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 "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/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/GlobalsModRef.h"
#include "llvm/Analysis/IVDescriptors.h"
#include "llvm/Analysis/LoopAccessAnalysis.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/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/Module.h"
#include "llvm/IR/Operator.h"
#include "llvm/IR/PatternMatch.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"
#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/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 <memory>
#include <optional>
#include <set>
#include <string>
#include <tuple>
#include <utility>
 
using namespace llvm;
using namespace llvm::PatternMatch;
using namespace slpvectorizer;
 
#define SV_NAME "slp-vectorizer"
#define DEBUG_TYPE "SLP"
 
STATISTIC(NumVectorInstructions, "Number of vector instructions generated");
 
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> SLPSkipEarlyProfitabilityCheck(
    "slp-skip-early-profitability-check", cl::init(false), cl::Hidden,
    cl::desc("When true, SLP vectorizer bypasses profitability checks based on "
             "heuristics and makes vectorization decision via cost modeling."));
 
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"));
 
// NOTE: If AllowHorRdxIdenityOptimization is true, the optimization will run
// even if we match a reduction but do not vectorize in the end.
static cl::opt<bool> AllowHorRdxIdenityOptimization(
    "slp-optimize-identity-hor-reduction-ops", cl::init(true), cl::Hidden,
    cl::desc("Allow optimization of original scalar identity operations on "
             "matched horizontal reductions."));
 
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"));
 
// 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> MaxProfitableLoadStride(
    "slp-max-stride", cl::init(8), cl::Hidden,
    cl::desc("The maximum stride, considered to be profitable."));
 
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."));
 
// 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;
 
// Limit of the number of uses for potentially transformed instructions/values,
// used in checks to avoid compile-time explode.
static constexpr int UsesLimit = 64;
 
// 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;
 
/// Predicate for the element types that the SLP vectorizer supports.
///
/// The most important thing to filter here are types which are invalid in LLVM
/// vectors. We also filter target specific types which have absolutely no
/// meaningful vectorization path such as x86_fp80 and ppc_f128. This just
/// avoids spending time checking the cost model and realizing that they will
/// be inevitably scalarized.
static bool isValidElementType(Type *Ty) {
  // TODO: Support ScalableVectorType.
  if (SLPReVec && isa<FixedVectorType>(Ty))
    Ty = Ty->getScalarType();
  return VectorType::isValidElementType(Ty) && !Ty->isX86_FP80Ty() &&
         !Ty->isPPC_FP128Ty();
}
 
/// \returns the number of elements for Ty.
static unsigned getNumElements(Type *Ty) {
  assert(!isa<ScalableVectorType>(Ty) &&
         "ScalableVectorType is not supported.");
  if (auto *VecTy = dyn_cast<FixedVectorType>(Ty))
    return VecTy->getNumElements();
  return 1;
}
 
/// \returns the vector type of ScalarTy based on vectorization factor.
static FixedVectorType *getWidenedType(Type *ScalarTy, unsigned VF) {
  return FixedVectorType::get(ScalarTy->getScalarType(),
                              VF * getNumElements(ScalarTy));
}
 
static void transformScalarShuffleIndiciesToVector(unsigned VecTyNumElements,
                                                   SmallVectorImpl<int> &Mask) {
  // The ShuffleBuilder implementation use shufflevector to splat an "element".
  // But the element have different meaning for SLP (scalar) and REVEC
  // (vector). We need to expand Mask into masks which shufflevector can use
  // directly.
  SmallVector<int> NewMask(Mask.size() * VecTyNumElements);
  for (unsigned I : seq<unsigned>(Mask.size()))
    for (auto [J, MaskV] : enumerate(MutableArrayRef(NewMask).slice(
             I * VecTyNumElements, VecTyNumElements)))
      MaskV = Mask[I] == PoisonMaskElem ? PoisonMaskElem
                                        : Mask[I] * VecTyNumElements + J;
  Mask.swap(NewMask);
}
 
/// \returns True if the value is a constant (but not globals/constant
/// expressions).
static bool isConstant(Value *V) {
  return isa<Constant>(V) && !isa<ConstantExpr, GlobalValue>(V);
}
 
/// Checks if \p V is one of vector-like instructions, i.e. undef,
/// insertelement/extractelement with constant indices for fixed vector type or
/// extractvalue instruction.
static bool isVectorLikeInstWithConstOps(Value *V) {
  if (!isa<InsertElementInst, ExtractElementInst>(V) &&
      !isa<ExtractValueInst, UndefValue>(V))
    return false;
  auto *I = dyn_cast<Instruction>(V);
  if (!I || isa<ExtractValueInst>(I))
    return true;
  if (!isa<FixedVectorType>(I->getOperand(0)->getType()))
    return false;
  if (isa<ExtractElementInst>(I))
    return isConstant(I->getOperand(1));
  assert(isa<InsertElementInst>(V) && "Expected only insertelement.");
  return isConstant(I->getOperand(2));
}
 
/// Returns power-of-2 number of elements in a single register (part), given the
/// total number of elements \p Size and number of registers (parts) \p
/// NumParts.
static unsigned getPartNumElems(unsigned Size, unsigned NumParts) {
  return PowerOf2Ceil(divideCeil(Size, NumParts));
}
 
/// Returns correct remaining number of elements, considering total amount \p
/// Size, (power-of-2 number) of elements in a single register \p PartNumElems
/// and current register (part) \p Part.
static unsigned getNumElems(unsigned Size, unsigned PartNumElems,
                            unsigned Part) {
  return std::min<unsigned>(PartNumElems, Size - Part * PartNumElems);
}
 
#if !defined(NDEBUG)
/// Print a short descriptor of the instruction bundle suitable for debug output.
static std::string shortBundleName(ArrayRef<Value *> VL, int Idx = -1) {
  std::string Result;
  raw_string_ostream OS(Result);
  if (Idx >= 0)
    OS << "Idx: " << Idx << ", ";
  OS << "n=" << VL.size() << " [" << *VL.front() << ", ..]";
  OS.flush();
  return Result;
}
#endif
 
/// \returns true if all of the instructions in \p VL are in the same block or
/// false otherwise.
static bool allSameBlock(ArrayRef<Value *> VL) {
  Instruction *I0 = dyn_cast<Instruction>(VL[0]);
  if (!I0)
    return false;
  if (all_of(VL, isVectorLikeInstWithConstOps))
    return true;
 
  BasicBlock *BB = I0->getParent();
  for (int I = 1, E = VL.size(); I < E; I++) {
    auto *II = dyn_cast<Instruction>(VL[I]);
    if (!II)
      return false;
 
    if (BB != II->getParent())
      return false;
  }
  return true;
}
 
/// \returns True if all of the values in \p VL are constants (but not
/// globals/constant expressions).
static bool allConstant(ArrayRef<Value *> VL) {
  // Constant expressions and globals can't be vectorized like normal integer/FP
  // constants.
  return all_of(VL, isConstant);
}
 
/// \returns True if all of the values in \p VL are identical or some of them
/// are UndefValue.
static bool isSplat(ArrayRef<Value *> VL) {
  Value *FirstNonUndef = nullptr;
  for (Value *V : VL) {
    if (isa<UndefValue>(V))
      continue;
    if (!FirstNonUndef) {
      FirstNonUndef = V;
      continue;
    }
    if (V != FirstNonUndef)
      return false;
  }
  return FirstNonUndef != nullptr;
}
 
/// \returns True if \p I is commutative, handles CmpInst and BinaryOperator.
static bool isCommutative(Instruction *I) {
  if (auto *Cmp = dyn_cast<CmpInst>(I))
    return Cmp->isCommutative();
  if (auto *BO = dyn_cast<BinaryOperator>(I))
    return BO->isCommutative() ||
           (BO->getOpcode() == Instruction::Sub &&
            !BO->hasNUsesOrMore(UsesLimit) &&
            all_of(
                BO->uses(),
                [](const Use &U) {
                  // Commutative, if icmp eq/ne sub, 0
                  ICmpInst::Predicate Pred;
                  if (match(U.getUser(),
                            m_ICmp(Pred, m_Specific(U.get()), m_Zero())) &&
                      (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE))
                    return true;
                  // Commutative, if abs(sub nsw, true) or abs(sub, false).
                  ConstantInt *Flag;
                  return match(U.getUser(),
                               m_Intrinsic<Intrinsic::abs>(
                                   m_Specific(U.get()), m_ConstantInt(Flag))) &&
                         (!cast<Instruction>(U.get())->hasNoSignedWrap() ||
                          Flag->isOne());
                })) ||
           (BO->getOpcode() == Instruction::FSub &&
            !BO->hasNUsesOrMore(UsesLimit) &&
            all_of(BO->uses(), [](const Use &U) {
              return match(U.getUser(),
                           m_Intrinsic<Intrinsic::fabs>(m_Specific(U.get())));
            }));
  return I->isCommutative();
}
 
template <typename T>
static std::optional<unsigned> getInsertExtractIndex(const Value *Inst,
                                                     unsigned Offset) {
  static_assert(std::is_same_v<T, InsertElementInst> ||
                    std::is_same_v<T, ExtractElementInst>,
                "unsupported T");
  int Index = Offset;
  if (const auto *IE = dyn_cast<T>(Inst)) {
    const auto *VT = dyn_cast<FixedVectorType>(IE->getType());
    if (!VT)
      return std::nullopt;
    const auto *CI = dyn_cast<ConstantInt>(IE->getOperand(2));
    if (!CI)
      return std::nullopt;
    if (CI->getValue().uge(VT->getNumElements()))
      return std::nullopt;
    Index *= VT->getNumElements();
    Index += CI->getZExtValue();
    return Index;
  }
  return std::nullopt;
}
 
/// \returns inserting or extracting index of InsertElement, ExtractElement or
/// InsertValue instruction, using Offset as base offset for index.
/// \returns std::nullopt if the index is not an immediate.
static std::optional<unsigned> getElementIndex(const Value *Inst,
                                               unsigned Offset = 0) {
  if (auto Index = getInsertExtractIndex<InsertElementInst>(Inst, Offset))
    return Index;
  if (auto Index = getInsertExtractIndex<ExtractElementInst>(Inst, Offset))
    return Index;
 
  int Index = Offset;
 
  const auto *IV = dyn_cast<InsertValueInst>(Inst);
  if (!IV)
    return std::nullopt;
 
  Type *CurrentType = IV->getType();
  for (unsigned I : IV->indices()) {
    if (const auto *ST = dyn_cast<StructType>(CurrentType)) {
      Index *= ST->getNumElements();
      CurrentType = ST->getElementType(I);
    } else if (const auto *AT = dyn_cast<ArrayType>(CurrentType)) {
      Index *= AT->getNumElements();
      CurrentType = AT->getElementType();
    } else {
      return std::nullopt;
    }
    Index += I;
  }
  return Index;
}
 
namespace {
/// Specifies the way the mask should be analyzed for undefs/poisonous elements
/// in the shuffle mask.
enum class UseMask {
  FirstArg, ///< The mask is expected to be for permutation of 1-2 vectors,
            ///< check for the mask elements for the first argument (mask
            ///< indices are in range [0:VF)).
  SecondArg, ///< The mask is expected to be for permutation of 2 vectors, check
             ///< for the mask elements for the second argument (mask indices
             ///< are in range [VF:2*VF))
  UndefsAsMask ///< Consider undef mask elements (-1) as placeholders for
               ///< future shuffle elements and mark them as ones as being used
               ///< in future. Non-undef elements are considered as unused since
               ///< they're already marked as used in the mask.
};
} // namespace
 
/// Prepares a use bitset for the given mask either for the first argument or
/// for the second.
static SmallBitVector buildUseMask(int VF, ArrayRef<int> Mask,
                                   UseMask MaskArg) {
  SmallBitVector UseMask(VF, true);
  for (auto [Idx, Value] : enumerate(Mask)) {
    if (Value == PoisonMaskElem) {
      if (MaskArg == UseMask::UndefsAsMask)
        UseMask.reset(Idx);
      continue;
    }
    if (MaskArg == UseMask::FirstArg && Value < VF)
      UseMask.reset(Value);
    else if (MaskArg == UseMask::SecondArg && Value >= VF)
      UseMask.reset(Value - VF);
  }
  return UseMask;
}
 
/// Checks if the given value is actually an undefined constant vector.
/// Also, if the \p UseMask is not empty, tries to check if the non-masked
/// elements actually mask the insertelement buildvector, if any.
template <bool IsPoisonOnly = false>
static SmallBitVector isUndefVector(const Value *V,
                                    const SmallBitVector &UseMask = {}) {
  SmallBitVector Res(UseMask.empty() ? 1 : UseMask.size(), true);
  using T = std::conditional_t<IsPoisonOnly, PoisonValue, UndefValue>;
  if (isa<T>(V))
    return Res;
  auto *VecTy = dyn_cast<FixedVectorType>(V->getType());
  if (!VecTy)
    return Res.reset();
  auto *C = dyn_cast<Constant>(V);
  if (!C) {
    if (!UseMask.empty()) {
      const Value *Base = V;
      while (auto *II = dyn_cast<InsertElementInst>(Base)) {
        Base = II->getOperand(0);
        if (isa<T>(II->getOperand(1)))
          continue;
        std::optional<unsigned> Idx = getElementIndex(II);
        if (!Idx) {
          Res.reset();
          return Res;
        }
        if (*Idx < UseMask.size() && !UseMask.test(*Idx))
          Res.reset(*Idx);
      }
      // TODO: Add analysis for shuffles here too.
      if (V == Base) {
        Res.reset();
      } else {
        SmallBitVector SubMask(UseMask.size(), false);
        Res &= isUndefVector<IsPoisonOnly>(Base, SubMask);
      }
    } else {
      Res.reset();
    }
    return Res;
  }
  for (unsigned I = 0, E = VecTy->getNumElements(); I != E; ++I) {
    if (Constant *Elem = C->getAggregateElement(I))
      if (!isa<T>(Elem) &&
          (UseMask.empty() || (I < UseMask.size() && !UseMask.test(I))))
        Res.reset(I);
  }
  return Res;
}
 
/// 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) {
  const auto *It = find_if(VL, IsaPred<ExtractElementInst>);
  if (It == VL.end())
    return std::nullopt;
  unsigned Size =
      std::accumulate(VL.begin(), VL.end(), 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);
  });
  enum ShuffleMode { Unknown, Select, Permute };
  ShuffleMode CommonShuffleMode = Unknown;
  Mask.assign(VL.size(), PoisonMaskElem);
  for (unsigned I = 0, E = VL.size(); I < E; ++I) {
    // Undef can be represented as an undef element in a vector.
    if (isa<UndefValue>(VL[I]))
      continue;
    auto *EI = cast<ExtractElementInst>(VL[I]);
    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 True if Extract{Value,Element} instruction extracts element Idx.
static std::optional<unsigned> getExtractIndex(Instruction *E) {
  unsigned Opcode = E->getOpcode();
  assert((Opcode == Instruction::ExtractElement ||
          Opcode == Instruction::ExtractValue) &&
         "Expected extractelement or extractvalue instruction.");
  if (Opcode == Instruction::ExtractElement) {
    auto *CI = dyn_cast<ConstantInt>(E->getOperand(1));
    if (!CI)
      return std::nullopt;
    return CI->getZExtValue();
  }
  auto *EI = cast<ExtractValueInst>(E);
  if (EI->getNumIndices() != 1)
    return std::nullopt;
  return *EI->idx_begin();
}
 
namespace {
 
/// Main data required for vectorization of instructions.
struct InstructionsState {
  /// The very first instruction in the list with the main opcode.
  Value *OpValue = nullptr;
 
  /// The main/alternate instruction.
  Instruction *MainOp = nullptr;
  Instruction *AltOp = nullptr;
 
  /// The main/alternate opcodes for the list of instructions.
  unsigned getOpcode() const {
    return MainOp ? MainOp->getOpcode() : 0;
  }
 
  unsigned getAltOpcode() const {
    return AltOp ? AltOp->getOpcode() : 0;
  }
 
  /// Some of the instructions in the list have alternate opcodes.
  bool isAltShuffle() const { return AltOp != MainOp; }
 
  bool isOpcodeOrAlt(Instruction *I) const {
    unsigned CheckedOpcode = I->getOpcode();
    return getOpcode() == CheckedOpcode || getAltOpcode() == CheckedOpcode;
  }
 
  InstructionsState() = delete;
  InstructionsState(Value *OpValue, Instruction *MainOp, Instruction *AltOp)
      : OpValue(OpValue), MainOp(MainOp), AltOp(AltOp) {}
};
 
} // end anonymous namespace
 
/// 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.
static Value *isOneOf(const InstructionsState &S, Value *Op) {
  auto *I = dyn_cast<Instruction>(Op);
  if (I && S.isOpcodeOrAlt(I))
    return Op;
  return S.OpValue;
}
 
/// \returns true if \p Opcode is allowed as part of the main/alternate
/// instruction for SLP vectorization.
///
/// Example of unsupported opcode is SDIV that can potentially cause UB if the
/// "shuffled out" lane would result in division by zero.
static bool isValidForAlternation(unsigned Opcode) {
  if (Instruction::isIntDivRem(Opcode))
    return false;
 
  return true;
}
 
static InstructionsState getSameOpcode(ArrayRef<Value *> VL,
                                       const TargetLibraryInfo &TLI,
                                       unsigned BaseIndex = 0);
 
/// Checks if the provided operands of 2 cmp instructions are compatible, i.e.
/// compatible instructions or constants, or just some other regular values.
static bool areCompatibleCmpOps(Value *BaseOp0, Value *BaseOp1, Value *Op0,
                                Value *Op1, const TargetLibraryInfo &TLI) {
  return (isConstant(BaseOp0) && isConstant(Op0)) ||
         (isConstant(BaseOp1) && isConstant(Op1)) ||
         (!isa<Instruction>(BaseOp0) && !isa<Instruction>(Op0) &&
          !isa<Instruction>(BaseOp1) && !isa<Instruction>(Op1)) ||
         BaseOp0 == Op0 || BaseOp1 == Op1 ||
         getSameOpcode({BaseOp0, Op0}, TLI).getOpcode() ||
         getSameOpcode({BaseOp1, Op1}, TLI).getOpcode();
}
 
/// \returns true if a compare instruction \p CI has similar "look" and
/// same predicate as \p BaseCI, "as is" or with its operands and predicate
/// swapped, false otherwise.
static bool isCmpSameOrSwapped(const CmpInst *BaseCI, const CmpInst *CI,
                               const TargetLibraryInfo &TLI) {
  assert(BaseCI->getOperand(0)->getType() == CI->getOperand(0)->getType() &&
         "Assessing comparisons of different types?");
  CmpInst::Predicate BasePred = BaseCI->getPredicate();
  CmpInst::Predicate Pred = CI->getPredicate();
  CmpInst::Predicate SwappedPred = CmpInst::getSwappedPredicate(Pred);
 
  Value *BaseOp0 = BaseCI->getOperand(0);
  Value *BaseOp1 = BaseCI->getOperand(1);
  Value *Op0 = CI->getOperand(0);
  Value *Op1 = CI->getOperand(1);
 
  return (BasePred == Pred &&
          areCompatibleCmpOps(BaseOp0, BaseOp1, Op0, Op1, TLI)) ||
         (BasePred == SwappedPred &&
          areCompatibleCmpOps(BaseOp0, BaseOp1, Op1, Op0, TLI));
}
 
/// \returns analysis of the Instructions in \p VL described in
/// InstructionsState, the Opcode that we suppose the whole list
/// could be vectorized even if its structure is diverse.
static InstructionsState getSameOpcode(ArrayRef<Value *> VL,
                                       const TargetLibraryInfo &TLI,
                                       unsigned BaseIndex) {
  // Make sure these are all Instructions.
  if (llvm::any_of(VL, [](Value *V) { return !isa<Instruction>(V); }))
    return InstructionsState(VL[BaseIndex], nullptr, nullptr);
 
  bool IsCastOp = isa<CastInst>(VL[BaseIndex]);
  bool IsBinOp = isa<BinaryOperator>(VL[BaseIndex]);
  bool IsCmpOp = isa<CmpInst>(VL[BaseIndex]);
  CmpInst::Predicate BasePred =
      IsCmpOp ? cast<CmpInst>(VL[BaseIndex])->getPredicate()
              : CmpInst::BAD_ICMP_PREDICATE;
  unsigned Opcode = cast<Instruction>(VL[BaseIndex])->getOpcode();
  unsigned AltOpcode = Opcode;
  unsigned AltIndex = BaseIndex;
 
  bool SwappedPredsCompatible = [&]() {
    if (!IsCmpOp)
      return false;
    SetVector<unsigned> UniquePreds, UniqueNonSwappedPreds;
    UniquePreds.insert(BasePred);
    UniqueNonSwappedPreds.insert(BasePred);
    for (Value *V : VL) {
      auto *I = dyn_cast<CmpInst>(V);
      if (!I)
        return false;
      CmpInst::Predicate CurrentPred = I->getPredicate();
      CmpInst::Predicate SwappedCurrentPred =
          CmpInst::getSwappedPredicate(CurrentPred);
      UniqueNonSwappedPreds.insert(CurrentPred);
      if (!UniquePreds.contains(CurrentPred) &&
          !UniquePreds.contains(SwappedCurrentPred))
        UniquePreds.insert(CurrentPred);
    }
    // Total number of predicates > 2, but if consider swapped predicates
    // compatible only 2, consider swappable predicates as compatible opcodes,
    // not alternate.
    return UniqueNonSwappedPreds.size() > 2 && UniquePreds.size() == 2;
  }();
  // Check for one alternate opcode from another BinaryOperator.
  // TODO - generalize to support all operators (types, calls etc.).
  auto *IBase = cast<Instruction>(VL[BaseIndex]);
  Intrinsic::ID BaseID = 0;
  SmallVector<VFInfo> BaseMappings;
  if (auto *CallBase = dyn_cast<CallInst>(IBase)) {
    BaseID = getVectorIntrinsicIDForCall(CallBase, &TLI);
    BaseMappings = VFDatabase(*CallBase).getMappings(*CallBase);
    if (!isTriviallyVectorizable(BaseID) && BaseMappings.empty())
      return InstructionsState(VL[BaseIndex], nullptr, nullptr);
  }
  for (int Cnt = 0, E = VL.size(); Cnt < E; Cnt++) {
    auto *I = cast<Instruction>(VL[Cnt]);
    unsigned InstOpcode = I->getOpcode();
    if (IsBinOp && isa<BinaryOperator>(I)) {
      if (InstOpcode == Opcode || InstOpcode == AltOpcode)
        continue;
      if (Opcode == AltOpcode && isValidForAlternation(InstOpcode) &&
          isValidForAlternation(Opcode)) {
        AltOpcode = InstOpcode;
        AltIndex = Cnt;
        continue;
      }
    } else if (IsCastOp && isa<CastInst>(I)) {
      Value *Op0 = IBase->getOperand(0);
      Type *Ty0 = Op0->getType();
      Value *Op1 = I->getOperand(0);
      Type *Ty1 = Op1->getType();
      if (Ty0 == Ty1) {
        if (InstOpcode == Opcode || InstOpcode == AltOpcode)
          continue;
        if (Opcode == AltOpcode) {
          assert(isValidForAlternation(Opcode) &&
                 isValidForAlternation(InstOpcode) &&
                 "Cast isn't safe for alternation, logic needs to be updated!");
          AltOpcode = InstOpcode;
          AltIndex = Cnt;
          continue;
        }
      }
    } else if (auto *Inst = dyn_cast<CmpInst>(VL[Cnt]); Inst && IsCmpOp) {
      auto *BaseInst = cast<CmpInst>(VL[BaseIndex]);
      Type *Ty0 = BaseInst->getOperand(0)->getType();
      Type *Ty1 = Inst->getOperand(0)->getType();
      if (Ty0 == Ty1) {
        assert(InstOpcode == Opcode && "Expected same CmpInst opcode.");
        // Check for compatible operands. If the corresponding operands are not
        // compatible - need to perform alternate vectorization.
        CmpInst::Predicate CurrentPred = Inst->getPredicate();
        CmpInst::Predicate SwappedCurrentPred =
            CmpInst::getSwappedPredicate(CurrentPred);
 
        if ((E == 2 || SwappedPredsCompatible) &&
            (BasePred == CurrentPred || BasePred == SwappedCurrentPred))
          continue;
 
        if (isCmpSameOrSwapped(BaseInst, Inst, TLI))
          continue;
        auto *AltInst = cast<CmpInst>(VL[AltIndex]);
        if (AltIndex != BaseIndex) {
          if (isCmpSameOrSwapped(AltInst, Inst, TLI))
            continue;
        } else if (BasePred != CurrentPred) {
          assert(
              isValidForAlternation(InstOpcode) &&
              "CmpInst isn't safe for alternation, logic needs to be updated!");
          AltIndex = Cnt;
          continue;
        }
        CmpInst::Predicate AltPred = AltInst->getPredicate();
        if (BasePred == CurrentPred || BasePred == SwappedCurrentPred ||
            AltPred == CurrentPred || AltPred == SwappedCurrentPred)
          continue;
      }
    } else if (InstOpcode == Opcode || InstOpcode == AltOpcode) {
      if (auto *Gep = dyn_cast<GetElementPtrInst>(I)) {
        if (Gep->getNumOperands() != 2 ||
            Gep->getOperand(0)->getType() != IBase->getOperand(0)->getType())
          return InstructionsState(VL[BaseIndex], nullptr, nullptr);
      } else if (auto *EI = dyn_cast<ExtractElementInst>(I)) {
        if (!isVectorLikeInstWithConstOps(EI))
          return InstructionsState(VL[BaseIndex], nullptr, nullptr);
      } else if (auto *LI = dyn_cast<LoadInst>(I)) {
        auto *BaseLI = cast<LoadInst>(IBase);
        if (!LI->isSimple() || !BaseLI->isSimple())
          return InstructionsState(VL[BaseIndex], nullptr, nullptr);
      } else if (auto *Call = dyn_cast<CallInst>(I)) {
        auto *CallBase = cast<CallInst>(IBase);
        if (Call->getCalledFunction() != CallBase->getCalledFunction())
          return InstructionsState(VL[BaseIndex], nullptr, nullptr);
        if (Call->hasOperandBundles() && (!CallBase->hasOperandBundles() ||
            !std::equal(Call->op_begin() + Call->getBundleOperandsStartIndex(),
                        Call->op_begin() + Call->getBundleOperandsEndIndex(),
                        CallBase->op_begin() +
                            CallBase->getBundleOperandsStartIndex())))
          return InstructionsState(VL[BaseIndex], nullptr, nullptr);
        Intrinsic::ID ID = getVectorIntrinsicIDForCall(Call, &TLI);
        if (ID != BaseID)
          return InstructionsState(VL[BaseIndex], nullptr, nullptr);
        if (!ID) {
          SmallVector<VFInfo> Mappings = VFDatabase(*Call).getMappings(*Call);
          if (Mappings.size() != BaseMappings.size() ||
              Mappings.front().ISA != BaseMappings.front().ISA ||
              Mappings.front().ScalarName != BaseMappings.front().ScalarName ||
              Mappings.front().VectorName != BaseMappings.front().VectorName ||
              Mappings.front().Shape.VF != BaseMappings.front().Shape.VF ||
              Mappings.front().Shape.Parameters !=
                  BaseMappings.front().Shape.Parameters)
            return InstructionsState(VL[BaseIndex], nullptr, nullptr);
        }
      }
      continue;
    }
    return InstructionsState(VL[BaseIndex], nullptr, nullptr);
  }
 
  return InstructionsState(VL[BaseIndex], cast<Instruction>(VL[BaseIndex]),
                           cast<Instruction>(VL[AltIndex]));
}
 
/// \returns true if all of the values in \p VL have the same type or false
/// otherwise.
static bool allSameType(ArrayRef<Value *> VL) {
  Type *Ty = VL.front()->getType();
  return all_of(VL.drop_front(), [&](Value *V) { return V->getType() == Ty; });
}
 
/// \returns True if in-tree use also needs extract. This refers to
/// possible scalar operand in vectorized instruction.
static bool doesInTreeUserNeedToExtract(Value *Scalar, Instruction *UserInst,
                                        TargetLibraryInfo *TLI) {
  unsigned Opcode = UserInst->getOpcode();
  switch (Opcode) {
  case Instruction::Load: {
    LoadInst *LI = cast<LoadInst>(UserInst);
    return (LI->getPointerOperand() == Scalar);
  }
  case Instruction::Store: {
    StoreInst *SI = cast<StoreInst>(UserInst);
    return (SI->getPointerOperand() == Scalar);
  }
  case Instruction::Call: {
    CallInst *CI = cast<CallInst>(UserInst);
    Intrinsic::ID ID = getVectorIntrinsicIDForCall(CI, TLI);
    return any_of(enumerate(CI->args()), [&](auto &&Arg) {
      return isVectorIntrinsicWithScalarOpAtArg(ID, Arg.index()) &&
             Arg.value().get() == Scalar;
    });
  }
  default:
    return false;
  }
}
 
/// \returns the AA location that is being access by the instruction.
static MemoryLocation getLocation(Instruction *I) {
  if (StoreInst *SI = dyn_cast<StoreInst>(I))
    return MemoryLocation::get(SI);
  if (LoadInst *LI = dyn_cast<LoadInst>(I))
    return MemoryLocation::get(LI);
  return MemoryLocation();
}
 
/// \returns True if the instruction is not a volatile or atomic load/store.
static bool isSimple(Instruction *I) {
  if (LoadInst *LI = dyn_cast<LoadInst>(I))
    return LI->isSimple();
  if (StoreInst *SI = dyn_cast<StoreInst>(I))
    return SI->isSimple();
  if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(I))
    return !MI->isVolatile();
  return true;
}
 
/// Shuffles \p Mask in accordance with the given \p SubMask.
/// \param ExtendingManyInputs Supports reshuffling of the mask with not only
/// one but two input vectors.
static void addMask(SmallVectorImpl<int> &Mask, ArrayRef<int> SubMask,
                    bool ExtendingManyInputs = false) {
  if (SubMask.empty())
    return;
  assert(
      (!ExtendingManyInputs || SubMask.size() > Mask.size() ||
       // Check if input scalars were extended to match the size of other node.
       (SubMask.size() == Mask.size() &&
        std::all_of(std::next(Mask.begin(), Mask.size() / 2), Mask.end(),
                    [](int Idx) { return Idx == PoisonMaskElem; }))) &&
      "SubMask with many inputs support must be larger than the mask.");
  if (Mask.empty()) {
    Mask.append(SubMask.begin(), SubMask.end());
    return;
  }
  SmallVector<int> NewMask(SubMask.size(), PoisonMaskElem);
  int TermValue = std::min(Mask.size(), SubMask.size());
  for (int I = 0, E = SubMask.size(); I < E; ++I) {
    if (SubMask[I] == PoisonMaskElem ||
        (!ExtendingManyInputs &&
         (SubMask[I] >= TermValue || Mask[SubMask[I]] >= TermValue)))
      continue;
    NewMask[I] = Mask[SubMask[I]];
  }
  Mask.swap(NewMask);
}
 
/// Order may have elements assigned special value (size) which is out of
/// bounds. Such indices only appear on places which correspond to undef values
/// (see canReuseExtract for details) and used in order to avoid undef values
/// have effect on operands ordering.
/// The first loop below simply finds all unused indices and then the next loop
/// nest assigns these indices for undef values positions.
/// As an example below Order has two undef positions and they have assigned
/// values 3 and 7 respectively:
/// before:  6 9 5 4 9 2 1 0
/// after:   6 3 5 4 7 2 1 0
static void fixupOrderingIndices(MutableArrayRef<unsigned> Order) {
  const unsigned Sz = Order.size();
  SmallBitVector UnusedIndices(Sz, /*t=*/true);
  SmallBitVector MaskedIndices(Sz);
  for (unsigned I = 0; I < Sz; ++I) {
    if (Order[I] < Sz)
      UnusedIndices.reset(Order[I]);
    else
      MaskedIndices.set(I);
  }
  if (MaskedIndices.none())
    return;
  assert(UnusedIndices.count() == MaskedIndices.count() &&
         "Non-synced masked/available indices.");
  int Idx = UnusedIndices.find_first();
  int MIdx = MaskedIndices.find_first();
  while (MIdx >= 0) {
    assert(Idx >= 0 && "Indices must be synced.");
    Order[MIdx] = Idx;
    Idx = UnusedIndices.find_next(Idx);
    MIdx = MaskedIndices.find_next(MIdx);
  }
}
 
/// \returns a bitset for selecting opcodes. false for Opcode0 and true for
/// Opcode1.
SmallBitVector getAltInstrMask(ArrayRef<Value *> VL, unsigned Opcode0,
                               unsigned Opcode1) {
  Type *ScalarTy = VL[0]->getType();
  unsigned ScalarTyNumElements = getNumElements(ScalarTy);
  SmallBitVector OpcodeMask(VL.size() * ScalarTyNumElements, false);
  for (unsigned Lane : seq<unsigned>(VL.size()))
    if (cast<Instruction>(VL[Lane])->getOpcode() == Opcode1)
      OpcodeMask.set(Lane * ScalarTyNumElements,
                     Lane * ScalarTyNumElements + ScalarTyNumElements);
  return OpcodeMask;
}
 
namespace llvm {
 
static void inversePermutation(ArrayRef<unsigned> Indices,
                               SmallVectorImpl<int> &Mask) {
  Mask.clear();
  const unsigned E = Indices.size();
  Mask.resize(E, PoisonMaskElem);
  for (unsigned I = 0; I < E; ++I)
    Mask[Indices[I]] = I;
}
 
/// Reorders the list of scalars in accordance with the given \p Mask.
static void reorderScalars(SmallVectorImpl<Value *> &Scalars,
                           ArrayRef<int> Mask) {
  assert(!Mask.empty() && "Expected non-empty mask.");
  SmallVector<Value *> Prev(Scalars.size(),
                            PoisonValue::get(Scalars.front()->getType()));
  Prev.swap(Scalars);
  for (unsigned I = 0, E = Prev.size(); I < E; ++I)
    if (Mask[I] != PoisonMaskElem)
      Scalars[Mask[I]] = Prev[I];
}
 
/// Checks if the provided value does not require scheduling. It does not
/// require scheduling if this is not an instruction or it is an instruction
/// that does not read/write memory and all operands are either not instructions
/// or phi nodes or instructions from different blocks.
static bool areAllOperandsNonInsts(Value *V) {
  auto *I = dyn_cast<Instruction>(V);
  if (!I)
    return true;
  return !mayHaveNonDefUseDependency(*I) &&
    all_of(I->operands(), [I](Value *V) {
      auto *IO = dyn_cast<Instruction>(V);
      if (!IO)
        return true;
      return isa<PHINode>(IO) || IO->getParent() != I->getParent();
    });
}
 
/// Checks if the provided value does not require scheduling. It does not
/// require scheduling if this is not an instruction or it is an instruction
/// that does not read/write memory and all users are phi nodes or instructions
/// from the different blocks.
static bool isUsedOutsideBlock(Value *V) {
  auto *I = dyn_cast<Instruction>(V);
  if (!I)
    return true;
  // Limits the number of uses to save compile time.
  return !I->mayReadOrWriteMemory() && !I->hasNUsesOrMore(UsesLimit) &&
         all_of(I->users(), [I](User *U) {
           auto *IU = dyn_cast<Instruction>(U);
           if (!IU)
             return true;
           return IU->getParent() != I->getParent() || isa<PHINode>(IU);
         });
}
 
/// Checks if the specified value does not require scheduling. It does not
/// require scheduling if all operands and all users do not need to be scheduled
/// in the current basic block.
static bool doesNotNeedToBeScheduled(Value *V) {
  return areAllOperandsNonInsts(V) && isUsedOutsideBlock(V);
}
 
/// Checks if the specified array of instructions does not require scheduling.
/// It is so if all either instructions have operands that do not require
/// scheduling or their users do not require scheduling since they are phis or
/// in other basic blocks.
static bool doesNotNeedToSchedule(ArrayRef<Value *> VL) {
  return !VL.empty() &&
         (all_of(VL, isUsedOutsideBlock) || all_of(VL, areAllOperandsNonInsts));
}
 
namespace slpvectorizer {
 
/// Bottom Up SLP Vectorizer.
class BoUpSLP {
  struct TreeEntry;
  struct ScheduleData;
  class ShuffleCostEstimator;
  class ShuffleInstructionBuilder;
 
public:
  /// Tracks the state we can represent the loads in the given sequence.
  enum class LoadsState {
    Gather,
    Vectorize,
    ScatterVectorize,
    StridedVectorize
  };
 
  using ValueList = SmallVector<Value *, 8>;
  using InstrList = SmallVector<Instruction *, 16>;
  using ValueSet = SmallPtrSet<Value *, 16>;
  using StoreList = SmallVector<StoreInst *, 8>;
  using ExtraValueToDebugLocsMap =
      MapVector<Value *, SmallVector<Instruction *, 2>>;
  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),
        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.
  /// \param ReplacedExternals containd list of replaced external values
  /// {scalar, replace} after emitting extractelement for external uses.
  Value *
  vectorizeTree(const ExtraValueToDebugLocsMap &ExternallyUsedValues,
                SmallVectorImpl<std::pair<Value *, Value *>> &ReplacedExternals,
                Instruction *ReductionRoot = nullptr);
 
  /// \returns the cost incurred by unwanted spills and fills, caused by
  /// holding live values over call sites.
  InstructionCost getSpillCost() const;
 
  /// \returns the vectorization cost of the subtree that starts at \p VL.
  /// A negative number means that this is profitable.
  InstructionCost getTreeCost(ArrayRef<Value *> VectorizedVals = std::nullopt);
 
  /// 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);
 
  /// Returns whether the root node has in-tree uses.
  bool doesRootHaveInTreeUses() const {
    return !VectorizableTree.empty() &&
           !VectorizableTree.front()->UserTreeIndices.empty();
  }
 
  /// Return the scalars of the root node.
  ArrayRef<Value *> getRootNodeScalars() const {
    assert(!VectorizableTree.empty() && "No graph to get the first node from");
    return VectorizableTree.front()->Scalars;
  }
 
  /// 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(VectorizableTree.front().get()).second;
  }
 
  /// 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();
    ScalarToTreeEntry.clear();
    MultiNodeScalars.clear();
    MustGather.clear();
    NonScheduledFirst.clear();
    EntryToLastInstruction.clear();
    ExternalUses.clear();
    ExternalUsesAsGEPs.clear();
    for (auto &Iter : BlocksSchedules) {
      BlockScheduling *BS = Iter.second.get();
      BS->clear();
    }
    MinBWs.clear();
    ReductionBitWidth = 0;
    CastMaxMinBWSizes.reset();
    ExtraBitWidthNodes.clear();
    InstrElementSize.clear();
    UserIgnoreList = nullptr;
    PostponedGathers.clear();
    ValueToGatherNodes.clear();
  }
 
  unsigned getTreeSize() const { return VectorizableTree.size(); }
 
  /// Perform LICM and CSE on the newly generated gather sequences.
  void optimizeGatherSequence();
 
  /// 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.
  std::optional<OrdersType> findReusedOrderedScalars(const TreeEntry &TE);
 
  /// 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.
  std::optional<OrdersType> getReorderingData(const TreeEntry &TE,
                                              bool TopToBottom);
 
  /// 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);
 
  /// \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;
  }
 
  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 VectorizableTree is both tiny and not fully
  /// vectorizable. We do not vectorize such trees.
  bool isTreeTinyAndNotFullyVectorizable(bool ForReduction = false) const;
 
  /// Assume that a legal-sized 'or'-reduction of shifted/zexted loaded values
  /// can be load combined in the backend. Load combining may not be allowed in
  /// the IR optimizer, so we do not want to alter the pattern. For example,
  /// partially transforming a scalar bswap() pattern into vector code is
  /// effectively impossible for the backend to undo.
  /// TODO: If load combining is allowed in the IR optimizer, this analysis
  ///       may not be necessary.
  bool isLoadCombineReductionCandidate(RecurKind RdxKind) const;
 
  /// Assume that a vector of stores of bitwise-or/shifted/zexted loaded values
  /// can be load combined in the backend. Load combining may not be allowed in
  /// the IR optimizer, so we do not want to alter the pattern. For example,
  /// partially transforming a scalar bswap() pattern into vector code is
  /// effectively impossible for the backend to undo.
  /// TODO: If load combining is allowed in the IR optimizer, this analysis
  ///       may not be necessary.
  bool isLoadCombineCandidate(ArrayRef<Value *> Stores) 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 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,
                               bool TryRecursiveCheck = true) const;
 
  OptimizationRemarkEmitter *getORE() { return ORE; }
 
  /// This structure holds any data we need about the edges being traversed
  /// during buildTree_rec(). 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;
    }
  };
 
  /// 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 const int ScoreConsecutiveLoads = 4;
    /// 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 const int ScoreSplatLoads = 3;
    /// Loads from reversed memory addresses, e.g. load(A[i+1]), load(A[i]).
    static const int ScoreReversedLoads = 3;
    /// A load candidate for masked gather.
    static const int ScoreMaskedGatherCandidate = 1;
    /// ExtractElementInst from same vector and consecutive indexes.
    static const int ScoreConsecutiveExtracts = 4;
    /// ExtractElementInst from same vector and reversed indices.
    static const int ScoreReversedExtracts = 3;
    /// Constants.
    static const int ScoreConstants = 2;
    /// Instructions with the same opcode.
    static const int ScoreSameOpcode = 2;
    /// Instructions with alt opcodes (e.g, add + sub).
    static const int ScoreAltOpcodes = 1;
    /// Identical instructions (a.k.a. splat or broadcast).
    static const int ScoreSplat = 1;
    /// Matching with an undef is preferable to failing.
    static const int ScoreUndef = 1;
    /// Score for failing to find a decent match.
    static const int ScoreFail = 0;
    /// Score if all users are vectorized.
    static const int ScoreAllUserVectorized = 1;
 
    /// \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()) ||
          !isValidElementType(V2->getType()))
        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.getTreeEntry(U) != nullptr;
              });
            };
            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;
        }
        return LookAheadHeuristics::ScoreSplat;
      }
 
      auto CheckSameEntryOrFail = [&]() {
        if (const TreeEntry *TE1 = R.getTreeEntry(V1);
            TE1 && TE1 == R.getTreeEntry(V2))
          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<int> 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;
 
      // 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();
        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.getOpcode() &&
            (S.MainOp->getNumOperands() <= 2 || !MainAltOps.empty() ||
             !S.isAltShuffle()) &&
            all_of(Ops, [&S](Value *V) {
              return cast<Instruction>(V)->getNumOperands() ==
                     S.MainOp->getNumOperands();
            }))
          return S.isAltShuffle() ? LookAheadHeuristics::ScoreAltOpcodes
                                  : 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.
      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, std::nullopt);
          // 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;
 
    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 {
      Value *IdxLaneV = getData(Idx, Lane).V;
      if (!isa<Instruction>(IdxLaneV) || IdxLaneV == getData(OpIdx, Lane).V)
        return 0;
      SmallPtrSet<Value *, 4> Uniques;
      for (unsigned Ln = 0, E = getNumLanes(); Ln < E; ++Ln) {
        if (Ln == Lane)
          continue;
        Value *OpIdxLnV = getData(OpIdx, Ln).V;
        if (!isa<Instruction>(OpIdxLnV))
          return 0;
        Uniques.insert(OpIdxLnV);
      }
      int UniquesCount = Uniques.size();
      int UniquesCntWithIdxLaneV =
          Uniques.contains(IdxLaneV) ? UniquesCount : UniquesCount + 1;
      Value *OpIdxLaneV = getData(OpIdx, Lane).V;
      int UniquesCntWithOpIdxLaneV =
          Uniques.contains(OpIdxLaneV) ? UniquesCount : UniquesCount + 1;
      if (UniquesCntWithIdxLaneV == UniquesCntWithOpIdxLaneV)
        return 0;
      return (PowerOf2Ceil(UniquesCntWithOpIdxLaneV) -
              UniquesCntWithOpIdxLaneV) -
             (PowerOf2Ceil(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 const int ScoreScaleFactor = 10;
 
    /// \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) {
      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);
        if (Score <= -SplatScore) {
          // Set the minimum score for splat-like sequence to avoid setting
          // failed state.
          Score = 1;
        } 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.
          Score *= ScoreScaleFactor;
          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) {
      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);
          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
      // lowest 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 = HashMap.find(NumFreeOpsHash.Hash);
          if (It == HashMap.end())
            HashMap[NumFreeOpsHash.Hash] = std::make_pair(1, Lane);
          else
            ++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).getOpcode() ||
              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 appendOperandsOfVL(ArrayRef<Value *> VL) {
      assert(!VL.empty() && "Bad VL");
      assert((empty() || VL.size() == getNumLanes()) &&
             "Expected same number of lanes");
      assert(isa<Instruction>(VL[0]) && "Expected instruction");
      unsigned NumOperands = cast<Instruction>(VL[0])->getNumOperands();
      constexpr unsigned IntrinsicNumOperands = 2;
      if (isa<IntrinsicInst>(VL[0]))
        NumOperands = IntrinsicNumOperands;
      OpsVec.resize(NumOperands);
      unsigned NumLanes = VL.size();
      for (unsigned OpIdx = 0; OpIdx != NumOperands; ++OpIdx) {
        OpsVec[OpIdx].resize(NumLanes);
        for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
          assert(isa<Instruction>(VL[Lane]) && "Expected instruction");
          // 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 VL[Lane] 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 VL[Lane] 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.
          bool IsInverseOperation = !isCommutative(cast<Instruction>(VL[Lane]));
          bool APO = (OpIdx == 0) ? false : IsInverseOperation;
          OpsVec[OpIdx][Lane] = {cast<Instruction>(VL[Lane])->getOperand(OpIdx),
                                 APO, false};
        }
      }
    }
 
    /// \returns the number of operands.
    unsigned getNumOperands() const { return OpsVec.size(); }
 
    /// \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) {
      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)
                      .getOpcode() &&
                 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).getOpcode() &&
               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 {
      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).getOpcode() &&
                      Op->getParent() == cast<Instruction>(OpILn)->getParent());
            }))
          return true;
      }
      return false;
    }
 
  public:
    /// Initialize with all the operands of the instruction vector \p RootVL.
    VLOperands(ArrayRef<Value *> RootVL, const BoUpSLP &R)
        : TLI(*R.TLI), DL(*R.DL), SE(*R.SE), R(R),
          L(R.LI->getLoopFor(
              (cast<Instruction>(RootVL.front())->getParent()))) {
      // Append all the operands of RootVL.
      appendOperandsOfVL(RootVL);
    }
 
    /// \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 (isa<LoadInst>(OpLane0))
          ReorderingModes[OpIdx] = ReorderingMode::Load;
        else 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
            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
          // NOTE: This should be unreachable.
          ReorderingModes[OpIdx] = ReorderingMode::Failed;
      }
 
      // 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 or non-power of 2 number of scalars (just for
      // now).
      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 : drop_begin(OpsVec, 1)) {
          if (any_of(Op, [&UniqueValues](const OperandData &Data) {
                return !UniqueValues.contains(Data.V);
              }))
            return false;
        }
        // TODO: Check if we can remove a check for non-power-2 number of
        // scalars after full support of non-power-2 vectorization.
        return UniqueValues.size() != 2 && isPowerOf2_32(UniqueValues.size());
      };
 
      // 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);
 
        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;
            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]);
              // 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.getOpcode() && 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::optional<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,
                                               /*Level=*/1, std::nullopt);
      if (Score > BestScore) {
        BestScore = Score;
        Index = I;
      }
    }
    return Index;
  }
 
  /// Checks if the instruction is marked for deletion.
  bool isDeleted(Instruction *I) const { return DeletedInstructions.count(I); }
 
  /// 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) {
    SmallVector<WeakTrackingVH> DeadInsts;
    for (T *V : DeadVals) {
      auto *I = cast<Instruction>(V);
      DeletedInstructions.insert(I);
    }
    DenseSet<Value *> Processed;
    for (T *V : DeadVals) {
      if (!V || !Processed.insert(V).second)
        continue;
      auto *I = cast<Instruction>(V);
      salvageDebugInfo(*I);
      SmallVector<const TreeEntry *> Entries;
      if (const TreeEntry *Entry = getTreeEntry(I)) {
        Entries.push_back(Entry);
        auto It = MultiNodeScalars.find(I);
        if (It != MultiNodeScalars.end())
          Entries.append(It->second.begin(), It->second.end());
      }
      for (Use &U : I->operands()) {
        if (auto *OpI = dyn_cast_if_present<Instruction>(U.get());
            OpI && !DeletedInstructions.contains(OpI) && OpI->hasOneUser() &&
            wouldInstructionBeTriviallyDead(OpI, TLI) &&
            (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) &&
              isInstructionTriviallyDead(OpI, TLI))
            DeadInsts.push_back(OpI);
      }
 
      VI->removeFromParent();
      DeletedInstructions.insert(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));
  }
  /// Clear the list of the analyzed reduction root instructions.
  void clearReductionData() {
    AnalyzedReductionsRoots.clear();
    AnalyzedReductionVals.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 the value is vectorized in the tree.
  bool isVectorized(Value *V) const { return getTreeEntry(V); }
 
  ~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,
                             unsigned &MaxDepthLevel,
                             bool &IsProfitableToDemote,
                             bool IsTruncRoot) const;
 
  /// Check if the operands on the edges \p Edges of the \p UserTE allows
  /// 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.
  bool
  canReorderOperands(TreeEntry *UserTE,
                     SmallVectorImpl<std::pair<unsigned, TreeEntry *>> &Edges,
                     ArrayRef<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;
 
  /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph,
  /// if any. If it is not vectorized (gather node), returns nullptr.
  TreeEntry *getVectorizedOperand(TreeEntry *UserTE, unsigned OpIdx) {
    ArrayRef<Value *> VL = UserTE->getOperand(OpIdx);
    TreeEntry *TE = nullptr;
    const auto *It = find_if(VL, [&](Value *V) {
      TE = getTreeEntry(V);
      if (TE && is_contained(TE->UserTreeIndices, EdgeInfo(UserTE, OpIdx)))
        return true;
      auto It = MultiNodeScalars.find(V);
      if (It != MultiNodeScalars.end()) {
        for (TreeEntry *E : It->second) {
          if (is_contained(E->UserTreeIndices, EdgeInfo(UserTE, OpIdx))) {
            TE = E;
            return true;
          }
        }
      }
      return false;
    });
    if (It != VL.end()) {
      assert(TE->isSame(VL) && "Expected same scalars.");
      return TE;
    }
    return nullptr;
  }
 
  /// Returns vectorized operand \p OpIdx of the node \p UserTE from the graph,
  /// if any. If it is not vectorized (gather node), returns nullptr.
  const TreeEntry *getVectorizedOperand(const TreeEntry *UserTE,
                                        unsigned OpIdx) const {
    return const_cast<BoUpSLP *>(this)->getVectorizedOperand(
        const_cast<TreeEntry *>(UserTE), OpIdx);
  }
 
  /// Checks if all users of \p I are the part of the vectorization tree.
  bool areAllUsersVectorized(
      Instruction *I,
      const SmallDenseSet<Value *> *VectorizedVals = nullptr) const;
 
  /// Return information about the vector formed for the specified index
  /// of a vector of (the same) instruction.
  TargetTransformInfo::OperandValueInfo getOperandInfo(ArrayRef<Value *> Ops);
 
  /// \ returns the graph entry for the \p Idx operand of the \p E entry.
  const TreeEntry *getOperandEntry(const TreeEntry *E, unsigned Idx) const;
 
  /// \returns Cast context for the given graph node.
  TargetTransformInfo::CastContextHint
  getCastContextHint(const TreeEntry &TE) const;
 
  /// \returns the cost of the vectorizable entry.
  InstructionCost getEntryCost(const TreeEntry *E,
                               ArrayRef<Value *> VectorizedVals,
                               SmallPtrSetImpl<Value *> &CheckedExtracts);
 
  /// This is the recursive part of buildTree.
  void buildTree_rec(ArrayRef<Value *> Roots, unsigned Depth,
                     const EdgeInfo &EI);
 
  /// \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, Value *OpValue,
                       SmallVectorImpl<unsigned> &CurrentOrder,
                       bool ResizeAllowed = false) const;
 
  /// Vectorize a single entry in the tree.
  /// \param PostponedPHIs true, if need to postpone emission of phi nodes to
  /// avoid issues with def-use order.
  Value *vectorizeTree(TreeEntry *E, bool PostponedPHIs);
 
  /// Vectorize a single entry in the tree, the \p Idx-th operand of the entry
  /// \p E.
  /// \param PostponedPHIs true, if need to postpone emission of phi nodes to
  /// avoid issues with def-use order.
  Value *vectorizeOperand(TreeEntry *E, unsigned NodeIdx, bool PostponedPHIs);
 
  /// 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);
 
  /// 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 scalarization cost for this list of values. Assuming that
  /// this subtree gets vectorized, we may need to extract the values from the
  /// roots. This method calculates the cost of extracting the values.
  /// \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);
 
  /// \returns whether the VectorizableTree is fully vectorizable and will
  /// be beneficial even the tree height is tiny.
  bool isFullyVectorizableTinyTree(bool ForReduction) const;
 
  /// Reorder commutative or alt operands to get better probability of
  /// generating vectorized code.
  static void reorderInputsAccordingToOpcode(ArrayRef<Value *> VL,
                                             SmallVectorImpl<Value *> &Left,
                                             SmallVectorImpl<Value *> &Right,
                                             const BoUpSLP &R);
 
  /// 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.
  DenseMap<Value *, 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;
 
  struct TreeEntry {
    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 {
      SmallVector<int> Mask;
      inversePermutation(ReorderIndices, Mask);
      ::addMask(Mask, ReuseShuffleIndices);
      return Mask;
    }
 
    /// \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<UndefValue>(V) &&
                                    Idx == PoisonMaskElem) ||
                                   (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);
    }
 
    bool isOperandGatherNode(const EdgeInfo &UserEI) const {
      return isGather() && UserTreeIndices.front().EdgeIdx == UserEI.EdgeIdx &&
             UserTreeIndices.front().UserTE == UserEI.UserTE;
    }
 
    /// \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;
 
    /// New vector phi instructions emitted for the vectorized phi nodes.
    PHINode *PHI = 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)
      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.
    };
    EntryState State;
 
    /// List of combined opcodes supported by the vectorizer.
    enum CombinedOpcode {
      NotCombinedOp = -1,
      MinMax = Instruction::OtherOpsEnd + 1,
    };
    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.  We can actually
    /// have multiple users so the data structure is not truly a tree.
    SmallVector<EdgeInfo, 1> UserTreeIndices;
 
    /// The index of this treeEntry in VectorizableTree.
    int Idx = -1;
 
  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 buildTree_rec() and vectorizeTree().
    SmallVector<ValueList, 2> Operands;
 
    /// The main/alternate instruction.
    Instruction *MainOp = nullptr;
    Instruction *AltOp = nullptr;
 
  public:
    /// 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());
    }
 
    /// Set the operands of this bundle in their original order.
    void setOperandsInOrder() {
      assert(Operands.empty() && "Already initialized?");
      auto *I0 = cast<Instruction>(Scalars[0]);
      Operands.resize(I0->getNumOperands());
      unsigned NumLanes = Scalars.size();
      for (unsigned OpIdx = 0, NumOperands = I0->getNumOperands();
           OpIdx != NumOperands; ++OpIdx) {
        Operands[OpIdx].resize(NumLanes);
        for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
          auto *I = cast<Instruction>(Scalars[Lane]);
          assert(I->getNumOperands() == NumOperands &&
                 "Expected same number of operands");
          Operands[OpIdx][Lane] = I->getOperand(OpIdx);
        }
      }
    }
 
    /// 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 MainOp != AltOp; }
 
    bool isOpcodeOrAlt(Instruction *I) const {
      unsigned CheckedOpcode = I->getOpcode();
      return (getOpcode() == CheckedOpcode ||
              getAltOpcode() == CheckedOpcode);
    }
 
    /// 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 && isOpcodeOrAlt(I))
        return Op;
      return MainOp;
    }
 
    void setOperations(const InstructionsState &S) {
      MainOp = S.MainOp;
      AltOp = S.AltOp;
    }
 
    Instruction *getMainOp() const {
      return MainOp;
    }
 
    Instruction *getAltOp() const {
      return AltOp;
    }
 
    /// The main/alternate opcodes for the list of instructions.
    unsigned getOpcode() const {
      return MainOp ? MainOp->getOpcode() : 0;
    }
 
    unsigned getAltOpcode() const {
      return AltOp ? AltOp->getOpcode() : 0;
    }
 
    /// When ReuseReorderShuffleIndices is empty it just returns position of \p
    /// V within vector of Scalars. Otherwise, try to remap on its reuse index.
    int findLaneForValue(Value *V) const {
      unsigned FoundLane = std::distance(Scalars.begin(), find(Scalars, V));
      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()) {
        FoundLane = std::distance(ReuseShuffleIndices.begin(),
                                  find(ReuseShuffleIndices, FoundLane));
      }
      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 = !isPowerOf2_32(Scalars.size());
      assert((!IsNonPowerOf2 || ReuseShuffleIndices.empty()) &&
             "Reshuffling not supported with non-power-of-2 vectors yet.");
      return IsNonPowerOf2;
    }
 
#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 << "\n";
      dbgs() << "State: ";
      switch (State) {
      case Vectorize:
        dbgs() << "Vectorize\n";
        break;
      case ScatterVectorize:
        dbgs() << "ScatterVectorize\n";
        break;
      case StridedVectorize:
        dbgs() << "StridedVectorize\n";
        break;
      case NeedToGather:
        dbgs() << "NeedToGather\n";
        break;
      case CombinedVectorize:
        dbgs() << "CombinedVectorize\n";
        break;
      }
      dbgs() << "MainOp: ";
      if (MainOp)
        dbgs() << *MainOp << "\n";
      else
        dbgs() << "NULL\n";
      dbgs() << "AltOp: ";
      if (AltOp)
        dbgs() << *AltOp << "\n";
      else
        dbgs() << "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() << "UserTreeIndices: ";
      for (const auto &EInfo : UserTreeIndices)
        dbgs() << EInfo << ", ";
      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 VectorizableTree entry.
  TreeEntry *newTreeEntry(ArrayRef<Value *> VL,
                          std::optional<ScheduleData *> Bundle,
                          const InstructionsState &S,
                          const EdgeInfo &UserTreeIdx,
                          ArrayRef<int> ReuseShuffleIndices = std::nullopt,
                          ArrayRef<unsigned> ReorderIndices = std::nullopt) {
    TreeEntry::EntryState EntryState =
        Bundle ? TreeEntry::Vectorize : TreeEntry::NeedToGather;
    return newTreeEntry(VL, EntryState, Bundle, S, UserTreeIdx,
                        ReuseShuffleIndices, ReorderIndices);
  }
 
  TreeEntry *newTreeEntry(ArrayRef<Value *> VL,
                          TreeEntry::EntryState EntryState,
                          std::optional<ScheduleData *> Bundle,
                          const InstructionsState &S,
                          const EdgeInfo &UserTreeIdx,
                          ArrayRef<int> ReuseShuffleIndices = std::nullopt,
                          ArrayRef<unsigned> ReorderIndices = std::nullopt) {
    assert(((!Bundle && EntryState == TreeEntry::NeedToGather) ||
            (Bundle && EntryState != TreeEntry::NeedToGather)) &&
           "Need to vectorize gather entry?");
    VectorizableTree.push_back(std::make_unique<TreeEntry>(VectorizableTree));
    TreeEntry *Last = VectorizableTree.back().get();
    Last->Idx = VectorizableTree.size() - 1;
    Last->State = EntryState;
    Last->ReuseShuffleIndices.append(ReuseShuffleIndices.begin(),
                                     ReuseShuffleIndices.end());
    if (ReorderIndices.empty()) {
      Last->Scalars.assign(VL.begin(), VL.end());
      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);
      Last->setOperations(S);
      Last->ReorderIndices.append(ReorderIndices.begin(), ReorderIndices.end());
    }
    if (!Last->isGather()) {
      for (Value *V : VL) {
        const TreeEntry *TE = getTreeEntry(V);
        assert((!TE || TE == Last || doesNotNeedToBeScheduled(V)) &&
               "Scalar already in tree!");
        if (TE) {
          if (TE != Last)
            MultiNodeScalars.try_emplace(V).first->getSecond().push_back(Last);
          continue;
        }
        ScalarToTreeEntry[V] = Last;
      }
      // Update the scheduler bundle to point to this TreeEntry.
      ScheduleData *BundleMember = *Bundle;
      assert((BundleMember || isa<PHINode>(S.MainOp) ||
              isVectorLikeInstWithConstOps(S.MainOp) ||
              doesNotNeedToSchedule(VL)) &&
             "Bundle and VL out of sync");
      if (BundleMember) {
        for (Value *V : VL) {
          if (doesNotNeedToBeScheduled(V))
            continue;
          if (!BundleMember)
            continue;
          BundleMember->TE = Last;
          BundleMember = BundleMember->NextInBundle;
        }
      }
      assert(!BundleMember && "Bundle and VL out of sync");
    } else {
      // Build a map for gathered scalars to the nodes where they are used.
      bool AllConstsOrCasts = true;
      for (Value *V : VL)
        if (!isConstant(V)) {
          auto *I = dyn_cast<CastInst>(V);
          AllConstsOrCasts &= I && I->getType()->isIntegerTy();
          ValueToGatherNodes.try_emplace(V).first->getSecond().insert(Last);
        }
      if (AllConstsOrCasts)
        CastMaxMinBWSizes =
            std::make_pair(std::numeric_limits<unsigned>::max(), 1);
      MustGather.insert(VL.begin(), VL.end());
    }
 
    if (UserTreeIdx.UserTE) {
      Last->UserTreeIndices.push_back(UserTreeIdx);
      assert((!Last->isNonPowOf2Vec() || Last->ReorderIndices.empty()) &&
             "Reordering isn't implemented for non-power-of-2 nodes yet");
    }
    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();
      dbgs() << "\n";
    }
  }
#endif
 
  TreeEntry *getTreeEntry(Value *V) { return ScalarToTreeEntry.lookup(V); }
 
  const TreeEntry *getTreeEntry(Value *V) const {
    return ScalarToTreeEntry.lookup(V);
  }
 
  /// Check that the operand node of alternate node does not generate
  /// buildvector sequence. If it is, then probably not worth it to build
  /// alternate shuffle, if number of buildvector operands + alternate
  /// instruction > than the number of buildvector instructions.
  /// \param S the instructions state of the analyzed values.
  /// \param VL list of the instructions with alternate opcodes.
  bool areAltOperandsProfitable(const InstructionsState &S,
                                ArrayRef<Value *> VL) 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(
      InstructionsState &S, ArrayRef<Value *> VL, bool IsScatterVectorizeUserTE,
      OrdersType &CurrentOrder, SmallVectorImpl<Value *> &PointerOps) const;
 
  /// Maps a specific scalar to its tree entry.
  SmallDenseMap<Value *, TreeEntry *> ScalarToTreeEntry;
 
  /// List of scalars, used in several vectorize nodes, and the list of the
  /// nodes.
  SmallDenseMap<Value *, SmallVector<TreeEntry *>> MultiNodeScalars;
 
  /// 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;
 
  /// 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.
  DenseMap<const TreeEntry *, Instruction *> EntryToLastInstruction;
 
  /// 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 *, SmallPtrSet<const TreeEntry *, 4>>;
  ValueToGatherNodesMap ValueToGatherNodes;
 
  /// This POD struct describes one external user in the vectorized tree.
  struct ExternalUser {
    ExternalUser(Value *S, llvm::User *U, int L)
        : Scalar(S), User(U), Lane(L) {}
 
    // Which scalar in our function.
    Value *Scalar;
 
    // Which user that uses the scalar.
    llvm::User *User;
 
    // Which lane does the scalar belong to.
    int 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) {
    if (!Loc1.Ptr || !isSimple(Inst1) || !isSimple(Inst2))
      return true;
    // First check if the result is already in the cache.
    AliasCacheKey Key = std::make_pair(Inst1, Inst2);
    auto It = AliasCache.find(Key);
    if (It != AliasCache.end())
      return It->second;
    bool Aliased = isModOrRefSet(BatchAA.getModRefInfo(Inst2, Loc1));
    // Store the result in the cache.
    AliasCache.try_emplace(Key, Aliased);
    AliasCache.try_emplace(std::make_pair(Inst2, Inst1), Aliased);
    return Aliased;
  }
 
  using AliasCacheKey = std::pair<Instruction *, Instruction *>;
 
  /// Cache for alias results.
  /// TODO: consider moving this to the AliasAnalysis itself.
  DenseMap<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;
 
  /// 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> ExternalUsesAsGEPs;
 
  /// 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;
 
  /// 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).
  struct ScheduleData {
    // The initial value for the dependency counters. It means that the
    // dependencies are not calculated yet.
    enum { InvalidDeps = -1 };
 
    ScheduleData() = default;
 
    void init(int BlockSchedulingRegionID, Value *OpVal) {
      FirstInBundle = this;
      NextInBundle = nullptr;
      NextLoadStore = nullptr;
      IsScheduled = false;
      SchedulingRegionID = BlockSchedulingRegionID;
      clearDependencies();
      OpValue = OpVal;
      TE = nullptr;
    }
 
    /// Verify basic self consistency properties
    void verify() {
      if (hasValidDependencies()) {
        assert(UnscheduledDeps <= Dependencies && "invariant");
      } else {
        assert(UnscheduledDeps == Dependencies && "invariant");
      }
 
      if (IsScheduled) {
        assert(isSchedulingEntity() &&
                "unexpected scheduled state");
        for (const ScheduleData *BundleMember = this; BundleMember;
             BundleMember = BundleMember->NextInBundle) {
          assert(BundleMember->hasValidDependencies() &&
                 BundleMember->UnscheduledDeps == 0 &&
                 "unexpected scheduled state");
          assert((BundleMember == this || !BundleMember->IsScheduled) &&
                 "only bundle is marked scheduled");
        }
      }
 
      assert(Inst->getParent() == FirstInBundle->Inst->getParent() &&
             "all bundle members must be in same basic block");
    }
 
    /// 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 for single instructions and for bundle representatives
    /// (= the head of a bundle).
    bool isSchedulingEntity() const { return FirstInBundle == this; }
 
    /// Returns true if it represents an instruction bundle and not only a
    /// single instruction.
    bool isPartOfBundle() const {
      return NextInBundle != nullptr || FirstInBundle != this || TE;
    }
 
    /// Returns true if it is ready for scheduling, i.e. it has no more
    /// unscheduled depending instructions/bundles.
    bool isReady() const {
      assert(isSchedulingEntity() &&
             "can't consider non-scheduling entity for ready list");
      return unscheduledDepsInBundle() == 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;
      return FirstInBundle->unscheduledDepsInBundle();
    }
 
    /// Sets the number of unscheduled dependencies to the number of
    /// dependencies.
    void resetUnscheduledDeps() {
      UnscheduledDeps = Dependencies;
    }
 
    /// Clears all dependency information.
    void clearDependencies() {
      Dependencies = InvalidDeps;
      resetUnscheduledDeps();
      MemoryDependencies.clear();
      ControlDependencies.clear();
    }
 
    int unscheduledDepsInBundle() const {
      assert(isSchedulingEntity() && "only meaningful on the bundle");
      int Sum = 0;
      for (const ScheduleData *BundleMember = this; BundleMember;
           BundleMember = BundleMember->NextInBundle) {
        if (BundleMember->UnscheduledDeps == InvalidDeps)
          return InvalidDeps;
        Sum += BundleMember->UnscheduledDeps;
      }
      return Sum;
    }
 
    void dump(raw_ostream &os) const {
      if (!isSchedulingEntity()) {
        os << "/ " << *Inst;
      } else if (NextInBundle) {
        os << '[' << *Inst;
        ScheduleData *SD = NextInBundle;
        while (SD) {
          os << ';' << *SD->Inst;
          SD = SD->NextInBundle;
        }
        os << ']';
      } else {
        os << *Inst;
      }
    }
 
    Instruction *Inst = nullptr;
 
    /// Opcode of the current instruction in the schedule data.
    Value *OpValue = nullptr;
 
    /// The TreeEntry that this instruction corresponds to.
    TreeEntry *TE = nullptr;
 
    /// Points to the head in an instruction bundle (and always to this for
    /// single instructions).
    ScheduleData *FirstInBundle = nullptr;
 
    /// Single linked list of all instructions in a bundle. Null if it is a
    /// single instruction.
    ScheduleData *NextInBundle = 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 *, 4> 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 *, 4> ControlDependencies;
 
    /// This ScheduleData is in the current scheduling region if this matches
    /// the current SchedulingRegionID of BlockScheduling.
    int SchedulingRegionID = 0;
 
    /// Used for getting a "good" final ordering of instructions.
    int SchedulingPriority = 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;
 
    /// True if this instruction is scheduled (or considered as scheduled in the
    /// dry-run).
    bool IsScheduled = false;
  };
 
#ifndef NDEBUG
  friend inline raw_ostream &operator<<(raw_ostream &os,
                                        const BoUpSLP::ScheduleData &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() {
      ReadyInsts.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 (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) {
      if (auto *I = dyn_cast<Instruction>(V))
        return getScheduleData(I);
      return nullptr;
    }
 
    ScheduleData *getScheduleData(Value *V, Value *Key) {
      if (V == Key)
        return getScheduleData(V);
      auto I = ExtraScheduleDataMap.find(V);
      if (I != ExtraScheduleDataMap.end()) {
        ScheduleData *SD = I->second.lookup(Key);
        if (SD && isInSchedulingRegion(SD))
          return SD;
      }
      return nullptr;
    }
 
    bool isInSchedulingRegion(ScheduleData *SD) const {
      return SD->SchedulingRegionID == SchedulingRegionID;
    }
 
    /// Marks an instruction as scheduled and puts all dependent ready
    /// instructions into the ready-list.
    template <typename ReadyListType>
    void schedule(ScheduleData *SD, ReadyListType &ReadyList) {
      SD->IsScheduled = true;
      LLVM_DEBUG(dbgs() << "SLP:   schedule " << *SD << "\n");
 
      for (ScheduleData *BundleMember = SD; BundleMember;
           BundleMember = BundleMember->NextInBundle) {
        if (BundleMember->Inst != BundleMember->OpValue)
          continue;
 
        // Handle the def-use chain dependencies.
 
        // Decrement the unscheduled counter and insert to ready list if ready.
        auto &&DecrUnsched = [this, &ReadyList](Instruction *I) {
          doForAllOpcodes(I, [&ReadyList](ScheduleData *OpDef) {
            if (OpDef && OpDef->hasValidDependencies() &&
                OpDef->incrementUnscheduledDeps(-1) == 0) {
              // There are no more unscheduled dependencies after
              // decrementing, so we can put the dependent instruction
              // into the ready list.
              ScheduleData *DepBundle = OpDef->FirstInBundle;
              assert(!DepBundle->IsScheduled &&
                     "already scheduled bundle gets ready");
              ReadyList.insert(DepBundle);
              LLVM_DEBUG(dbgs()
                         << "SLP:    gets ready (def): " << *DepBundle << "\n");
            }
          });
        };
 
        // 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 (TreeEntry *TE = BundleMember->TE) {
          // Need to search for the lane since the tree entry can be reordered.
          int Lane = std::distance(TE->Scalars.begin(),
                                   find(TE->Scalars, BundleMember->Inst));
          assert(Lane >= 0 && "Lane not set");
 
          // 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.
          auto *In = BundleMember->Inst;
          assert(
              In &&
              (isa<ExtractValueInst, ExtractElementInst, IntrinsicInst>(In) ||
               In->getNumOperands() == TE->getNumOperands()) &&
              "Missed TreeEntry operands?");
          (void)In; // fake use to avoid build failure when assertions disabled
 
          for (unsigned OpIdx = 0, NumOperands = TE->getNumOperands();
               OpIdx != NumOperands; ++OpIdx)
            if (auto *I = dyn_cast<Instruction>(TE->getOperand(OpIdx)[Lane]))
              DecrUnsched(I);
        } else {
          // If BundleMember is a stand-alone instruction, no operand reordering
          // has taken place, so we directly access its operands.
          for (Use &U : BundleMember->Inst->operands())
            if (auto *I = dyn_cast<Instruction>(U.get()))
              DecrUnsched(I);
        }
        // Handle the memory dependencies.
        for (ScheduleData *MemoryDepSD : BundleMember->MemoryDependencies) {
          if (MemoryDepSD->hasValidDependencies() &&
              MemoryDepSD->incrementUnscheduledDeps(-1) == 0) {
            // There are no more unscheduled dependencies after decrementing,
            // so we can put the dependent instruction into the ready list.
            ScheduleData *DepBundle = MemoryDepSD->FirstInBundle;
            assert(!DepBundle->IsScheduled &&
                   "already scheduled bundle gets ready");
            ReadyList.insert(DepBundle);
            LLVM_DEBUG(dbgs()
                       << "SLP:    gets ready (mem): " << *DepBundle << "\n");
          }
        }
        // Handle the control dependencies.
        for (ScheduleData *DepSD : BundleMember->ControlDependencies) {
          if (DepSD->incrementUnscheduledDeps(-1) == 0) {
            // There are no more unscheduled dependencies after decrementing,
            // so we can put the dependent instruction into the ready list.
            ScheduleData *DepBundle = DepSD->FirstInBundle;
            assert(!DepBundle->IsScheduled &&
                   "already scheduled bundle gets ready");
            ReadyList.insert(DepBundle);
            LLVM_DEBUG(dbgs()
                       << "SLP:    gets ready (ctl): " << *DepBundle << "\n");
          }
        }
      }
    }
 
    /// 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()) {
        auto *SD = getScheduleData(I);
        if (!SD)
          continue;
        assert(isInSchedulingRegion(SD) &&
               "primary schedule data not in window?");
        assert(isInSchedulingRegion(SD->FirstInBundle) &&
               "entire bundle in window!");
        (void)SD;
        doForAllOpcodes(I, [](ScheduleData *SD) { SD->verify(); });
      }
 
      for (auto *SD : ReadyInsts) {
        assert(SD->isSchedulingEntity() && SD->isReady() &&
               "item in ready list not ready?");
        (void)SD;
      }
    }
 
    void doForAllOpcodes(Value *V,
                         function_ref<void(ScheduleData *SD)> Action) {
      if (ScheduleData *SD = getScheduleData(V))
        Action(SD);
      auto I = ExtraScheduleDataMap.find(V);
      if (I != ExtraScheduleDataMap.end())
        for (auto &P : I->second)
          if (isInSchedulingRegion(P.second))
            Action(P.second);
    }
 
    /// Put all instructions into the ReadyList which are ready for scheduling.
    template <typename ReadyListType>
    void initialFillReadyList(ReadyListType &ReadyList) {
      for (auto *I = ScheduleStart; I != ScheduleEnd; I = I->getNextNode()) {
        doForAllOpcodes(I, [&](ScheduleData *SD) {
          if (SD->isSchedulingEntity() && SD->hasValidDependencies() &&
              SD->isReady()) {
            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.
    ScheduleData *buildBundle(ArrayRef<Value *> VL);
 
    /// 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<ScheduleData *>
    tryScheduleBundle(ArrayRef<Value *> VL, BoUpSLP *SLP,
                      const InstructionsState &S);
 
    /// Un-bundles a group of instructions.
    void cancelScheduling(ArrayRef<Value *> VL, Value *OpValue);
 
    /// 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(ScheduleData *SD, bool InsertInReadyList,
                               BoUpSLP *SLP);
 
    /// 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.
    DenseMap<Instruction *, ScheduleData *> ScheduleDataMap;
 
    /// Attaches ScheduleData to Instruction with the leading key.
    DenseMap<Value *, SmallDenseMap<Value *, ScheduleData *>>
        ExtraScheduleDataMap;
 
    /// The ready-list for scheduling (only used for the dry-run).
    SetVector<ScheduleData *> 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;
 
    /// 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(BlockScheduling *BS);
 
  /// List of users to ignore during scheduling and that don't need extracting.
  const SmallDenseSet<Value *> *UserIgnoreList = nullptr;
 
  /// A DenseMapInfo implementation for holding DenseMaps and DenseSets of
  /// sorted SmallVectors of unsigned.
  struct OrdersTypeDenseMapInfo {
    static OrdersType getEmptyKey() {
      OrdersType V;
      V.push_back(~1U);
      return V;
    }
 
    static OrdersType getTombstoneKey() {
      OrdersType V;
      V.push_back(~2U);
      return V;
    }
 
    static unsigned getHashValue(const OrdersType &V) {
      return static_cast<unsigned>(hash_combine_range(V.begin(), V.end()));
    }
 
    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;
 
  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;
 
  /// 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;
};
 
} // end namespace slpvectorizer
 
template <> struct 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.VectorizableTree[0].get();
  }
 
  static ChildIteratorType child_begin(NodeRef N) {
    return {N->UserTreeIndices.begin(), N->Container};
  }
 
  static ChildIteratorType child_end(NodeRef N) {
    return {N->UserTreeIndices.end(), 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 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)
      return "color=blue";
    return "";
  }
};
 
} // end namespace llvm
 
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());
      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) {
  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();
  int NumScalars = GatheredScalars.size();
  if (!isValidElementType(ScalarTy))
    return std::nullopt;
  auto *VecTy = getWidenedType(ScalarTy, NumScalars);
  int NumParts = TTI->getNumberOfParts(VecTy);
  if (NumParts == 0 || NumParts >= NumScalars)
    NumParts = 1;
  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)) {
    // 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>(0, 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, [&](int I) { return I != NumScalars; })) {
        std::fill(Slice.begin(), Slice.end(), 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) {
        std::fill(Slice.begin(), Slice.end(), 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) {
          SecondVecFound = true;
          break;
        }
        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) {
        std::fill(Slice.begin(), Slice.end(), NumScalars);
        ShuffledSubMasks.set(I);
        continue;
      }
    }
  };
  int PartSz = getPartNumElems(NumScalars, NumParts);
  if (!ExtractShuffles.empty())
    TransformMaskToOrder(
        CurrentOrder, ExtractMask, PartSz, NumParts, [&](unsigned I) {
          if (!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 (ExtractMask[K] == PoisonMaskElem)
              continue;
            if (!TE.ReuseShuffleIndices.empty())
              K = TE.ReuseShuffleIndices[K];
            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, cast<VectorType>(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 (!GatherShuffles[I])
        return 0U;
      return std::max(Entries[I].front()->getVectorFactor(),
                      Entries[I].back()->getVectorFactor());
    });
  int NumUndefs =
      count_if(CurrentOrder, [&](int Idx) { return Idx == 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) != getUnderlyingObject(Ptr2))
    return false;
  auto *GEP1 = dyn_cast<GetElementPtrInst>(Ptr1);
  if (!GEP1)
    return false;
  auto *GEP2 = dyn_cast<GetElementPtrInst>(Ptr2);
  if (!GEP2)
    return false;
  return GEP1->getNumOperands() == 2 && GEP2->getNumOperands() == 2 &&
         ((isConstant(GEP1->getOperand(1)) &&
           isConstant(GEP2->getOperand(1))) ||
          !CompareOpcodes ||
          getSameOpcode({GEP1->getOperand(1), GEP2->getOperand(1)}, TLI)
              .getOpcode());
}
 
/// Calculates minimal alignment as a common alignment.
template <typename T>
static Align computeCommonAlignment(ArrayRef<Value *> VL) {
  Align CommonAlignment = cast<T>(VL.front())->getAlign();
  for (Value *V : VL.drop_front())
    CommonAlignment = std::min(CommonAlignment, cast<T>(V)->getAlign());
  return CommonAlignment;
}
 
/// Check if \p Order represents reverse order.
static bool isReverseOrder(ArrayRef<unsigned> Order) {
  unsigned Sz = Order.size();
  return !Order.empty() && 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, std::nullopt is returned.
/// Otherwise, if \p Inst is not specified, just initialized optional value is
/// returned to show that the pointers represent strided pointers. If \p Inst
/// specified, the runtime stride is materialized before the given \p Inst.
/// \returns std::nullopt if the pointers are not pointers with the runtime
/// stride, nullptr or actual stride value, otherwise.
static std::optional<Value *>
calculateRtStride(ArrayRef<Value *> PointerOps, Type *ElemTy,
                  const DataLayout &DL, ScalarEvolution &SE,
                  SmallVectorImpl<unsigned> &SortedIndices,
                  Instruction *Inst = nullptr) {
  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 std::nullopt;
    SCEVs.push_back(PtrSCEV);
    if (!PtrSCEVLowest && !PtrSCEVHighest) {
      PtrSCEVLowest = PtrSCEVHighest = PtrSCEV;
      continue;
    }
    const SCEV *Diff = SE.getMinusSCEV(PtrSCEV, PtrSCEVLowest);
    if (isa<SCEVCouldNotCompute>(Diff))
      return std::nullopt;
    if (Diff->isNonConstantNegative()) {
      PtrSCEVLowest = PtrSCEV;
      continue;
    }
    const SCEV *Diff1 = SE.getMinusSCEV(PtrSCEVHighest, PtrSCEV);
    if (isa<SCEVCouldNotCompute>(Diff1))
      return std::nullopt;
    if (Diff1->isNonConstantNegative()) {
      PtrSCEVHighest = PtrSCEV;
      continue;
    }
  }
  // Dist = PtrSCEVHighest - PtrSCEVLowest;
  const SCEV *Dist = SE.getMinusSCEV(PtrSCEVHighest, PtrSCEVLowest);
  if (isa<SCEVCouldNotCompute>(Dist))
    return std::nullopt;
  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() > 2) {
    const SCEV *Sz = SE.getConstant(Dist->getType(), Size * (SCEVs.size() - 1));
    Stride = TryGetStride(Dist, Sz);
    if (!Stride)
      return std::nullopt;
  }
  if (!Stride || isa<SCEVConstant>(Stride))
    return std::nullopt;
  // 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);
  int Cnt = 0;
  bool IsConsecutive = true;
  for (const SCEV *PtrSCEV : SCEVs) {
    unsigned Dist = 0;
    if (PtrSCEV != PtrSCEVLowest) {
      const SCEV *Diff = SE.getMinusSCEV(PtrSCEV, PtrSCEVLowest);
      const SCEV *Coeff = TryGetStride(Diff, Stride);
      if (!Coeff)
        return std::nullopt;
      const auto *SC = dyn_cast<SCEVConstant>(Coeff);
      if (!SC || isa<SCEVCouldNotCompute>(SC))
        return std::nullopt;
      if (!SE.getMinusSCEV(PtrSCEV, SE.getAddExpr(PtrSCEVLowest,
                                                  SE.getMulExpr(Stride, SC)))
               ->isZero())
        return std::nullopt;
      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 std::nullopt;
    auto Res = Offsets.emplace(Dist, Cnt);
    if (!Res.second)
      return std::nullopt;
    // Consecutive order if the inserted element is the last one.
    IsConsecutive = IsConsecutive && std::next(Res.first) == Offsets.end();
    ++Cnt;
  }
  if (Offsets.size() != SCEVs.size())
    return std::nullopt;
  SortedIndices.clear();
  if (!IsConsecutive) {
    // Fill SortedIndices array only if it is non-consecutive.
    SortedIndices.resize(PointerOps.size());
    Cnt = 0;
    for (const std::pair<int64_t, int> &Pair : Offsets) {
      SortedIndices[Cnt] = Pair.second;
      ++Cnt;
    }
  }
  if (!Inst)
    return nullptr;
  SCEVExpander Expander(SE, DL, "strided-load-vec");
  return Expander.expandCodeFor(Stride, Stride->getType(), Inst);
}
 
static std::pair<InstructionCost, InstructionCost>
getGEPCosts(const TargetTransformInfo &TTI, ArrayRef<Value *> Ptrs,
            Value *BasePtr, unsigned Opcode, TTI::TargetCostKind CostKind,
            Type *ScalarTy, VectorType *VecTy);
 
BoUpSLP::LoadsState BoUpSLP::canVectorizeLoads(
    ArrayRef<Value *> VL, const Value *VL0, SmallVectorImpl<unsigned> &Order,
    SmallVectorImpl<Value *> &PointerOps, 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.
  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 unsigned Sz = VL.size();
  PointerOps.resize(Sz);
  auto *POIter = PointerOps.begin();
  for (Value *V : VL) {
    auto *L = cast<LoadInst>(V);
    if (!L->isSimple())
      return LoadsState::Gather;
    *POIter = L->getPointerOperand();
    ++POIter;
  }
 
  Order.clear();
  auto *VecTy = getWidenedType(ScalarTy, Sz);
  // Check the order of pointer operands or that all pointers are the same.
  bool IsSorted = sortPtrAccesses(PointerOps, ScalarTy, *DL, *SE, Order);
  // FIXME: Reordering isn't implemented for non-power-of-2 nodes yet.
  if (!Order.empty() && !isPowerOf2_32(VL.size())) {
    assert(VectorizeNonPowerOf2 && "non-power-of-2 number of loads only "
                                   "supported with VectorizeNonPowerOf2");
    return LoadsState::Gather;
  }
 
  Align CommonAlignment = computeCommonAlignment<LoadInst>(VL);
  if (!IsSorted && Sz > MinProfitableStridedLoads && TTI->isTypeLegal(VecTy) &&
      TTI->isLegalStridedLoadStore(VecTy, CommonAlignment) &&
      calculateRtStride(PointerOps, ScalarTy, *DL, *SE, Order))
    return LoadsState::StridedVectorize;
  if (IsSorted || all_of(PointerOps, [&](Value *P) {
        return arePointersCompatible(P, PointerOps.front(), *TLI);
      })) {
    if (IsSorted) {
      Value *Ptr0;
      Value *PtrN;
      if (Order.empty()) {
        Ptr0 = PointerOps.front();
        PtrN = PointerOps.back();
      } else {
        Ptr0 = PointerOps[Order.front()];
        PtrN = PointerOps[Order.back()];
      }
      std::optional<int> Diff =
          getPointersDiff(ScalarTy, Ptr0, ScalarTy, PtrN, *DL, *SE);
      // Check that the sorted loads are consecutive.
      if (static_cast<unsigned>(*Diff) == Sz - 1)
        return LoadsState::Vectorize;
      // Simple check if not a strided access - clear order.
      bool IsPossibleStrided = *Diff % (Sz - 1) == 0;
      // Try to generate strided load node if:
      // 1. Target with strided load support is detected.
      // 2. The number of loads is greater than MinProfitableStridedLoads,
      // or the potential stride <= MaxProfitableLoadStride 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).
      auto IsAnyPointerUsedOutGraph =
          IsPossibleStrided && any_of(PointerOps, [&](Value *V) {
            return isa<Instruction>(V) && any_of(V->users(), [&](User *U) {
                     return !getTreeEntry(U) && !MustGather.contains(U);
                   });
          });
      if (IsPossibleStrided && (IsAnyPointerUsedOutGraph ||
                                ((Sz > MinProfitableStridedLoads ||
                                  (static_cast<unsigned>(std::abs(*Diff)) <=
                                       MaxProfitableLoadStride * Sz &&
                                   isPowerOf2_32(std::abs(*Diff)))) &&
                                 static_cast<unsigned>(std::abs(*Diff)) > Sz) ||
                                *Diff == -(static_cast<int>(Sz) - 1))) {
        int Stride = *Diff / static_cast<int>(Sz - 1);
        if (*Diff == Stride * static_cast<int>(Sz - 1)) {
          Align Alignment =
              cast<LoadInst>(Order.empty() ? VL.front() : VL[Order.front()])
                  ->getAlign();
          if (TTI->isLegalStridedLoadStore(VecTy, Alignment)) {
            // Iterate through all pointers and check if all distances are
            // unique multiple of Dist.
            SmallSet<int, 4> Dists;
            for (Value *Ptr : PointerOps) {
              int Dist = 0;
              if (Ptr == PtrN)
                Dist = *Diff;
              else if (Ptr != Ptr0)
                Dist =
                    *getPointersDiff(ScalarTy, Ptr0, ScalarTy, Ptr, *DL, *SE);
              // If the strides are not the same or repeated, we can't
              // vectorize.
              if (((Dist / Stride) * Stride) != Dist ||
                  !Dists.insert(Dist).second)
                break;
            }
            if (Dists.size() == Sz)
              return LoadsState::StridedVectorize;
          }
        }
      }
    }
    auto CheckForShuffledLoads = [&, &TTI = *TTI](Align CommonAlignment) {
      unsigned Sz = DL->getTypeSizeInBits(ScalarTy);
      unsigned MinVF = getMinVF(Sz);
      unsigned MaxVF = std::max<unsigned>(bit_floor(VL.size() / 2), MinVF);
      MaxVF = std::min(getMaximumVF(Sz, Instruction::Load), MaxVF);
      for (unsigned VF = MaxVF; VF >= MinVF; VF /= 2) {
        unsigned VectorizedCnt = 0;
        SmallVector<LoadsState> States;
        for (unsigned Cnt = 0, End = VL.size(); Cnt + VF <= End;
             Cnt += VF, ++VectorizedCnt) {
          ArrayRef<Value *> Slice = VL.slice(Cnt, VF);
          SmallVector<unsigned> Order;
          SmallVector<Value *> PointerOps;
          LoadsState LS =
              canVectorizeLoads(Slice, Slice.front(), Order, PointerOps,
                                /*TryRecursiveCheck=*/false);
          // Check that the sorted loads are consecutive.
          if (LS == LoadsState::Gather)
            break;
          // If need the reorder - consider as high-cost masked gather for now.
          if ((LS == LoadsState::Vectorize ||
               LS == LoadsState::StridedVectorize) &&
              !Order.empty() && !isReverseOrder(Order))
            LS = LoadsState::ScatterVectorize;
          States.push_back(LS);
        }
        // Can be vectorized later as a serie of loads/insertelements.
        if (VectorizedCnt == VL.size() / VF) {
          // Compare masked gather cost and loads + insersubvector costs.
          TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
          auto [ScalarGEPCost, VectorGEPCost] = getGEPCosts(
              TTI, PointerOps, PointerOps.front(), Instruction::GetElementPtr,
              CostKind, ScalarTy, VecTy);
          InstructionCost MaskedGatherCost =
              TTI.getGatherScatterOpCost(
                  Instruction::Load, VecTy,
                  cast<LoadInst>(VL0)->getPointerOperand(),
                  /*VariableMask=*/false, CommonAlignment, CostKind) +
              VectorGEPCost - ScalarGEPCost;
          InstructionCost VecLdCost = 0;
          auto *SubVecTy = getWidenedType(ScalarTy, VF);
          for (auto [I, LS] : enumerate(States)) {
            auto *LI0 = cast<LoadInst>(VL[I * VF]);
            switch (LS) {
            case LoadsState::Vectorize: {
              auto [ScalarGEPCost, VectorGEPCost] =
                  getGEPCosts(TTI, ArrayRef(PointerOps).slice(I * VF, VF),
                              LI0->getPointerOperand(), Instruction::Load,
                              CostKind, ScalarTy, SubVecTy);
              VecLdCost += TTI.getMemoryOpCost(
                               Instruction::Load, SubVecTy, LI0->getAlign(),
                               LI0->getPointerAddressSpace(), CostKind,
                               TTI::OperandValueInfo()) +
                           VectorGEPCost - ScalarGEPCost;
              break;
            }
            case LoadsState::StridedVectorize: {
              auto [ScalarGEPCost, VectorGEPCost] =
                  getGEPCosts(TTI, ArrayRef(PointerOps).slice(I * VF, VF),
                              LI0->getPointerOperand(), Instruction::Load,
                              CostKind, ScalarTy, SubVecTy);
              VecLdCost +=
                  TTI.getStridedMemoryOpCost(
                      Instruction::Load, SubVecTy, LI0->getPointerOperand(),
                      /*VariableMask=*/false, CommonAlignment, CostKind) +
                  VectorGEPCost - ScalarGEPCost;
              break;
            }
            case LoadsState::ScatterVectorize: {
              auto [ScalarGEPCost, VectorGEPCost] = getGEPCosts(
                  TTI, ArrayRef(PointerOps).slice(I * VF, VF),
                  LI0->getPointerOperand(), Instruction::GetElementPtr,
                  CostKind, ScalarTy, SubVecTy);
              VecLdCost +=
                  TTI.getGatherScatterOpCost(
                      Instruction::Load, SubVecTy, LI0->getPointerOperand(),
                      /*VariableMask=*/false, CommonAlignment, CostKind) +
                  VectorGEPCost - ScalarGEPCost;
              break;
            }
            case LoadsState::Gather:
              llvm_unreachable(
                  "Expected only consecutive, strided or masked gather loads.");
            }
            SmallVector<int> ShuffleMask(VL.size());
            for (int Idx : seq<int>(0, VL.size()))
              ShuffleMask[Idx] = Idx / VF == I ? VL.size() + Idx % VF : Idx;
            VecLdCost +=
                TTI.getShuffleCost(TTI::SK_InsertSubvector, VecTy, ShuffleMask,
                                   CostKind, I * VF, 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)
            return true;
        }
      }
      return false;
    };
    // 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, [IsSorted](Value *P) {
          auto *GEP = dyn_cast<GetElementPtrInst>(P);
          return (IsSorted && !GEP && doesNotNeedToBeScheduled(P)) ||
                 (GEP && GEP->getNumOperands() == 2 &&
                  isa<Constant, Instruction>(GEP->getOperand(1)));
        })) {
      Align CommonAlignment = computeCommonAlignment<LoadInst>(VL);
      if (TTI->isLegalMaskedGather(VecTy, CommonAlignment) &&
          !TTI->forceScalarizeMaskedGather(VecTy, CommonAlignment)) {
        // Check if potential masked gather can be represented as series
        // of loads + insertsubvectors.
        if (TryRecursiveCheck && CheckForShuffledLoads(CommonAlignment)) {
          // If masked gather cost is higher - better to vectorize, so
          // consider it as a gather node. It will be better estimated
          // later.
          return LoadsState::Gather;
        }
        return LoadsState::ScatterVectorize;
      }
    }
  }
 
  return LoadsState::Gather;
}
 
static bool clusterSortPtrAccesses(ArrayRef<Value *> VL, Type *ElemTy,
                                   const DataLayout &DL, ScalarEvolution &SE,
                                   SmallVectorImpl<unsigned> &SortedIndices) {
  assert(llvm::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.
  MapVector<Value *, SmallVector<std::tuple<Value *, int, unsigned>>> Bases;
  Bases[VL[0]].push_back(std::make_tuple(VL[0], 0U, 0U));
 
  unsigned Cnt = 1;
  for (Value *Ptr : VL.drop_front()) {
    bool Found = any_of(Bases, [&](auto &Base) {
      std::optional<int> Diff =
          getPointersDiff(ElemTy, Base.first, ElemTy, Ptr, DL, SE,
                          /*StrictCheck=*/true);
      if (!Diff)
        return false;
 
      Base.second.emplace_back(Ptr, *Diff, Cnt++);
      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[Ptr].emplace_back(Ptr, 0, Cnt++);
    }
  }
 
  // For each of the bases sort the pointers by Offset and check if any of the
  // base become consecutively allocated.
  bool AnyConsecutive = false;
  for (auto &Base : Bases) {
    auto &Vec = Base.second;
    if (Vec.size() > 1) {
      llvm::stable_sort(Vec, [](const std::tuple<Value *, int, unsigned> &X,
                                const std::tuple<Value *, int, unsigned> &Y) {
        return std::get<1>(X) < std::get<1>(Y);
      });
      int InitialOffset = std::get<1>(Vec[0]);
      AnyConsecutive |= all_of(enumerate(Vec), [InitialOffset](const auto &P) {
        return std::get<1>(P.value()) == int(P.index()) + InitialOffset;
      });
    }
  }
 
  // Fill SortedIndices array only if it looks worth-while to sort the ptrs.
  SortedIndices.clear();
  if (!AnyConsecutive)
    return false;
 
  // If we have a better order, also sort the base pointers by increasing
  // (variable) values if possible, to try and keep the order more regular. In
  // order to create a valid strict-weak order we cluster by the Root of gep
  // chains and sort within each.
  SmallVector<std::tuple<Value *, Value *, Value *>> SortedBases;
  for (auto &Base : Bases) {
    Value *Strip = Base.first->stripInBoundsConstantOffsets();
    Value *Root = Strip;
    while (auto *Gep = dyn_cast<GetElementPtrInst>(Root))
      Root = Gep->getOperand(0);
    SortedBases.emplace_back(Base.first, Strip, Root);
  }
  auto *Begin = SortedBases.begin();
  auto *End = SortedBases.end();
  while (Begin != End) {
    Value *Root = std::get<2>(*Begin);
    auto *Mid = std::stable_partition(
        Begin, End, [&Root](auto V) { return std::get<2>(V) == Root; });
    DenseMap<Value *, DenseMap<Value *, bool>> LessThan;
    for (auto I = Begin; I < Mid; ++I)
      LessThan.try_emplace(std::get<1>(*I));
    for (auto I = Begin; I < Mid; ++I) {
      Value *V = std::get<1>(*I);
      while (auto *Gep = dyn_cast<GetElementPtrInst>(V)) {
        V = Gep->getOperand(0);
        if (LessThan.contains(V))
          LessThan[V][std::get<1>(*I)] = true;
      }
    }
    std::stable_sort(Begin, Mid, [&LessThan](auto &V1, auto &V2) {
      return LessThan[std::get<1>(V1)][std::get<1>(V2)];
    });
    Begin = Mid;
  }
 
  // Collect the final order of sorted indices
  for (auto Base : SortedBases)
    for (auto &T : Bases[std::get<0>(Base)])
      SortedIndices.push_back(std::get<2>(T));
 
  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());
  for (Value *V : TE.Scalars) {
    auto *L = dyn_cast<LoadInst>(V);
    if (!L || !L->isSimple())
      return std::nullopt;
    Ptrs.push_back(L->getPointerOperand());
  }
 
  BoUpSLP::OrdersType Order;
  if (clusterSortPtrAccesses(Ptrs, 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(
      cast<VectorType>(VU->getType())->getElementCount().getKnownMinValue());