSeq.insert(I);
        CSEBlocks.insert(I->getParent());
      }
      return Vec;
    }
    /// Creates permutation of the single vector operand with the given mask, if
    /// it is not identity mask.
    Value *createShuffleVector(Value *V1, ArrayRef<int> Mask) {
      if (Mask.empty())
        return V1;
      unsigned VF = Mask.size();
      unsigned LocalVF = cast<FixedVectorType>(V1->getType())->getNumElements();
      if (VF == LocalVF && ShuffleVectorInst::isIdentityMask(Mask, VF))
        return V1;
      Value *Vec = Builder.CreateShuffleVector(V1, Mask);
      if (auto *I = dyn_cast<Instruction>(Vec)) {
        GatherShuffleExtractSeq.insert(I);
        CSEBlocks.insert(I->getParent());
      }
      return Vec;
    }
    Value *createIdentity(Value *V) { return V; }
    Value *createPoison(Type *Ty, unsigned VF) {
      return PoisonValue::get(getWidenedType(Ty, VF));
    }
    /// Resizes 2 input vector to match the sizes, if the they are not equal
    /// yet. The smallest vector is resized to the size of the larger vector.
    void resizeToMatch(Value *&V1, Value *&V2) {
      if (V1->getType() == V2->getType())
        return;
      int V1VF = cast<FixedVectorType>(V1->getType())->getNumElements();
      int V2VF = cast<FixedVectorType>(V2->getType())->getNumElements();
      int VF = std::max(V1VF, V2VF);
      int MinVF = std::min(V1VF, V2VF);
      SmallVector<int> IdentityMask(VF, PoisonMaskElem);
      std::iota(IdentityMask.begin(), std::next(IdentityMask.begin(), MinVF),
                0);
      Value *&Op = MinVF == V1VF ? V1 : V2;
      Op = Builder.CreateShuffleVector(Op, IdentityMask);
      if (auto *I = dyn_cast<Instruction>(Op)) {
        GatherShuffleExtractSeq.insert(I);
        CSEBlocks.insert(I->getParent());
      }
      if (MinVF == V1VF)
        V1 = Op;
      else
        V2 = Op;
    }
  };
 
  /// Smart shuffle instruction emission, walks through shuffles trees and
  /// tries to find the best matching vector for the actual shuffle
  /// instruction.
  Value *createShuffle(Value *V1, Value *V2, ArrayRef<int> Mask) {
    assert(V1 && "Expected at least one vector value.");
    ShuffleIRBuilder ShuffleBuilder(Builder, R.GatherShuffleExtractSeq,
                                    R.CSEBlocks, *R.DL);
    return BaseShuffleAnalysis::createShuffle<Value *>(
        V1, V2, Mask, ShuffleBuilder, ScalarTy);
  }
 
  /// Cast value \p V to the vector type with the same number of elements, but
  /// the base type \p ScalarTy.
  Value *castToScalarTyElem(Value *V,
                            std::optional<bool> IsSigned = std::nullopt) {
    auto *VecTy = cast<VectorType>(V->getType());
    assert(getNumElements(VecTy) % getNumElements(ScalarTy) == 0);
    if (VecTy->getElementType() == ScalarTy->getScalarType())
      return V;
    return Builder.CreateIntCast(
        V, VectorType::get(ScalarTy->getScalarType(), VecTy->getElementCount()),
        IsSigned.value_or(!isKnownNonNegative(V, SimplifyQuery(*R.DL))));
  }
 
  Value *getVectorizedValue(const TreeEntry &E) {
    Value *Vec = E.VectorizedValue;
    if (!Vec->getType()->isIntOrIntVectorTy())
      return Vec;
    return castToScalarTyElem(Vec, any_of(E.Scalars, [&](Value *V) {
                                return !isa<PoisonValue>(V) &&
                                       !isKnownNonNegative(
                                           V, SimplifyQuery(*R.DL));
                              }));
  }
 
public:
  ShuffleInstructionBuilder(Type *ScalarTy, IRBuilderBase &Builder, BoUpSLP &R)
      : BaseShuffleAnalysis(ScalarTy), Builder(Builder), R(R) {}
 
  /// Adjusts extractelements after reusing them.
  Value *adjustExtracts(const TreeEntry *E, MutableArrayRef<int> Mask,
                        ArrayRef<std::optional<TTI::ShuffleKind>> ShuffleKinds,
                        unsigned NumParts, bool &UseVecBaseAsInput) {
    UseVecBaseAsInput = false;
    SmallPtrSet<Value *, 4> UniqueBases;
    Value *VecBase = nullptr;
    SmallVector<Value *> VL(E->Scalars.begin(), E->Scalars.end());
    if (!E->ReorderIndices.empty()) {
      SmallVector<int> ReorderMask(E->ReorderIndices.begin(),
                                   E->ReorderIndices.end());
      reorderScalars(VL, ReorderMask);
    }
    for (int I = 0, Sz = Mask.size(); I < Sz; ++I) {
      int Idx = Mask[I];
      if (Idx == PoisonMaskElem)
        continue;
      auto *EI = cast<ExtractElementInst>(VL[I]);
      VecBase = EI->getVectorOperand();
      if (ArrayRef<TreeEntry *> TEs = R.getTreeEntries(VecBase); !TEs.empty())
        VecBase = TEs.front()->VectorizedValue;
      assert(VecBase && "Expected vectorized value.");
      UniqueBases.insert(VecBase);
      // If the only one use is vectorized - can delete the extractelement
      // itself.
      if (!EI->hasOneUse() || R.ExternalUsesAsOriginalScalar.contains(EI) ||
          (E->UserTreeIndex && E->UserTreeIndex.EdgeIdx == UINT_MAX &&
           !R.isVectorized(EI) &&
           count_if(E->Scalars, [&](Value *V) { return V == EI; }) !=
               count_if(E->UserTreeIndex.UserTE->Scalars,
                        [&](Value *V) { return V == EI; })) ||
          (NumParts != 1 && count(VL, EI) > 1) ||
          any_of(EI->users(), [&](User *U) {
            ArrayRef<TreeEntry *> UTEs = R.getTreeEntries(U);
            return UTEs.empty() || UTEs.size() > 1 ||
                   any_of(UTEs,
                          [&](const TreeEntry *TE) {
                            return R.DeletedNodes.contains(TE) ||
                                   R.TransformedToGatherNodes.contains(TE);
                          }) ||
                   (isa<GetElementPtrInst>(U) &&
                    !R.areAllUsersVectorized(cast<Instruction>(U))) ||
                   (!UTEs.empty() &&
                    count_if(R.VectorizableTree,
                             [&](const std::unique_ptr<TreeEntry> &TE) {
                               return TE->UserTreeIndex.UserTE ==
                                          UTEs.front() &&
                                      is_contained(VL, EI);
                             }) != 1);
          }))
        continue;
      R.eraseInstruction(EI);
    }
    if (NumParts == 1 || UniqueBases.size() == 1) {
      assert(VecBase && "Expected vectorized value.");
      return castToScalarTyElem(VecBase);
    }
    UseVecBaseAsInput = true;
    auto TransformToIdentity = [](MutableArrayRef<int> Mask) {
      for (auto [I, Idx] : enumerate(Mask))
        if (Idx != PoisonMaskElem)
          Idx = I;
    };
    // Perform multi-register vector shuffle, joining them into a single virtual
    // long vector.
    // Need to shuffle each part independently and then insert all this parts
    // into a long virtual vector register, forming the original vector.
    Value *Vec = nullptr;
    SmallVector<int> VecMask(Mask.size(), PoisonMaskElem);
    unsigned SliceSize = getPartNumElems(VL.size(), NumParts);
    for (unsigned Part : seq<unsigned>(NumParts)) {
      unsigned Limit = getNumElems(VL.size(), SliceSize, Part);
      ArrayRef<Value *> SubVL = ArrayRef(VL).slice(Part * SliceSize, Limit);
      MutableArrayRef<int> SubMask = Mask.slice(Part * SliceSize, Limit);
      constexpr int MaxBases = 2;
      SmallVector<Value *, MaxBases> Bases(MaxBases);
      auto VLMask = zip(SubVL, SubMask);
      const unsigned VF =
          accumulate(VLMask, 0U, [&](unsigned S, const auto &D) {
            if (std::get<1>(D) == PoisonMaskElem)
              return S;
            Value *VecOp =
                cast<ExtractElementInst>(std::get<0>(D))->getVectorOperand();
            if (ArrayRef<TreeEntry *> TEs = R.getTreeEntries(VecOp);
                !TEs.empty())
              VecOp = TEs.front()->VectorizedValue;
            assert(VecOp && "Expected vectorized value.");
            const unsigned Size =
                cast<FixedVectorType>(VecOp->getType())->getNumElements();
            return std::max(S, Size);
          });
      for (const auto [V, I] : VLMask) {
        if (I == PoisonMaskElem)
          continue;
        Value *VecOp = cast<ExtractElementInst>(V)->getVectorOperand();
        if (ArrayRef<TreeEntry *> TEs = R.getTreeEntries(VecOp); !TEs.empty())
          VecOp = TEs.front()->VectorizedValue;
        assert(VecOp && "Expected vectorized value.");
        VecOp = castToScalarTyElem(VecOp);
        Bases[I / VF] = VecOp;
      }
      if (!Bases.front())
        continue;
      Value *SubVec;
      if (Bases.back()) {
        SubVec = createShuffle(Bases.front(), Bases.back(), SubMask);
        TransformToIdentity(SubMask);
      } else {
        SubVec = Bases.front();
      }
      if (!Vec) {
        Vec = SubVec;
        assert((Part == 0 || all_of(seq<unsigned>(0, Part),
                                    [&](unsigned P) {
                                      ArrayRef<int> SubMask =
                                          Mask.slice(P * SliceSize,
                                                     getNumElems(Mask.size(),
                                                                 SliceSize, P));
                                      return all_of(SubMask, [](int Idx) {
                                        return Idx == PoisonMaskElem;
                                      });
                                    })) &&
               "Expected first part or all previous parts masked.");
        copy(SubMask, std::next(VecMask.begin(), Part * SliceSize));
      } else {
        unsigned NewVF =
            cast<FixedVectorType>(Vec->getType())->getNumElements();
        if (Vec->getType() != SubVec->getType()) {
          unsigned SubVecVF =
              cast<FixedVectorType>(SubVec->getType())->getNumElements();
          NewVF = std::max(NewVF, SubVecVF);
        }
        // Adjust SubMask.
        for (int &Idx : SubMask)
          if (Idx != PoisonMaskElem)
            Idx += NewVF;
        copy(SubMask, std::next(VecMask.begin(), Part * SliceSize));
        Vec = createShuffle(Vec, SubVec, VecMask);
        TransformToIdentity(VecMask);
      }
    }
    copy(VecMask, Mask.begin());
    return Vec;
  }
  /// Checks if the specified entry \p E needs to be delayed because of its
  /// dependency nodes.
  std::optional<Value *>
  needToDelay(const TreeEntry *E,
              ArrayRef<SmallVector<const TreeEntry *>> Deps) const {
    // No need to delay emission if all deps are ready.
    if (all_of(Deps, [](ArrayRef<const TreeEntry *> TEs) {
          return all_of(
              TEs, [](const TreeEntry *TE) { return TE->VectorizedValue; });
        }))
      return std::nullopt;
    // Postpone gather emission, will be emitted after the end of the
    // process to keep correct order.
    auto *ResVecTy = getWidenedType(ScalarTy, E->getVectorFactor());
    return Builder.CreateAlignedLoad(
        ResVecTy,
        PoisonValue::get(PointerType::getUnqual(ScalarTy->getContext())),
        MaybeAlign());
  }
  /// Reset the builder to handle perfect diamond match.
  void resetForSameNode() {
    IsFinalized = false;
    CommonMask.clear();
    InVectors.clear();
  }
  /// Adds 2 input vectors (in form of tree entries) and the mask for their
  /// shuffling.
  void add(const TreeEntry &E1, const TreeEntry &E2, ArrayRef<int> Mask) {
    Value *V1 = getVectorizedValue(E1);
    Value *V2 = getVectorizedValue(E2);
    add(V1, V2, Mask);
  }
  /// Adds single input vector (in form of tree entry) and the mask for its
  /// shuffling.
  void add(const TreeEntry &E1, ArrayRef<int> Mask) {
    Value *V1 = getVectorizedValue(E1);
    add(V1, Mask);
  }
  /// Adds 2 input vectors and the mask for their shuffling.
  void add(Value *V1, Value *V2, ArrayRef<int> Mask) {
    assert(V1 && V2 && !Mask.empty() && "Expected non-empty input vectors.");
    assert(isa<FixedVectorType>(V1->getType()) &&
           isa<FixedVectorType>(V2->getType()) &&
           "castToScalarTyElem expects V1 and V2 to be FixedVectorType");
    V1 = castToScalarTyElem(V1);
    V2 = castToScalarTyElem(V2);
    if (InVectors.empty()) {
      InVectors.push_back(V1);
      InVectors.push_back(V2);
      CommonMask.assign(Mask.begin(), Mask.end());
      return;
    }
    Value *Vec = InVectors.front();
    if (InVectors.size() == 2) {
      Vec = createShuffle(Vec, InVectors.back(), CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
    } else if (cast<FixedVectorType>(Vec->getType())->getNumElements() !=
               Mask.size()) {
      Vec = createShuffle(Vec, nullptr, CommonMask);
      transformMaskAfterShuffle(CommonMask, CommonMask);
    }
    V1 = createShuffle(V1, V2, Mask);
    unsigned VF = std::max(getVF(V1), getVF(Vec));
    for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
      if (Mask[Idx] != PoisonMaskElem)
        CommonMask[Idx] = Idx + VF;
    InVectors.front() = Vec;
    if (InVectors.size() == 2)
      InVectors.back() = V1;
    else
      InVectors.push_back(V1);
  }
  /// Adds another one input vector and the mask for the shuffling.
  void add(Value *V1, ArrayRef<int> Mask, bool = false) {
    assert(isa<FixedVectorType>(V1->getType()) &&
           "castToScalarTyElem expects V1 to be FixedVectorType");
    V1 = castToScalarTyElem(V1);
    if (InVectors.empty()) {
      InVectors.push_back(V1);
      CommonMask.assign(Mask.begin(), Mask.end());
      return;
    }
    const auto *It = find(InVectors, V1);
    if (It == InVectors.end()) {
      if (InVectors.size() == 2 ||
          InVectors.front()->getType() != V1->getType()) {
        Value *V = InVectors.front();
        if (InVectors.size() == 2) {
          V = createShuffle(InVectors.front(), InVectors.back(), CommonMask);
          transformMaskAfterShuffle(CommonMask, CommonMask);
        } else if (cast<FixedVectorType>(V->getType())->getNumElements() !=
                   CommonMask.size()) {
          V = createShuffle(InVectors.front(), nullptr, CommonMask);
          transformMaskAfterShuffle(CommonMask, CommonMask);
        }
        unsigned VF = std::max(CommonMask.size(), Mask.size());
        for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
          if (CommonMask[Idx] == PoisonMaskElem && Mask[Idx] != PoisonMaskElem)
            CommonMask[Idx] = V->getType() != V1->getType()
                                  ? Idx + VF
                                  : Mask[Idx] + getVF(V1);
        if (V->getType() != V1->getType())
          V1 = createShuffle(V1, nullptr, Mask);
        InVectors.front() = V;
        if (InVectors.size() == 2)
          InVectors.back() = V1;
        else
          InVectors.push_back(V1);
        return;
      }
      // Check if second vector is required if the used elements are already
      // used from the first one.
      for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
        if (Mask[Idx] != PoisonMaskElem && CommonMask[Idx] == PoisonMaskElem) {
          InVectors.push_back(V1);
          break;
        }
    }
    unsigned VF = 0;
    for (Value *V : InVectors)
      VF = std::max(VF, getVF(V));
    for (unsigned Idx = 0, Sz = CommonMask.size(); Idx < Sz; ++Idx)
      if (Mask[Idx] != PoisonMaskElem && CommonMask[Idx] == PoisonMaskElem)
        CommonMask[Idx] = Mask[Idx] + (It == InVectors.begin() ? 0 : VF);
  }
  /// Adds another one input vector and the mask for the shuffling.
  void addOrdered(Value *V1, ArrayRef<unsigned> Order) {
    SmallVector<int> NewMask;
    inversePermutation(Order, NewMask);
    add(V1, NewMask);
  }
  Value *gather(ArrayRef<Value *> VL, unsigned MaskVF = 0,
                Value *Root = nullptr) {
    return R.gather(VL, Root, ScalarTy,
                    [&](Value *V1, Value *V2, ArrayRef<int> Mask) {
                      return createShuffle(V1, V2, Mask);
                    });
  }
  Value *createFreeze(Value *V) { return Builder.CreateFreeze(V); }
  /// Finalize emission of the shuffles.
  /// \param Action the action (if any) to be performed before final applying of
  /// the \p ExtMask mask.
  Value *finalize(
      ArrayRef<int> ExtMask,
      ArrayRef<std::pair<const TreeEntry *, unsigned>> SubVectors,
      ArrayRef<int> SubVectorsMask, unsigned VF = 0,
      function_ref<void(Value *&, SmallVectorImpl<int> &,
                        function_ref<Value *(Value *, Value *, ArrayRef<int>)>)>
          Action = {}) {
    IsFinalized = true;
    if (Action) {
      Value *Vec = InVectors.front();
      if (InVectors.size() == 2) {
        Vec = createShuffle(Vec, InVectors.back(), CommonMask);
        InVectors.pop_back();
      } else {
        Vec = createShuffle(Vec, nullptr, CommonMask);
      }
      transformMaskAfterShuffle(CommonMask, CommonMask);
      assert(VF > 0 &&
             "Expected vector length for the final value before action.");
      unsigned VecVF = cast<FixedVectorType>(Vec->getType())->getNumElements();
      if (VecVF < VF) {
        SmallVector<int> ResizeMask(VF, PoisonMaskElem);
        std::iota(ResizeMask.begin(), std::next(ResizeMask.begin(), VecVF), 0);
        Vec = createShuffle(Vec, nullptr, ResizeMask);
      }
      Action(Vec, CommonMask, [this](Value *V1, Value *V2, ArrayRef<int> Mask) {
        return createShuffle(V1, V2, Mask);
      });
      InVectors.front() = Vec;
    }
    if (!SubVectors.empty()) {
      Value *Vec = InVectors.front();
      if (InVectors.size() == 2) {
        Vec = createShuffle(Vec, InVectors.back(), CommonMask);
        InVectors.pop_back();
      } else {
        Vec = createShuffle(Vec, nullptr, CommonMask);
      }
      transformMaskAfterShuffle(CommonMask, CommonMask);
      auto CreateSubVectors = [&](Value *Vec,
                                  SmallVectorImpl<int> &CommonMask) {
        for (auto [E, Idx] : SubVectors) {
          Value *V = getVectorizedValue(*E);
          unsigned InsertionIndex = Idx * getNumElements(ScalarTy);
          // Use scalar version of the SCalarType to correctly handle shuffles
          // for revectorization. The revectorization mode operates by the
          // vectors, but here we need to operate on the scalars, because the
          // masks were already transformed for the vector elements and we don't
          // need doing this transformation again.
          Type *OrigScalarTy = ScalarTy;
          ScalarTy = ScalarTy->getScalarType();
          Vec = createInsertVector(
              Builder, Vec, V, InsertionIndex,
              std::bind(&ShuffleInstructionBuilder::createShuffle, this, _1, _2,
                        _3));
          ScalarTy = OrigScalarTy;
          if (!CommonMask.empty()) {
            std::iota(std::next(CommonMask.begin(), Idx),
                      std::next(CommonMask.begin(), Idx + E->getVectorFactor()),
                      Idx);
          }
        }
        return Vec;
      };
      if (SubVectorsMask.empty()) {
        Vec = CreateSubVectors(Vec, CommonMask);
      } else {
        SmallVector<int> SVMask(CommonMask.size(), PoisonMaskElem);
        copy(SubVectorsMask, SVMask.begin());
        for (auto [I1, I2] : zip(SVMask, CommonMask)) {
          if (I2 != PoisonMaskElem) {
            assert(I1 == PoisonMaskElem && "Expected unused subvectors mask");
            I1 = I2 + CommonMask.size();
          }
        }
        Value *InsertVec =
            CreateSubVectors(PoisonValue::get(Vec->getType()), CommonMask);
        Vec = createShuffle(InsertVec, Vec, SVMask);
        transformMaskAfterShuffle(CommonMask, SVMask);
      }
      InVectors.front() = Vec;
    }
 
    if (!ExtMask.empty()) {
      if (CommonMask.empty()) {
        CommonMask.assign(ExtMask.begin(), ExtMask.end());
      } else {
        SmallVector<int> NewMask(ExtMask.size(), PoisonMaskElem);
        for (int I = 0, Sz = ExtMask.size(); I < Sz; ++I) {
          if (ExtMask[I] == PoisonMaskElem)
            continue;
          NewMask[I] = CommonMask[ExtMask[I]];
        }
        CommonMask.swap(NewMask);
      }
    }
    if (CommonMask.empty()) {
      assert(InVectors.size() == 1 && "Expected only one vector with no mask");
      return InVectors.front();
    }
    if (InVectors.size() == 2)
      return createShuffle(InVectors.front(), InVectors.back(), CommonMask);
    return createShuffle(InVectors.front(), nullptr, CommonMask);
  }
 
  ~ShuffleInstructionBuilder() {
    assert((IsFinalized || CommonMask.empty()) &&
           "Shuffle construction must be finalized.");
  }
};
 
Value *BoUpSLP::vectorizeOperand(TreeEntry *E, unsigned NodeIdx) {
  return vectorizeTree(getOperandEntry(E, NodeIdx));
}
 
template <typename BVTy, typename ResTy, typename... Args>
ResTy BoUpSLP::processBuildVector(const TreeEntry *E, Type *ScalarTy,
                                  Args &...Params) {
  assert((E->isGather() || TransformedToGatherNodes.contains(E)) &&
         "Expected gather node.");
  unsigned VF = E->getVectorFactor();
 
  bool NeedFreeze = false;
  SmallVector<Value *> GatheredScalars(E->Scalars.begin(), E->Scalars.end());
  // Do not process split vectorize node, marked to be gathers/buildvectors.
  SmallVector<std::pair<const TreeEntry *, unsigned>> SubVectors(
      E->CombinedEntriesWithIndices.size());
  if (E->State == TreeEntry::SplitVectorize &&
      TransformedToGatherNodes.contains(E)) {
    SubVectors.clear();
  } else {
    // Clear values, to be replaced by insertvector instructions.
    for (auto [EIdx, Idx] : E->CombinedEntriesWithIndices)
      for_each(MutableArrayRef(GatheredScalars)
                   .slice(Idx, VectorizableTree[EIdx]->getVectorFactor()),
               [&](Value *&V) { V = PoisonValue::get(V->getType()); });
    transform(
        E->CombinedEntriesWithIndices, SubVectors.begin(), [&](const auto &P) {
          return std::make_pair(VectorizableTree[P.first].get(), P.second);
        });
  }
  // Build a mask out of the reorder indices and reorder scalars per this
  // mask.
  SmallVector<int> ReorderMask(E->ReorderIndices.begin(),
                               E->ReorderIndices.end());
  if (!ReorderMask.empty())
    reorderScalars(GatheredScalars, ReorderMask);
  SmallVector<int> SubVectorsMask;
  inversePermutation(E->ReorderIndices, SubVectorsMask);
  // Transform non-clustered elements in the mask to poison (-1).
  // "Clustered" operations will be reordered using this mask later.
  if (!SubVectors.empty() && !SubVectorsMask.empty()) {
    for (unsigned I : seq<unsigned>(GatheredScalars.size()))
      if (E->Scalars[I] == GatheredScalars[ReorderMask[I]])
        SubVectorsMask[ReorderMask[I]] = PoisonMaskElem;
  } else {
    SubVectorsMask.clear();
  }
  SmallVector<Value *> StoredGS(GatheredScalars);
  auto FindReusedSplat = [&](MutableArrayRef<int> Mask, unsigned InputVF,
                             unsigned I, unsigned SliceSize,
                             bool IsNotPoisonous) {
    if (!isSplat(E->Scalars) || none_of(E->Scalars, [](Value *V) {
          return isa<UndefValue>(V) && !isa<PoisonValue>(V);
        }))
      return false;
    TreeEntry *UserTE = E->UserTreeIndex.UserTE;
    unsigned EdgeIdx = E->UserTreeIndex.EdgeIdx;
    if (UserTE->getNumOperands() != 2)
      return false;
    if (!IsNotPoisonous) {
      auto *It = find_if(ArrayRef(VectorizableTree).drop_front(UserTE->Idx + 1),
                         [=](const std::unique_ptr<TreeEntry> &TE) {
                           return TE->UserTreeIndex.UserTE == UserTE &&
                                  TE->UserTreeIndex.EdgeIdx != EdgeIdx;
                         });
      if (It == VectorizableTree.end())
        return false;
      SmallVector<Value *> GS((*It)->Scalars.begin(), (*It)->Scalars.end());
      if (!(*It)->ReorderIndices.empty()) {
        inversePermutation((*It)->ReorderIndices, ReorderMask);
        reorderScalars(GS, ReorderMask);
      }
      if (!all_of(zip(GatheredScalars, GS), [&](const auto &P) {
            Value *V0 = std::get<0>(P);
            Value *V1 = std::get<1>(P);
            return !isa<UndefValue>(V0) || isa<PoisonValue>(V0) ||
                   (isa<UndefValue>(V0) && !isa<PoisonValue>(V0) &&
                    is_contained(E->Scalars, V1));
          }))
        return false;
    }
    int Idx;
    if ((Mask.size() < InputVF &&
         ShuffleVectorInst::isExtractSubvectorMask(Mask, InputVF, Idx) &&
         Idx == 0) ||
        (Mask.size() == InputVF &&
         ShuffleVectorInst::isIdentityMask(Mask, Mask.size()))) {
      std::iota(
          std::next(Mask.begin(), I * SliceSize),
          std::next(Mask.begin(),
                    I * SliceSize + getNumElems(Mask.size(), SliceSize, I)),
          0);
    } else {
      unsigned IVal =
          *find_if_not(Mask, [](int Idx) { return Idx == PoisonMaskElem; });
      std::fill(
          std::next(Mask.begin(), I * SliceSize),
          std::next(Mask.begin(),
                    I * SliceSize + getNumElems(Mask.size(), SliceSize, I)),
          IVal);
    }
    return true;
  };
  BVTy ShuffleBuilder(ScalarTy, Params...);
  ResTy Res = ResTy();
  SmallVector<int> Mask;
  SmallVector<int> ExtractMask(GatheredScalars.size(), PoisonMaskElem);
  SmallVector<std::optional<TTI::ShuffleKind>> ExtractShuffles;
  Value *ExtractVecBase = nullptr;
  bool UseVecBaseAsInput = false;
  SmallVector<std::optional<TargetTransformInfo::ShuffleKind>> GatherShuffles;
  SmallVector<SmallVector<const TreeEntry *>> Entries;
  Type *OrigScalarTy = GatheredScalars.front()->getType();
  auto *VecTy = getWidenedType(ScalarTy, GatheredScalars.size());
  unsigned NumParts = getNumberOfParts(VecTy, ScalarTy, GatheredScalars.size());
  if (!all_of(GatheredScalars, IsaPred<UndefValue>)) {
    // Check for gathered extracts.
    bool Resized = false;
    ExtractShuffles =
        tryToGatherExtractElements(GatheredScalars, ExtractMask, NumParts);
    if (!ExtractShuffles.empty()) {
      SmallVector<const TreeEntry *> ExtractEntries;
      for (auto [Idx, I] : enumerate(ExtractMask)) {
        if (I == PoisonMaskElem)
          continue;
        if (ArrayRef<TreeEntry *> TEs = getTreeEntries(
                cast<ExtractElementInst>(StoredGS[Idx])->getVectorOperand());
            !TEs.empty())
          ExtractEntries.append(TEs.begin(), TEs.end());
      }
      if (std::optional<ResTy> Delayed =
              ShuffleBuilder.needToDelay(E, ExtractEntries)) {
        // Delay emission of gathers which are not ready yet.
        PostponedGathers.insert(E);
        // Postpone gather emission, will be emitted after the end of the
        // process to keep correct order.
        return *Delayed;
      }
      if (Value *VecBase = ShuffleBuilder.adjustExtracts(
              E, ExtractMask, ExtractShuffles, NumParts, UseVecBaseAsInput)) {
        ExtractVecBase = VecBase;
        if (auto *VecBaseTy = dyn_cast<FixedVectorType>(VecBase->getType()))
          if (VF == VecBaseTy->getNumElements() &&
              GatheredScalars.size() != VF) {
            Resized = true;
            GatheredScalars.append(VF - GatheredScalars.size(),
                                   PoisonValue::get(OrigScalarTy));
            NumParts = getNumberOfParts(getWidenedType(OrigScalarTy, VF),
                                        OrigScalarTy, VF);
          }
      }
    }
    // Gather extracts after we check for full matched gathers only.
    if (!ExtractShuffles.empty() || !E->hasState() ||
        E->getOpcode() != Instruction::Load ||
        (((E->hasState() && E->getOpcode() == Instruction::Load) ||
          any_of(E->Scalars, IsaPred<LoadInst>)) &&
         any_of(E->Scalars,
                [this](Value *V) {
                  return isa<LoadInst>(V) && isVectorized(V);
                })) ||
        (E->hasState() && E->isAltShuffle()) ||
        all_of(E->Scalars, [this](Value *V) { return isVectorized(V); }) ||
        isSplat(E->Scalars) ||
        (E->Scalars != GatheredScalars && GatheredScalars.size() <= 2)) {
      GatherShuffles =
          isGatherShuffledEntry(E, GatheredScalars, Mask, Entries, NumParts);
    }
    if (!GatherShuffles.empty()) {
      if (std::optional<ResTy> Delayed =
              ShuffleBuilder.needToDelay(E, Entries)) {
        // Delay emission of gathers which are not ready yet.
        PostponedGathers.insert(E);
        // Postpone gather emission, will be emitted after the end of the
        // process to keep correct order.
        return *Delayed;
      }
      // Match against the expanded vector of the gather node, so that poison
      // lanes of the matched entry cannot wildcard-match its real scalars.
      auto IsFullVectorMatch = [&](const TreeEntry *FrontTE) {
        SmallVector<int> CommonMask = E->getCommonMask();
        SmallVector<Value *> Expanded(E->getVectorFactor());
        for (unsigned I : seq<unsigned>(E->getVectorFactor()))
          Expanded[I] = CommonMask.empty() ? E->Scalars[I]
                        : CommonMask[I] == PoisonMaskElem
                            ? PoisonValue::get(OrigScalarTy)
                            : E->Scalars[CommonMask[I]];
        return FrontTE->isSame(Expanded);
      };
      if (GatherShuffles.size() == 1 &&
          *GatherShuffles.front() == TTI::SK_PermuteSingleSrc &&
          (Entries.front().front()->isSame(E->Scalars) ||
           IsFullVectorMatch(Entries.front().front()))) {
        // Perfect match in the graph, will reuse the previously vectorized
        // node. Cost is 0.
        LLVM_DEBUG(dbgs() << "SLP: perfect diamond match for gather bundle "
                          << shortBundleName(E->Scalars, E->Idx) << ".\n");
        // Restore the mask for previous partially matched values.
        Mask.resize(E->Scalars.size());
        const TreeEntry *FrontTE = Entries.front().front();
        if (FrontTE->ReorderIndices.empty() && E->ReorderIndices.empty() &&
            ((FrontTE->ReuseShuffleIndices.empty() &&
              E->Scalars.size() == FrontTE->Scalars.size()) ||
             (E->Scalars.size() == FrontTE->ReuseShuffleIndices.size()))) {
          std::iota(Mask.begin(), Mask.end(), 0);
        } else {
          for (auto [I, V] : enumerate(E->Scalars)) {
            if (isa<PoisonValue>(V)) {
              Mask[I] = PoisonMaskElem;
              continue;
            }
            Mask[I] = FrontTE->findLaneForValue(V);
          }
        }
        // Reset the builder(s) to correctly handle perfect diamond matched
        // nodes.
        ShuffleBuilder.resetForSameNode();
        // Full matched entry found, no need to insert subvectors.
        if ((IsFullVectorMatch(FrontTE) &&
             FrontTE->ReuseShuffleIndices.empty() &&
             FrontTE->ReorderIndices.empty() &&
             E->getVectorFactor() == FrontTE->getVectorFactor()) ||
            (equal(E->Scalars, FrontTE->Scalars) &&
             equal(E->ReorderIndices, FrontTE->ReorderIndices) &&
             equal(E->ReuseShuffleIndices, FrontTE->ReuseShuffleIndices))) {
          Mask.resize(FrontTE->getVectorFactor());
          std::iota(Mask.begin(), Mask.end(), 0);
          ShuffleBuilder.add(*FrontTE, Mask);
          Res = ShuffleBuilder.finalize({}, {}, {});
        } else {
          ShuffleBuilder.add(*FrontTE, Mask);
          Res = ShuffleBuilder.finalize(E->getCommonMask(), {}, {});
        }
        return Res;
      }
      if (!Resized) {
        if (GatheredScalars.size() != VF &&
            any_of(Entries, [&](ArrayRef<const TreeEntry *> TEs) {
              return any_of(TEs, [&](const TreeEntry *TE) {
                return TE->getVectorFactor() == VF;
              });
            }))
          GatheredScalars.append(VF - GatheredScalars.size(),
                                 PoisonValue::get(OrigScalarTy));
      }
      // Remove shuffled elements from list of gathers.
      for (int I = 0, Sz = Mask.size(); I < Sz; ++I) {
        if (Mask[I] != PoisonMaskElem)
          GatheredScalars[I] = PoisonValue::get(OrigScalarTy);
      }
    }
  }
  auto TryPackScalars = [&](SmallVectorImpl<Value *> &Scalars,
                            SmallVectorImpl<int> &ReuseMask,
                            bool IsRootPoison) {
    // For splats with can emit broadcasts instead of gathers, so try to find
    // such sequences.
    bool IsSplat = IsRootPoison && isSplat(Scalars) &&
                   (Scalars.size() > 2 || Scalars.front() == Scalars.back());
    Scalars.append(VF - Scalars.size(), PoisonValue::get(OrigScalarTy));
    SmallVector<int> UndefPos;
    DenseMap<Value *, unsigned> UniquePositions;
    // Gather unique non-const values and all constant values.
    // For repeated values, just shuffle them.
    int NumNonConsts = 0;
    int SinglePos = 0;
    for (auto [I, V] : enumerate(Scalars)) {
      if (isa<UndefValue>(V)) {
        if (!isa<PoisonValue>(V)) {
          ReuseMask[I] = I;
          UndefPos.push_back(I);
        }
        continue;
      }
      if (isConstant(V)) {
        ReuseMask[I] = I;
        continue;
      }
      ++NumNonConsts;
      SinglePos = I;
      Value *OrigV = V;
      Scalars[I] = PoisonValue::get(OrigScalarTy);
      if (IsSplat) {
        Scalars.front() = OrigV;
        ReuseMask[I] = 0;
      } else {
        const auto Res = UniquePositions.try_emplace(OrigV, I);
        Scalars[Res.first->second] = OrigV;
        ReuseMask[I] = Res.first->second;
      }
    }
    if (NumNonConsts == 1) {
      // Restore single insert element.
      if (IsSplat) {
        ReuseMask.assign(VF, PoisonMaskElem);
        std::swap(Scalars.front(), Scalars[SinglePos]);
        if (!UndefPos.empty() && UndefPos.front() == 0)
          Scalars.front() = UndefValue::get(OrigScalarTy);
      }
      ReuseMask[SinglePos] = SinglePos;
    } else if (!UndefPos.empty() && IsSplat) {
      // For undef values, try to replace them with the simple broadcast.
      // We can do it if the broadcasted value is guaranteed to be
      // non-poisonous, or by freezing the incoming scalar value first.
      auto *It = find_if(Scalars, [this, E](Value *V) {
        return !isa<UndefValue>(V) &&
               (isVectorized(V) || isGuaranteedNotToBePoison(V, AC) ||
                (E->UserTreeIndex && any_of(V->uses(), [E](const Use &U) {
                   // Check if the value already used in the same operation in
                   // one of the nodes already.
                   return E->UserTreeIndex.EdgeIdx != U.getOperandNo() &&
                          is_contained(E->UserTreeIndex.UserTE->Scalars,
                                       U.getUser());
                 })));
      });
      if (It != Scalars.end()) {
        // Replace undefs by the non-poisoned scalars and emit broadcast.
        int Pos = std::distance(Scalars.begin(), It);
        for (int I : UndefPos) {
          // Set the undef position to the non-poisoned scalar.
          ReuseMask[I] = Pos;
          // Replace the undef by the poison, in the mask it is replaced by
          // non-poisoned scalar already.
          if (I != Pos)
            Scalars[I] = PoisonValue::get(OrigScalarTy);
        }
      } else {
        // Replace undefs by the poisons, emit broadcast and then emit
        // freeze.
        for (int I : UndefPos) {
          ReuseMask[I] = PoisonMaskElem;
          if (isa<UndefValue>(Scalars[I]))
            Scalars[I] = PoisonValue::get(OrigScalarTy);
        }
        NeedFreeze = true;
      }
    }
  };
  if (!ExtractShuffles.empty() || !GatherShuffles.empty()) {
    bool IsNonPoisoned = true;
    bool IsUsedInExpr = true;
    Value *Vec1 = nullptr;
    if (!ExtractShuffles.empty()) {
      // Gather of extractelements can be represented as just a shuffle of
      // a single/two vectors the scalars are extracted from.
      // Find input vectors.
      Value *Vec2 = nullptr;
      for (unsigned I = 0, Sz = ExtractMask.size(); I < Sz; ++I) {
        if (!Mask.empty() && Mask[I] != PoisonMaskElem)
          ExtractMask[I] = PoisonMaskElem;
      }
      if (UseVecBaseAsInput) {
        Vec1 = ExtractVecBase;
      } else {
        for (unsigned I = 0, Sz = ExtractMask.size(); I < Sz; ++I) {
          if (ExtractMask[I] == PoisonMaskElem)
            continue;
          if (isa<UndefValue>(StoredGS[I]))
            continue;
          auto *EI = cast<ExtractElementInst>(StoredGS[I]);
          Value *VecOp = EI->getVectorOperand();
          if (ArrayRef<TreeEntry *> TEs = getTreeEntries(VecOp);
              !TEs.empty() && TEs.front()->VectorizedValue)
            VecOp = TEs.front()->VectorizedValue;
          if (!Vec1) {
            Vec1 = VecOp;
          } else if (Vec1 != VecOp) {
            assert((!Vec2 || Vec2 == VecOp) &&
                   "Expected only 1 or 2 vectors shuffle.");
            Vec2 = VecOp;
          }
        }
      }
      if (Vec2) {
        IsUsedInExpr = false;
        IsNonPoisoned &= isGuaranteedNotToBePoison(Vec1, AC) &&
                         isGuaranteedNotToBePoison(Vec2, AC);
        ShuffleBuilder.add(Vec1, Vec2, ExtractMask);
      } else if (Vec1) {
        bool IsNotPoisonedVec = isGuaranteedNotToBePoison(Vec1, AC);
        IsUsedInExpr &= FindReusedSplat(
            ExtractMask,
            cast<FixedVectorType>(Vec1->getType())->getNumElements(), 0,
            ExtractMask.size(), IsNotPoisonedVec);
        ShuffleBuilder.add(Vec1, ExtractMask, /*ForExtracts=*/true);
        IsNonPoisoned &= IsNotPoisonedVec;
      } else {
        IsUsedInExpr = false;
        ShuffleBuilder.add(PoisonValue::get(VecTy), ExtractMask,
                           /*ForExtracts=*/true);
      }
    }
    if (!GatherShuffles.empty()) {
      unsigned SliceSize = getPartNumElems(VF, NumParts);
      if (Mask.size() == E->Scalars.size())
        SliceSize = getPartNumElems(
            E->Scalars.size(),
            getNumberOfParts(VecTy, ScalarTy, E->Scalars.size()));
      SmallVector<int> VecMask(Mask.size(), PoisonMaskElem);
      for (const auto [I, TEs] : enumerate(Entries)) {
        if (TEs.empty()) {
          assert(!GatherShuffles[I] &&
                 "No shuffles with empty entries list expected.");
          continue;
        }
        assert((TEs.size() == 1 || TEs.size() == 2) &&
               "Expected shuffle of 1 or 2 entries.");
        unsigned Limit = getNumElems(Mask.size(), SliceSize, I);
        auto SubMask = ArrayRef(Mask).slice(I * SliceSize, Limit);
        VecMask.assign(VecMask.size(), PoisonMaskElem);
        copy(SubMask, std::next(VecMask.begin(), I * SliceSize));
        if (TEs.size() == 1) {
          bool IsNotPoisonedVec =
              TEs.front()->VectorizedValue
                  ? isGuaranteedNotToBePoison(TEs.front()->VectorizedValue, AC)
                  : true;
          IsUsedInExpr &=
              FindReusedSplat(VecMask, TEs.front()->getVectorFactor(), I,
                              SliceSize, IsNotPoisonedVec);
          ShuffleBuilder.add(*TEs.front(), VecMask);
          IsNonPoisoned &= IsNotPoisonedVec;
        } else {
          IsUsedInExpr = false;
          ShuffleBuilder.add(*TEs.front(), *TEs.back(), VecMask);
          if (TEs.front()->VectorizedValue && TEs.back()->VectorizedValue)
            IsNonPoisoned &=
                isGuaranteedNotToBePoison(TEs.front()->VectorizedValue, AC) &&
                isGuaranteedNotToBePoison(TEs.back()->VectorizedValue, AC);
        }
      }
    }
    // Try to figure out best way to combine values: build a shuffle and insert
    // elements or just build several shuffles.
    // Insert non-constant scalars.
    SmallVector<Value *> NonConstants(GatheredScalars);
    int EMSz = ExtractMask.size();
    int MSz = Mask.size();
    // Try to build constant vector and shuffle with it only if currently we
    // have a single permutation and more than 1 scalar constants.
    bool IsSingleShuffle = ExtractShuffles.empty() || GatherShuffles.empty();
    bool IsIdentityShuffle =
        ((UseVecBaseAsInput ||
          all_of(ExtractShuffles,
                 [](const std::optional<TTI::ShuffleKind> &SK) {
                   return SK.value_or(TTI::SK_PermuteTwoSrc) ==
                          TTI::SK_PermuteSingleSrc;
                 })) &&
         none_of(ExtractMask, [&](int I) { return I >= EMSz; }) &&
         ShuffleVectorInst::isIdentityMask(ExtractMask, EMSz)) ||
        (!GatherShuffles.empty() &&
         all_of(GatherShuffles,
                [](const std::optional<TTI::ShuffleKind> &SK) {
                  return SK.value_or(TTI::SK_PermuteTwoSrc) ==
                         TTI::SK_PermuteSingleSrc;
                }) &&
         none_of(Mask, [&](int I) { return I >= MSz; }) &&
         ShuffleVectorInst::isIdentityMask(Mask, MSz));
    bool EnoughConstsForShuffle =
        IsSingleShuffle &&
        (none_of(GatheredScalars,
                 [](Value *V) {
                   return isa<UndefValue>(V) && !isa<PoisonValue>(V);
                 }) ||
         any_of(GatheredScalars,
                [](Value *V) {
                  return isa<Constant>(V) && !isa<UndefValue>(V);
                })) &&
        (!IsIdentityShuffle ||
         (GatheredScalars.size() == 2 &&
          any_of(GatheredScalars,
                 [](Value *V) { return !isa<UndefValue>(V); })) ||
         count_if(GatheredScalars, [](Value *V) {
           return isa<Constant>(V) && !isa<PoisonValue>(V);
         }) > 1);
    // NonConstants array contains just non-constant values, GatheredScalars
    // contains only constant to build final vector and then shuffle.
    for (int I = 0, Sz = GatheredScalars.size(); I < Sz; ++I) {
      if (EnoughConstsForShuffle && isa<Constant>(GatheredScalars[I]))
        NonConstants[I] = PoisonValue::get(OrigScalarTy);
      else
        GatheredScalars[I] = PoisonValue::get(OrigScalarTy);
    }
    // Generate constants for final shuffle and build a mask for them.
    if (!all_of(GatheredScalars, IsaPred<PoisonValue>)) {
      SmallVector<int> BVMask(GatheredScalars.size(), PoisonMaskElem);
      TryPackScalars(GatheredScalars, BVMask, /*IsRootPoison=*/true);
      Value *BV = ShuffleBuilder.gather(GatheredScalars, BVMask.size());
      ShuffleBuilder.add(BV, BVMask);
    }
    if (all_of(NonConstants, [=](Value *V) {
          return isa<PoisonValue>(V) ||
                 (IsSingleShuffle && ((IsIdentityShuffle &&
                  IsNonPoisoned) || IsUsedInExpr) && isa<UndefValue>(V));
        }))
      Res = ShuffleBuilder.finalize(E->ReuseShuffleIndices, SubVectors,
                                    SubVectorsMask);
    else
      Res = ShuffleBuilder.finalize(
          E->ReuseShuffleIndices, SubVectors, SubVectorsMask, E->Scalars.size(),
          [&](Value *&Vec, SmallVectorImpl<int> &Mask, auto CreateShuffle) {
            bool IsSplat = isSplat(NonConstants);
            SmallVector<int> BVMask(Mask.size(), PoisonMaskElem);
            TryPackScalars(NonConstants, BVMask, /*IsRootPoison=*/false);
            auto CheckIfSplatIsProfitable = [&]() {
              // Estimate the cost of splatting + shuffle and compare with
              // insert + shuffle.
              Value *V = *find_if_not(NonConstants, IsaPred<UndefValue>);
              if (isa<ExtractElementInst>(V) || isVectorized(V))
                return false;
              InstructionCost SplatCost = TTI->getVectorInstrCost(
                  Instruction::InsertElement, VecTy, CostKind, /*Index=*/0,
                  PoisonValue::get(VecTy), V);
              SmallVector<int> NewMask(Mask.begin(), Mask.end());
              for (auto [Idx, I] : enumerate(BVMask))
                if (I != PoisonMaskElem)
                  NewMask[Idx] = Mask.size();
              SplatCost +=
                  getShuffleCost(*TTI, TTI::SK_PermuteTwoSrc,
                                 cast<VectorType>(VecTy), CostKind, NewMask);
              InstructionCost BVCost = TTI->getVectorInstrCost(
                  Instruction::InsertElement, VecTy, CostKind,
                  *find_if(Mask, not_equal_to(PoisonMaskElem)), Vec, V);
              // Shuffle required?
              if (count(BVMask, PoisonMaskElem) <
                  static_cast<int>(BVMask.size() - 1)) {
                SmallVector<int> NewMask(Mask.begin(), Mask.end());
                for (auto [Idx, I] : enumerate(BVMask))
                  if (I != PoisonMaskElem)
                    NewMask[Idx] = I;
                BVCost +=
                    getShuffleCost(*TTI, TTI::SK_PermuteSingleSrc,
                                   cast<VectorType>(VecTy), CostKind, NewMask);
              }
              return SplatCost <= BVCost;
            };
            if (!IsSplat || Mask.size() <= 2 || !CheckIfSplatIsProfitable()) {
              for (auto [Idx, I] : enumerate(BVMask))
                if (I != PoisonMaskElem)
                  Mask[Idx] = I;
              Vec = ShuffleBuilder.gather(NonConstants, Mask.size(), Vec);
            } else {
              Value *V = *find_if_not(NonConstants, IsaPred<UndefValue>);
              SmallVector<Value *> Values(NonConstants.size(),
                                          PoisonValue::get(ScalarTy));
              Values[0] = V;
              Value *BV = ShuffleBuilder.gather(Values, BVMask.size());
              SmallVector<int> SplatMask(BVMask.size(), PoisonMaskElem);
              transform(BVMask, SplatMask.begin(), [](int I) {
                return I == PoisonMaskElem ? PoisonMaskElem : 0;
              });
              if (!ShuffleVectorInst::isIdentityMask(SplatMask, VF))
                BV = CreateShuffle(BV, nullptr, SplatMask);
              for (auto [Idx, I] : enumerate(BVMask))
                if (I != PoisonMaskElem)
                  Mask[Idx] = BVMask.size() + Idx;
              Vec = CreateShuffle(Vec, BV, Mask);
              for (auto [Idx, I] : enumerate(Mask))
                if (I != PoisonMaskElem)
                  Mask[Idx] = Idx;
            }
          });
  } else if (!allConstant(GatheredScalars)) {
    // Gather unique scalars and all constants.
    SmallVector<int> ReuseMask(GatheredScalars.size(), PoisonMaskElem);
    TryPackScalars(GatheredScalars, ReuseMask, /*IsRootPoison=*/true);
    Value *BV = ShuffleBuilder.gather(GatheredScalars, ReuseMask.size());
    ShuffleBuilder.add(BV, ReuseMask);
    Res = ShuffleBuilder.finalize(E->ReuseShuffleIndices, SubVectors,
                                  SubVectorsMask);
  } else {
    // Gather all constants.
    SmallVector<int> Mask(GatheredScalars.size(), PoisonMaskElem);
    for (auto [I, V] : enumerate(GatheredScalars)) {
      if (!isa<PoisonValue>(V))
        Mask[I] = I;
    }
    Value *BV = ShuffleBuilder.gather(GatheredScalars);
    ShuffleBuilder.add(BV, Mask);
    Res = ShuffleBuilder.finalize(E->ReuseShuffleIndices, SubVectors,
                                  SubVectorsMask);
  }
 
  if (NeedFreeze)
    Res = ShuffleBuilder.createFreeze(Res);
  return Res;
}
 
Value *BoUpSLP::createBuildVector(const TreeEntry *E, Type *ScalarTy) {
  // Do not do this for split vectorize node, marked to be gathers/buildvectors.
  if (E->State != TreeEntry::SplitVectorize ||
      !TransformedToGatherNodes.contains(E)) {
    for (auto [EIdx, _] : E->CombinedEntriesWithIndices)
      (void)vectorizeTree(VectorizableTree[EIdx].get());
  }
  return processBuildVector<ShuffleInstructionBuilder, Value *>(E, ScalarTy,
                                                                Builder, *this);
}
 
/// \returns \p I after propagating metadata from \p VL only for instructions in
/// \p VL.
static Instruction *propagateMetadata(Instruction *Inst, ArrayRef<Value *> VL) {
  SmallVector<Value *> Insts;
  for (Value *V : VL)
    if (isa<Instruction>(V))
      Insts.push_back(V);
  return llvm::propagateMetadata(Inst, Insts);
}
 
static DebugLoc getDebugLocFromPHI(PHINode &PN) {
  if (DebugLoc DL = PN.getDebugLoc())
    return DL;
  return DebugLoc::getUnknown();
}
 
Value *BoUpSLP::vectorizeTree(TreeEntry *E) {
  IRBuilderBase::InsertPointGuard Guard(Builder);
 
  Value *V = E->Scalars.front();
  Type *ScalarTy = getValueType(V, SLPReVec);
  auto It = MinBWs.find(E);
  if (It != MinBWs.end()) {
    auto *VecTy = dyn_cast<FixedVectorType>(ScalarTy);
    ScalarTy = IntegerType::get(F->getContext(), It->second.first);
    if (VecTy)
      ScalarTy = getWidenedType(ScalarTy, VecTy->getNumElements());
  }
  if (E->VectorizedValue)
    return E->VectorizedValue;
  auto *VecTy = getWidenedType(ScalarTy, E->Scalars.size());
  if (E->isGather() || TransformedToGatherNodes.contains(E)) {
    // Set insert point for non-reduction initial nodes.
    if (E->hasState() && E->Idx == 0 && !UserIgnoreList)
      setInsertPointAfterBundle(E);
    Value *Vec = createBuildVector(E, ScalarTy);
    E->VectorizedValue = Vec;
    return Vec;
  }
  if (E->State == TreeEntry::SplitVectorize) {
    assert(E->CombinedEntriesWithIndices.size() == 2 &&
           "Expected exactly 2 combined entries.");
    setInsertPointAfterBundle(E);
    TreeEntry &OpTE1 =
        *VectorizableTree[E->CombinedEntriesWithIndices.front().first];
    assert(OpTE1.isSame(
               ArrayRef(E->Scalars).take_front(OpTE1.getVectorFactor())) &&
           "Expected same first part of scalars.");
    Value *Op1 = vectorizeTree(&OpTE1);
    TreeEntry &OpTE2 =
        *VectorizableTree[E->CombinedEntriesWithIndices.back().first];
    assert(
        OpTE2.isSame(ArrayRef(E->Scalars).take_back(OpTE2.getVectorFactor())) &&
        "Expected same second part of scalars.");
    Value *Op2 = vectorizeTree(&OpTE2);
    auto GetOperandSignedness = [&](const TreeEntry *OpE) {
      bool IsSigned = false;
      auto It = MinBWs.find(OpE);
      if (It != MinBWs.end())
        IsSigned = It->second.second;
      else
        IsSigned = any_of(OpE->Scalars, [&](Value *R) {
          if (isa<PoisonValue>(V))
            return false;
          return !isKnownNonNegative(R, SimplifyQuery(*DL));
        });
      return IsSigned;
    };
    if (cast<VectorType>(Op1->getType())->getElementType() !=
        ScalarTy->getScalarType()) {
      assert(ScalarTy->isIntegerTy() && "Expected item in MinBWs.");
      Op1 = Builder.CreateIntCast(
          Op1,
          getWidenedType(
              ScalarTy,
              cast<FixedVectorType>(Op1->getType())->getNumElements()),
          GetOperandSignedness(&OpTE1));
    }
    if (cast<VectorType>(Op2->getType())->getElementType() !=
        ScalarTy->getScalarType()) {
      assert(ScalarTy->isIntegerTy() && "Expected item in MinBWs.");
      Op2 = Builder.CreateIntCast(
          Op2,
          getWidenedType(
              ScalarTy,
              cast<FixedVectorType>(Op2->getType())->getNumElements()),
          GetOperandSignedness(&OpTE2));
    }
    if (E->ReorderIndices.empty()) {
      SmallVector<int> Mask(E->getVectorFactor(), PoisonMaskElem);
      std::iota(
          Mask.begin(),
          std::next(Mask.begin(), E->CombinedEntriesWithIndices.back().second),
          0);
      unsigned ScalarTyNumElements = getNumElements(ScalarTy);
      if (ScalarTyNumElements != 1) {
        assert(SLPReVec && "Only supported by REVEC.");
        transformScalarShuffleIndiciesToVector(ScalarTyNumElements, Mask);
      }
      Value *Vec = Builder.CreateShuffleVector(Op1, Mask);
      Vec = createInsertVector(Builder, Vec, Op2,
                               E->CombinedEntriesWithIndices.back().second *
                                   ScalarTyNumElements);
      E->VectorizedValue = Vec;
      return Vec;
    }
    unsigned CommonVF =
        std::max(OpTE1.getVectorFactor(), OpTE2.getVectorFactor());
    const unsigned Scale = getNumElements(ScalarTy);
    CommonVF *= Scale;
    if (getNumElements(Op1->getType()) != CommonVF) {
      SmallVector<int> Mask(CommonVF, PoisonMaskElem);
      copy(createReplicatedMask(Scale, OpTE1.getVectorFactor() * Scale),
           Mask.begin());
      Op1 = Builder.CreateShuffleVector(Op1, Mask);
    }
    if (getNumElements(Op2->getType()) != CommonVF) {
      SmallVector<int> Mask(CommonVF, PoisonMaskElem);
      copy(createReplicatedMask(Scale, OpTE2.getVectorFactor() * Scale),
           Mask.begin());
      Op2 = Builder.CreateShuffleVector(Op2, Mask);
    }
    Value *Vec = Builder.CreateShuffleVector(Op1, Op2, E->getSplitMask());
    E->VectorizedValue = Vec;
    return Vec;
  }
 
  bool IsReverseOrder =
      !E->ReorderIndices.empty() && isReverseOrder(E->ReorderIndices);
  auto FinalShuffle = [&](Value *V, const TreeEntry *E) {
    if (isa<StructType>(ScalarTy)) {
      // TODO: Reordering of struct types is not supported.
      assert(E->ReorderIndices.empty() &&
             "Expected no reordering for struct types.");
      assert(E->ReuseShuffleIndices.empty() &&
             "Expected no reuse shuffle indices for struct types.");
      return V;
    }
    ShuffleInstructionBuilder ShuffleBuilder(ScalarTy, Builder, *this);
    if ((E->State == TreeEntry::StridedVectorize && IsReverseOrder) ||
        E->State == TreeEntry::CompressVectorize) {
      ShuffleBuilder.addOrdered(V, {});
    } else if (E->getOpcode() == Instruction::Store &&
               (E->State == TreeEntry::Vectorize ||
                E->State == TreeEntry::StridedVectorize ||
                E->State == TreeEntry::ExpandVectorize)) {
      ArrayRef<int> Mask =
          ArrayRef(reinterpret_cast<const int *>(E->ReorderIndices.begin()),
                   E->ReorderIndices.size());
      ShuffleBuilder.add(V, Mask);
    } else {
      ShuffleBuilder.addOrdered(V, E->ReorderIndices);
    }
    SmallVector<std::pair<const TreeEntry *, unsigned>> SubVectors(
        E->CombinedEntriesWithIndices.size());
    transform(
        E->CombinedEntriesWithIndices, SubVectors.begin(), [&](const auto &P) {
          return std::make_pair(VectorizableTree[P.first].get(), P.second);
        });
    assert(
        (E->CombinedEntriesWithIndices.empty() || E->ReorderIndices.empty()) &&
        "Expected either combined subnodes or reordering");
    return ShuffleBuilder.finalize(E->ReuseShuffleIndices, SubVectors, {});
  };
 
  assert(!E->isGather() && "Unhandled state");
  unsigned ShuffleOrOp =
      E->isAltShuffle() ? (unsigned)Instruction::ShuffleVector : E->getOpcode();
  if (!E->isAltShuffle()) {
    switch (E->CombinedOp) {
    case TreeEntry::ReducedBitcast:
    case TreeEntry::ReducedBitcastBSwap:
    case TreeEntry::ReducedBitcastLoads:
    case TreeEntry::ReducedBitcastBSwapLoads:
    case TreeEntry::ReducedCmpBitcast:
      ShuffleOrOp = E->CombinedOp;
      break;
    default:
      break;
    }
  }
  Instruction *VL0 = E->getMainOp();
  auto GetOperandSignedness = [&](unsigned Idx) {
    const TreeEntry *OpE = getOperandEntry(E, Idx);
    bool IsSigned = false;
    auto It = MinBWs.find(OpE);
    if (It != MinBWs.end())
      IsSigned = It->second.second;
    else
      IsSigned = any_of(OpE->Scalars, [&](Value *R) {
        if (isa<PoisonValue>(V))
          return false;
        return !isKnownNonNegative(R, SimplifyQuery(*DL));
      });
    return IsSigned;
  };
  auto PropagateIRFlags = [&](Value *V, unsigned Opcode = 0,
                              ArrayRef<Value *> VL = {}) -> Value * {
    ArrayRef<Value *> Scalars = VL.empty() ? E->Scalars : VL;
    SmallSetVector<Value *, 4> UniqueInsts;
    for (Value *Scalar : Scalars) {
      auto *I = dyn_cast<Instruction>(Scalar);
      if (!I)
        continue;
      if (E->hasCopyableElements() && E->isCopyableElement(I))
        continue;
      UniqueInsts.insert(I);
    }
    if (!Opcode)
      Opcode = E->getOpcode();
    propagateIRFlags(V, UniqueInsts.getArrayRef(), nullptr, It == MinBWs.end());
    auto *I = dyn_cast<Instruction>(V);
    if (!I)
      return V;
    I = ::propagateMetadata(I, UniqueInsts.getArrayRef());
    // For copyable elements the lane is synthesized using a binop identity
    // value, so the operand at that lane is the copyable scalar's value.
    // fast-math flags that turn defined NaN/Inf inputs into poison (nnan,
    // ninf) are only justified if every copyable scalar at the synthesized
    // lane is itself provably non-NaN/non-Inf - either an FPMathOperator
    // with the matching flag set, or a constant FP that is not NaN/Inf.
    if (E->hasCopyableElements() && isa<FPMathOperator>(I)) {
      bool AllNoNaNs = true;
      bool AllNoInfs = true;
      for (Value *Scalar : Scalars) {
        if (!E->isCopyableElement(Scalar))
          continue;
        if (auto *FPMO = dyn_cast<FPMathOperator>(Scalar)) {
          AllNoNaNs &= FPMO->hasNoNaNs();
          AllNoInfs &= FPMO->hasNoInfs();
          continue;
        }
        if (auto *CFP = dyn_cast<ConstantFP>(Scalar)) {
          AllNoNaNs &= !CFP->isNaN();
          AllNoInfs &= !CFP->isInfinity();
          continue;
        }
        AllNoNaNs = false;
        AllNoInfs = false;
        break;
      }
      if (!AllNoNaNs)
        I->setHasNoNaNs(false);
      if (!AllNoInfs)
        I->setHasNoInfs(false);
    }
    // Drop nuw flags for abs(sub(commutative), true).
    if (!MinBWs.contains(E) && Opcode == Instruction::Sub &&
        (E->hasCopyableElements() || any_of(Scalars, [](Value *Scalar) {
           auto *SI = dyn_cast<Instruction>(Scalar);
           return !SI || isCommutative(SI);
         })))
      I->setHasNoUnsignedWrap(/*b=*/false);
    // A sub feeding icmp eq/ne 0 may have its operands swapped; nsw does not
    // survive a - b -> b - a (a - b can be INT_MIN while b - a overflows).
    if (!MinBWs.contains(E) && Opcode == Instruction::Sub &&
        any_of(Scalars, [](Value *Scalar) {
          auto *SI = dyn_cast<Instruction>(Scalar);
          if (!SI || SI->getOpcode() != Instruction::Sub || !isCommutative(SI))
            return false;
          return any_of(SI->uses(), [](const Use &U) {
            CmpPredicate Pred;
            return match(U.getUser(),
                         m_ICmp(Pred, m_Specific(U.get()), m_Zero())) &&
                   ICmpInst::isEquality(Pred);
          });
        }))
      I->setHasNoSignedWrap(/*b=*/false);
    // Interchanging add/sub negates the constant: nsw only survives if the
    // constant isn't INT_MIN (negating it would overflow); nuw never
    // survives a nonzero constant, since that flips the valid range from
    // X >= C to X < C.
    if (!MinBWs.contains(E) &&
        (Opcode == Instruction::Add || Opcode == Instruction::Sub)) {
      for (Value *V : UniqueInsts) {
        auto *SI = cast<Instruction>(V);
        if (SI->getOpcode() == Opcode ||
            !is_contained({Instruction::Add, Instruction::Sub},
                          SI->getOpcode()))
          continue;
        for (Value *Op : SI->operands()) {
          const auto *CI = dyn_cast<ConstantInt>(Op);
          if (!CI || CI->isZero())
            continue;
          I->setHasNoUnsignedWrap(/*b=*/false);
          if (CI->getValue().isMinSignedValue())
            I->setHasNoSignedWrap(/*b=*/false);
        }
      }
    }
    // mul nsw X, INT_MIN is not equivalent to shl nsw X, BW-1, because shl nsw
    // poisons when the sign bit is shifted out of a positive value. When a mul
    // lane is converted to shl with shift amount BW-1, nsw must be dropped.
    if (!MinBWs.contains(E) && Opcode == Instruction::Shl &&
        any_of(UniqueInsts, [&](Value *V) {
          auto *SI = cast<Instruction>(V);
          if (SI->getOpcode() != Instruction::Mul)
            return false;
          return any_of(SI->operands(), [](Value *Op) {
            const auto *CI = dyn_cast<ConstantInt>(Op);
            return CI && CI->getValue().isMinSignedValue();
          });
        }))
      I->setHasNoSignedWrap(/*b=*/false);
    // shl nsw X, BW-1 is not equivalent to mul nsw X, INT_MIN: shl nsw never
    // poisons at X = -1 (shifted-out bits match the sign bit), but mul nsw
    // does. A shl lane converted to mul this way must drop nsw.
    if (!MinBWs.contains(E) && Opcode == Instruction::Mul &&
        any_of(UniqueInsts, [&](Value *V) {
          auto *SI = cast<Instruction>(V);
          if (SI->getOpcode() != Instruction::Shl)
            return false;
          const auto *CI = dyn_cast<ConstantInt>(SI->getOperand(1));
          return CI && CI->getValue() == CI->getBitWidth() - 1;
        }))
      I->setHasNoSignedWrap(/*b=*/false);
    // Keep the intersected samesign unless narrowing changed the sign of some
    // compared operand (it neither sign-extends from the narrow type nor has
    // a known-zero narrow sign bit), or a converted lane's adjusted constant
    // flips the sign.
    if (auto *ICmp = dyn_cast<ICmpInst>(I); ICmp && It == MinBWs.end()) {
      bool SignChange =
          ICmp->getOperand(0)->getType()->getScalarType() !=
              VL0->getOperand(0)->getType()->getScalarType() &&
          ((!GetOperandSignedness(0) &&
            !isKnownNonNegative(ICmp->getOperand(0), SimplifyQuery(*DL))) ||
           (!GetOperandSignedness(1) &&
            !isKnownNonNegative(ICmp->getOperand(1), SimplifyQuery(*DL))));
      CmpInst::Predicate P0 = cast<CmpInst>(E->getMainOp())->getPredicate();
      bool SignFlip = !SignChange && any_of(E->Scalars, [&](Value *Scalar) {
        auto *LaneCI = dyn_cast<ICmpInst>(Scalar);
        if (!LaneCI)
          return false;
        auto *OrigC = dyn_cast<ConstantInt>(LaneCI->getOperand(1));
        if (!OrigC)
          return false;
        ConstantInt *AdjC =
            CmpSamePredicateHelper::getAdjustedConstant(LaneCI, P0);
        return AdjC && AdjC->isNegative() != OrigC->isNegative();
      });
      if (SignChange || SignFlip)
        ICmp->setSameSign(/*B=*/false);
    }
    return I;
  };
  switch (ShuffleOrOp) {
    case Instruction::PHI: {
      assert((E->ReorderIndices.empty() || !E->ReuseShuffleIndices.empty() ||
              E != &getRootNode() || E->UserTreeIndex) &&
             "PHI reordering is free.");
      auto *PH = cast<PHINode>(VL0);
      Builder.SetInsertPoint(PH->getParent(),
                             PH->getParent()->getFirstNonPHIIt());
      Builder.SetCurrentDebugLocation(getDebugLocFromPHI(*PH));
      PHINode *NewPhi = Builder.CreatePHI(VecTy, PH->getNumIncomingValues());
      Value *V = NewPhi;
 
      // Adjust insertion point once all PHI's have been generated.
      Builder.SetInsertPoint(PH->getParent(),
                             PH->getParent()->getFirstInsertionPt());
      Builder.SetCurrentDebugLocation(getDebugLocFromPHI(*PH));
 
      V = FinalShuffle(V, E);
 
      E->VectorizedValue = V;
      // If phi node is fully emitted - exit.
      if (NewPhi->getNumIncomingValues() != 0)
        return NewPhi;
 
      // PHINodes may have multiple entries from the same block. We want to
      // visit every block once.
      SmallDenseMap<BasicBlock *, unsigned, 4> VisitedBBs;
      for (unsigned I : seq<unsigned>(PH->getNumIncomingValues())) {
        BasicBlock *IBB = PH->getIncomingBlock(I);
 
        // Stop emission if all incoming values are generated.
        if (NewPhi->getNumIncomingValues() == PH->getNumIncomingValues()) {
          LLVM_DEBUG(dbgs() << "SLP: Diamond merged for " << *VL0 << ".\n");
          return NewPhi;
        }
 
        auto Res = VisitedBBs.try_emplace(IBB, I);
        if (!Res.second) {
          TreeEntry *OpTE = getOperandEntry(E, I);
          if (OpTE->isGather() || DeletedNodes.contains(OpTE) ||
              TransformedToGatherNodes.contains(OpTE)) {
            Value *VecOp = NewPhi->getIncomingValue(Res.first->getSecond());
            NewPhi->addIncoming(VecOp, IBB);
            assert(!OpTE->VectorizedValue && "Expected no vectorized value.");
            OpTE->VectorizedValue = VecOp;
            continue;
          }
        }
 
        Builder.SetInsertPoint(IBB->getTerminator());
        Builder.SetCurrentDebugLocation(getDebugLocFromPHI(*PH));
        Value *Vec = vectorizeOperand(E, I);
        if (VecTy != Vec->getType()) {
          assert((It != MinBWs.end() || getOperandEntry(E, I)->isGather() ||
                  MinBWs.contains(getOperandEntry(E, I))) &&
                 "Expected item in MinBWs.");
          Vec = Builder.CreateIntCast(Vec, VecTy, GetOperandSignedness(I));
        }
        NewPhi->addIncoming(Vec, IBB);
      }
 
      assert(NewPhi->getNumIncomingValues() == PH->getNumIncomingValues() &&
             "Invalid number of incoming values");
      assert(E->VectorizedValue && "Expected vectorized value.");
      return E->VectorizedValue;
    }
 
    case Instruction::ExtractElement: {
      Value *V = E->getSingleOperand(0);
      setInsertPointAfterBundle(E);
      V = FinalShuffle(V, E);
      // Insert the copyable lanes (non-extract scalars modeled on the extract
      // main op) into the reused source vector. The identity extract order and
      // absence of a reuse shuffle are guaranteed when the node is created, so
      // a lane maps to its own index.
      if (E->hasCopyableElements()) {
        assert(E->ReorderIndices.empty() && E->ReuseShuffleIndices.empty() &&
               "Copyable extract lanes require identity order and no reuse.");
        for (auto [Idx, Scalar] : enumerate(E->Scalars))
          if (E->isCopyableElement(Scalar))
            V = Builder.CreateInsertElement(V, Scalar, Builder.getInt32(Idx));
      }
      E->VectorizedValue = V;
      return V;
    }
    case Instruction::ExtractValue: {
      if (!E->StructEVIndices.empty()) {
        setInsertPointAfterBundle(E);
        Value *V = vectorizeOperand(E, 0);
        V = Builder.CreateExtractValue(V, E->StructEVIndices);
        if (auto *I = dyn_cast<Instruction>(V))
          V = ::propagateMetadata(I, E->Scalars);
        V = FinalShuffle(V, E);
        E->VectorizedValue = V;
        return V;
      }
      auto *LI = cast<LoadInst>(E->getSingleOperand(0));
      Builder.SetInsertPoint(LI);
      Value *Ptr = LI->getPointerOperand();
      LoadInst *V = Builder.CreateAlignedLoad(VecTy, Ptr, LI->getAlign());
      Value *NewV = PropagateIRFlags(V);
      NewV = FinalShuffle(NewV, E);
      E->VectorizedValue = NewV;
      return NewV;
    }
    case Instruction::InsertValue:
    case Instruction::InsertElement: {
      assert(E->ReuseShuffleIndices.empty() && "All inserts should be unique");
      if (const TreeEntry *OpE = getOperandEntry(E, 1);
          OpE && !OpE->isGather() && OpE->hasState() &&
          !OpE->hasCopyableElements())
        Builder.SetInsertPoint(cast<Instruction>(E->Scalars.back()));
      else
        setInsertPointAfterBundle(E);
      Value *V = vectorizeOperand(E, 1);
      ArrayRef<Value *> Op = E->getOperand(1);
      Type *ScalarTy = Op.front()->getType();
      if (cast<VectorType>(V->getType())->getElementType() != ScalarTy) {
        assert(ScalarTy->isIntegerTy() && "Expected item in MinBWs.");
        std::pair<unsigned, bool> Res = MinBWs.lookup(getOperandEntry(E, 1));
        assert(Res.first > 0 && "Expected item in MinBWs.");
        V = Builder.CreateIntCast(
            V,
            getWidenedType(
                ScalarTy,
                cast<FixedVectorType>(V->getType())->getNumElements()),
            Res.second);
      }
 
      // Create InsertVector shuffle if necessary
      auto *FirstInsert = cast<Instruction>(*find_if(E->Scalars, [E](Value *V) {
        return !is_contained(E->Scalars, cast<Instruction>(V)->getOperand(0));
      }));
      const unsigned NumScalars = E->Scalars.size();
      FixedVectorType *SrcVecTy = getInsertBuildVectorSrcTy(E);
      const unsigned NumElts = getNumElements(SrcVecTy);
 
      unsigned Offset = *getElementIndex(VL0);
      assert(Offset < NumElts && "Failed to find vector index offset");
 
      // Create shuffle to resize vector
      SmallVector<int> Mask;
      if (!E->ReorderIndices.empty()) {
        inversePermutation(E->ReorderIndices, Mask);
        Mask.append(NumElts - NumScalars, PoisonMaskElem);
      } else {
        Mask.assign(NumElts, PoisonMaskElem);
        std::iota(Mask.begin(), std::next(Mask.begin(), NumScalars), 0);
      }
      // Create InsertVector shuffle if necessary
      bool IsIdentity = true;
      SmallVector<int> PrevMask(NumElts, PoisonMaskElem);
      Mask.swap(PrevMask);
      for (unsigned I = 0; I < NumScalars; ++I) {
        Value *Scalar = E->Scalars[PrevMask[I]];
        unsigned InsertIdx = *getElementIndex(Scalar);
        IsIdentity &= InsertIdx - Offset == I;
        Mask[InsertIdx - Offset] = I;
      }
      if (!IsIdentity || NumElts != NumScalars) {
        Value *V2 = nullptr;
        bool IsVNonPoisonous =
            !isConstant(V) && isGuaranteedNotToBePoison(V, AC);
        SmallVector<int> InsertMask(Mask);
        if (NumElts != NumScalars && Offset == 0) {
          // Follow all insert element instructions from the current buildvector
          // sequence.
          Instruction *Ins = VL0;
          Instruction *Op;
          do {
            std::optional<unsigned> InsertIdx = getElementIndex(Ins);
            if (!InsertIdx)
              break;
            if (InsertMask[*InsertIdx] == PoisonMaskElem)
              InsertMask[*InsertIdx] = *InsertIdx;
            if (!Ins->hasOneUse())
              break;
            Op = Ins;
            Ins =
                dyn_cast_or_null<Instruction>(Ins->getUniqueUndroppableUser());
          } while (Ins && Ins->getOperand(0) == Op);
          SmallBitVector UseMask =
              buildUseMask(NumElts, InsertMask, UseMask::UndefsAsMask);
          SmallBitVector IsFirstPoison =
              isUndefVector<true>(FirstInsert->getOperand(0), UseMask);
          SmallBitVector IsFirstUndef =
              isUndefVector(FirstInsert->getOperand(0), UseMask);
          if (!IsFirstPoison.all()) {
            unsigned Idx = 0;
            for (unsigned I = 0; I < NumElts; I++) {
              if (InsertMask[I] == PoisonMaskElem && !IsFirstPoison.test(I) &&
                  IsFirstUndef.test(I)) {
                if (IsVNonPoisonous) {
                  InsertMask[I] = I < NumScalars ? I : 0;
                  continue;
                }
                if (!V2)
                  V2 = UndefValue::get(V->getType());
                if (Idx >= NumScalars)
                  Idx = NumScalars - 1;
                InsertMask[I] = NumScalars + Idx;
                ++Idx;
              } else if (InsertMask[I] != PoisonMaskElem &&
                         Mask[I] == PoisonMaskElem) {
                InsertMask[I] = PoisonMaskElem;
              }
            }
          } else {
            InsertMask = Mask;
          }
        }
        if (!V2)
          V2 = PoisonValue::get(V->getType());
        V = Builder.CreateShuffleVector(V, V2, InsertMask);
        if (auto *I = dyn_cast<Instruction>(V)) {
          GatherShuffleExtractSeq.insert(I);
          CSEBlocks.insert(I->getParent());
        }
      }
 
      SmallVector<int> InsertMask(NumElts, PoisonMaskElem);
      for (unsigned I = 0; I < NumElts; I++) {
        if (Mask[I] != PoisonMaskElem)
          InsertMask[Offset + I] = I;
      }
      SmallBitVector UseMask =
          buildUseMask(NumElts, InsertMask, UseMask::UndefsAsMask);
      SmallBitVector IsFirstUndef =
          isUndefVector(FirstInsert->getOperand(0), UseMask);
      if ((!IsIdentity || Offset != 0 || !IsFirstUndef.all()) &&
          NumElts != NumScalars) {
        if (IsFirstUndef.all()) {
          if (!ShuffleVectorInst::isIdentityMask(InsertMask, NumElts)) {
            SmallBitVector IsFirstPoison =
                isUndefVector<true>(FirstInsert->getOperand(0), UseMask);
            if (!IsFirstPoison.all()) {
              for (unsigned I = 0; I < NumElts; I++) {
                if (InsertMask[I] == PoisonMaskElem && !IsFirstPoison.test(I))
                  InsertMask[I] = I + NumElts;
              }
            }
            V = Builder.CreateShuffleVector(
                V,
                IsFirstPoison.all() ? PoisonValue::get(V->getType())
                                    : FirstInsert->getOperand(0),
                InsertMask, cast<Instruction>(E->Scalars.back())->getName());
            if (auto *I = dyn_cast<Instruction>(V)) {
              GatherShuffleExtractSeq.insert(I);
              CSEBlocks.insert(I->getParent());
            }
          }
        } else {
          SmallBitVector IsFirstPoison =
              isUndefVector<true>(FirstInsert->getOperand(0), UseMask);
          for (unsigned I = 0; I < NumElts; I++) {
            if (InsertMask[I] == PoisonMaskElem)
              InsertMask[I] = IsFirstPoison.test(I) ? PoisonMaskElem : I;
            else
              InsertMask[I] += NumElts;
          }
          V = Builder.CreateShuffleVector(
              FirstInsert->getOperand(0), V, InsertMask,
              cast<Instruction>(E->Scalars.back())->getName());
          if (auto *I = dyn_cast<Instruction>(V)) {
            GatherShuffleExtractSeq.insert(I);
            CSEBlocks.insert(I->getParent());
          }
        }
      }
 
      ++NumVectorInstructions;
      if (ShuffleOrOp == Instruction::InsertValue &&
          !canVectorStoreInsertValue(E)) {
        Type *AggTy = E->getMainOp()->getType();
        Align SlotAlign = std::max(DL->getPrefTypeAlign(V->getType()),
                                   DL->getPrefTypeAlign(AggTy));
        AllocaInst *Slot;
        {
          // Keep the alloca in the function entry block so it is not executed
          // (and does not grow the stack) on every iteration of a loop.
          IRBuilderBase::InsertPointGuard Guard(Builder);
          Builder.SetInsertPointPastAllocas(F);
          Slot = Builder.CreateAlloca(AggTy, /*ArraySize=*/nullptr,
                                      "vec2struct.slot");
          Slot->setAlignment(SlotAlign);
        }
        (void)Builder.CreateAlignedStore(V, Slot, SlotAlign);
        V = Builder.CreateAlignedLoad(AggTy, Slot, SlotAlign, "vec2struct");
        NumVectorInstructions += 2;
      }
      E->VectorizedValue = V;
      return V;
    }
    case Instruction::ZExt:
    case Instruction::SExt:
    case Instruction::FPToUI:
    case Instruction::FPToSI:
    case Instruction::FPExt:
    case Instruction::PtrToInt:
    case Instruction::PtrToAddr:
    case Instruction::IntToPtr:
    case Instruction::SIToFP:
    case Instruction::UIToFP:
    case Instruction::Trunc:
    case Instruction::FPTrunc:
    case Instruction::BitCast: {
      setInsertPointAfterBundle(E);
 
      Value *InVec = vectorizeOperand(E, 0);
 
      auto *CI = cast<CastInst>(VL0);
      Instruction::CastOps VecOpcode = CI->getOpcode();
      Type *SrcScalarTy = cast<VectorType>(InVec->getType())->getElementType();
      auto SrcIt = MinBWs.find(getOperandEntry(E, 0));
      if (!ScalarTy->isFPOrFPVectorTy() && !SrcScalarTy->isFPOrFPVectorTy() &&
          (SrcIt != MinBWs.end() || It != MinBWs.end() ||
           SrcScalarTy != CI->getOperand(0)->getType()->getScalarType())) {
        // Check if the values are candidates to demote.
        unsigned SrcBWSz = DL->getTypeSizeInBits(SrcScalarTy);
        if (SrcIt != MinBWs.end())
          SrcBWSz = SrcIt->second.first;
        unsigned BWSz = DL->getTypeSizeInBits(ScalarTy->getScalarType());
        if (BWSz == SrcBWSz) {
          VecOpcode = Instruction::BitCast;
        } else if (BWSz < SrcBWSz) {
          VecOpcode = Instruction::Trunc;
        } else if (It != MinBWs.end()) {
          assert(BWSz > SrcBWSz && "Invalid cast!");
          VecOpcode = It->second.second ? Instruction::SExt : Instruction::ZExt;
        } else if (SrcIt != MinBWs.end()) {
          assert(BWSz > SrcBWSz && "Invalid cast!");
          VecOpcode =
              SrcIt->second.second ? Instruction::SExt : Instruction::ZExt;
        }
      } else if (VecOpcode == Instruction::SIToFP && SrcIt != MinBWs.end() &&
                 !SrcIt->second.second) {
        VecOpcode = Instruction::UIToFP;
      } else if (VecOpcode == Instruction::BitCast && SrcIt != MinBWs.end() &&
                 ScalarTy->isFPOrFPVectorTy()) {
        Type *OrigSrcScalarTy = CI->getSrcTy();
        auto *OrigSrcVectorTy =
            getWidenedType(OrigSrcScalarTy, E->Scalars.size());
        InVec =
            Builder.CreateIntCast(InVec, OrigSrcVectorTy, SrcIt->second.second);
      }
      Value *V = (VecOpcode != ShuffleOrOp && VecOpcode == Instruction::BitCast)
                     ? InVec
                     : Builder.CreateCast(VecOpcode, InVec, VecTy);
      V = FinalShuffle(V, E);
 
      E->VectorizedValue = V;
      ++NumVectorInstructions;
      return V;
    }
    case Instruction::FCmp:
    case Instruction::ICmp: {
      setInsertPointAfterBundle(E);
 
      Value *L = vectorizeOperand(E, 0);
      Value *R = vectorizeOperand(E, 1);
      if (L->getType() != R->getType()) {
        assert((getOperandEntry(E, 0)->isGather() ||
                getOperandEntry(E, 1)->isGather() ||
                MinBWs.contains(getOperandEntry(E, 0)) ||
                MinBWs.contains(getOperandEntry(E, 1))) &&
               "Expected item in MinBWs.");
        const unsigned LBW = cast<VectorType>(L->getType())
                                 ->getElementType()
                                 ->getIntegerBitWidth();
        const unsigned RBW = cast<VectorType>(R->getType())
                                 ->getElementType()
                                 ->getIntegerBitWidth();
        // A signed comparison needs the sign bit getActiveBits() ignores.
        const bool Signed = cast<CmpInst>(VL0)->isSigned();
        auto GetRequiredBits = [Signed](const APInt &V) {
          return Signed ? V.getSignificantBits() : V.getActiveBits();
        };
        if ((LBW < RBW && (!allConstant(E->getOperand(1)) ||
                           any_of(
                               E->getOperand(1),
                               [&](Value *V) {
                                 auto *CI = dyn_cast<ConstantInt>(V);
                                 return !CI ||
                                        GetRequiredBits(CI->getValue()) > LBW;
                               }))) ||
            (LBW > RBW && allConstant(E->getOperand(0)) &&
             all_of(E->getOperand(1), [&](Value *V) {
               auto *CI = dyn_cast<ConstantInt>(V);
               return CI && GetRequiredBits(CI->getValue()) <= RBW;
             }))) {
          Type *CastTy = R->getType();
          L = Builder.CreateIntCast(L, CastTy, GetOperandSignedness(0));
        } else {
          Type *CastTy = L->getType();
          R = Builder.CreateIntCast(R, CastTy, GetOperandSignedness(1));
        }
      }
 
      CmpInst::Predicate P0 = cast<CmpInst>(VL0)->getPredicate();
      // A compare with both operands narrowed to the same type preserves the
      // result only if the predicate matches the operand signedness: signed
      // predicates need both operands sign-extended from the narrow type,
      // unsigned predicates need both zero-extended, equality needs the same
      // extension on both sides; operands with a known-zero narrow sign bit
      // match any predicate. Otherwise, extend the operands back to the
      // original compared type.
      auto *LTy = cast<FixedVectorType>(L->getType());
      auto *OrigOpTy =
          dyn_cast<IntegerType>(VL0->getOperand(0)->getType()->getScalarType());
      if (OrigOpTy && L->getType() == R->getType() &&
          LTy->getElementType()->getIntegerBitWidth() <
              OrigOpTy->getBitWidth()) {
        bool IsSigned0 = GetOperandSignedness(0);
        bool IsSigned1 = GetOperandSignedness(1);
        bool NonNeg0 = isKnownNonNegative(L, SimplifyQuery(*DL));
        bool NonNeg1 = isKnownNonNegative(R, SimplifyQuery(*DL));
        bool KeepNarrowTy;
        if (ICmpInst::isSigned(P0))
          KeepNarrowTy = (IsSigned0 || NonNeg0) && (IsSigned1 || NonNeg1);
        else if (ICmpInst::isUnsigned(P0))
          KeepNarrowTy = (!IsSigned0 || NonNeg0) && (!IsSigned1 || NonNeg1);
        else
          KeepNarrowTy = IsSigned0 == IsSigned1 || NonNeg0 || NonNeg1;
        if (!KeepNarrowTy) {
          Type *CastTy = getWidenedType(OrigOpTy, LTy->getNumElements());
          L = Builder.CreateIntCast(L, CastTy, IsSigned0);
          R = Builder.CreateIntCast(R, CastTy, IsSigned1);
        }
      }
 
      Value *V = Builder.CreateCmp(P0, L, R);
      V = PropagateIRFlags(V);
      // Do not cast for cmps.
      VecTy = cast<FixedVectorType>(V->getType());
      V = FinalShuffle(V, E);
 
      E->VectorizedValue = V;
      ++NumVectorInstructions;
      return V;
    }
    case Instruction::Select: {
      setInsertPointAfterBundle(E);
 
      Value *Cond = vectorizeOperand(E, 0);
      Value *True = vectorizeOperand(E, 1);
      Value *False = vectorizeOperand(E, 2);
      if (True->getType() != VecTy || False->getType() != VecTy) {
        assert((It != MinBWs.end() || getOperandEntry(E, 1)->isGather() ||
                getOperandEntry(E, 2)->isGather() ||
                MinBWs.contains(getOperandEntry(E, 1)) ||
                MinBWs.contains(getOperandEntry(E, 2))) &&
               "Expected item in MinBWs.");
        if (True->getType() != VecTy)
          True = Builder.CreateIntCast(True, VecTy, GetOperandSignedness(1));
        if (False->getType() != VecTy)
          False = Builder.CreateIntCast(False, VecTy, GetOperandSignedness(2));
      }
 
      unsigned CondNumElements = getNumElements(Cond->getType());
      unsigned TrueNumElements = getNumElements(True->getType());
      assert(TrueNumElements >= CondNumElements &&
             TrueNumElements % CondNumElements == 0 &&
             "Cannot vectorize Instruction::Select");
      assert(TrueNumElements == getNumElements(False->getType()) &&
             "Cannot vectorize Instruction::Select");
      if (CondNumElements != TrueNumElements) {
        // When the return type is i1 but the source is fixed vector type, we
        // need to duplicate the condition value.
        Cond = Builder.CreateShuffleVector(
            Cond, createReplicatedMask(TrueNumElements / CondNumElements,
                                       CondNumElements));
      }
      assert(getNumElements(Cond->getType()) == TrueNumElements &&
             "Cannot vectorize Instruction::Select");
      Value *V =
          Builder.CreateSelectWithUnknownProfile(Cond, True, False, DEBUG_TYPE);
      V = FinalShuffle(V, E);
 
      E->VectorizedValue = V;
      ++NumVectorInstructions;
      return V;
    }
    case Instruction::FNeg: {
      setInsertPointAfterBundle(E);
 
      Value *Op = vectorizeOperand(E, 0);
 
      Value *V = Builder.CreateUnOp(
          static_cast<Instruction::UnaryOps>(E->getOpcode()), Op);
      V = PropagateIRFlags(V);
 
      V = FinalShuffle(V, E);
 
      E->VectorizedValue = V;
      ++NumVectorInstructions;
 
      return V;
    }
    case Instruction::Freeze: {
      setInsertPointAfterBundle(E);
 
      Value *Op = vectorizeOperand(E, 0);
 
      if (Op->getType() != VecTy) {
        assert((It != MinBWs.end() || getOperandEntry(E, 0)->isGather() ||
                MinBWs.contains(getOperandEntry(E, 0))) &&
               "Expected item in MinBWs.");
        Op = Builder.CreateIntCast(Op, VecTy, GetOperandSignedness(0));
      }
      Value *V = Builder.CreateFreeze(Op);
      V = FinalShuffle(V, E);
 
      E->VectorizedValue = V;
      ++NumVectorInstructions;
 
      return V;
    }
    case Instruction::Add:
    case Instruction::FAdd:
    case Instruction::Sub:
    case Instruction::FSub:
    case Instruction::Mul:
    case Instruction::FMul:
    case Instruction::UDiv:
    case Instruction::SDiv:
    case Instruction::FDiv:
    case Instruction::URem:
    case Instruction::SRem:
    case Instruction::FRem:
    case Instruction::Shl:
    case Instruction::LShr:
    case Instruction::AShr:
    case Instruction::And:
    case Instruction::Or:
    case Instruction::Xor: {
      setInsertPointAfterBundle(E);
 
      if (E->hasReassocScalars()) {
        // Vectorize operand columns, then combine pairwise in a balanced
        // tree; negated columns (flattened subtracts)